question
stringlengths
101
4.47k
chosen
stringlengths
0
4.25k
rejected
stringlengths
39
9.29k
**Action Production**: Different from ordinary LLMs, where the model input and output are directly associated, the agent may take actions via different strategies and sources. In the following, we introduce two types of commonly used action production strategies. (1) _Action via Memory Recollection_. In this strategy, the action is generated by extracting information from the agent memory according to the current task. The task and the extracted memories are used as prompts to trigger the agent actions. For example, in Generative Agents [20], the agent maintains a memory stream, and before taking each action, it retrieves recent, relevant and important information from the memory steam to guide the agent actions. In GITM [16], in order to achieve a low-level sub-goal, the agent queries its memory to determine if there are any successful experiences related to the task. If similar tasks have been completed previously, the agent invokes the previously successful actions to handle the current task directly. In collaborative agents such as ChatDev [18] and MetaGPT [23], different agents may communicate with each other. In this process, the conversation history in a dialog is remembered in the agent memories. Each utterance generated by the agent is influenced by its memory. (2) _Action via Plan Following_. In this strategy, the agent takes actions following its pre-generated plans. For instance, in DEPS [33], for a given task, the agent first makes action plans. If there are no signals indicating plan failure, the agent will strictly adhere to these plans. In GITM [16], the agent makes high-level plans by decomposing the task into many sub-goals. Based on these plans, the agent takes actions to solve each sub-goal sequentially to complete the final task.
**Action Production**: ๋ชจ๋ธ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์ด ์ง์ ‘ ์—ฐ๊ด€๋˜์–ด ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ, ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต๊ณผ ์†Œ์Šค๋ฅผ ํ†ตํ•ด actions์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ action ์ƒ์„ฑ ์ „๋žต์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. (1) _Action via Memory Recollection_. ์ด ์ „๋žต์—์„œ๋Š” ํ˜„์žฌ ์ž‘์—…์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜์—ฌ action์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…๊ณผ ์ถ”์ถœ๋œ ๋ฉ”๋ชจ๋ฆฌ๋Š” ์—์ด์ „ํŠธ ๋™์ž‘์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๋Š” prompts๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Generative Agents [20]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์„ ์œ ์ง€ํ•˜๋ฉฐ, ๊ฐ ๋™์ž‘์„ ์ทจํ•˜๊ธฐ ์ „์— ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์—์„œ ์ตœ๊ทผ์˜ ๊ด€๋ จ์„ฑ ์žˆ๊ณ  ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์—์ด์ „ํŠธ ๋™์ž‘์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ๊ฒฝํ—˜์ด ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์— ์œ ์‚ฌํ•œ ์ž‘์—…์„ ์™„๋ฃŒํ•œ ์ ์ด ์žˆ๋‹ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด์ „์— ์„ฑ๊ณตํ•œ actions์„ ํ˜ธ์ถœํ•˜์—ฌ ํ˜„์žฌ ์ž‘์—…์„ ์ง์ ‘ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatDev [18] ๋ฐ MetaGPT [23]๊ณผ ๊ฐ™์€ ํ˜‘์—… ์—์ด์ „ํŠธ์—์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋Œ€ํ™”์˜ ๋Œ€ํ™” ๊ธฐ๋ก์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ์–ต๋ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ ๋ฐœํ™”๋Š” ๋ฉ”๋ชจ๋ฆฌ์˜ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. (2) _Action via Plan Following_. ์ด ์ „๋žต์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฏธ๋ฆฌ ์ƒ์„ฑ๋œ ๊ณ„ํš์— ๋”ฐ๋ผ actions์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, DEPS [33]์—์„œ๋Š” ์ฃผ์–ด์ง„ ์ž‘์—…์— ๋Œ€ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € action ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํŒจ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์‹ ํ˜ธ๊ฐ€ ์—†์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์ƒ์œ„ ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ๊ฐ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ actions์„ ์ทจํ•˜์—ฌ ์ตœ์ข… ์ž‘์—…์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.
**์•ก์…˜ ํ”„๋กœ๋•์…˜**: ๋ชจ๋ธ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์ด ์ง์ ‘ ์—ฐ๊ด€๋˜์–ด ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ LLM๊ณผ ๋‹ฌ๋ฆฌ, ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ์ „๋žต๊ณผ ์†Œ์Šค๋ฅผ ํ†ตํ•ด ์•ก์…˜์„ ์ทจํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐ€์ง€ ์œ ํ˜•์˜ ์•ก์…˜ ์ƒ์„ฑ ์ „๋žต์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. (1) _๋ฉ”๋ชจ๋ฆฌ ํšŒ์ƒ์„ ํ†ตํ•œ ์•ก์…˜_. ์ด ์ „๋žต์—์„œ๋Š” ํ˜„์žฌ ์ž‘์—…์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜์—ฌ ์•ก์…˜์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ž‘์—…๊ณผ ์ถ”์ถœ๋œ ๋ฉ”๋ชจ๋ฆฌ๋Š” ์—์ด์ „ํŠธ ๋™์ž‘์„ ํŠธ๋ฆฌ๊ฑฐํ•˜๋Š” ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ƒ์„ฑ ์—์ด์ „ํŠธ [20]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์„ ์œ ์ง€ํ•˜๋ฉฐ, ๊ฐ ๋™์ž‘์„ ์ทจํ•˜๊ธฐ ์ „์— ๋ฉ”๋ชจ๋ฆฌ ์ŠคํŠธ๋ฆผ์—์„œ ์ตœ๊ทผ์˜ ๊ด€๋ จ์„ฑ ์žˆ๊ณ  ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์—์ด์ „ํŠธ ๋™์ž‘์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ž‘์—…๊ณผ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ๊ฒฝํ—˜์ด ์žˆ๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์— ์œ ์‚ฌํ•œ ์ž‘์—…์„ ์™„๋ฃŒํ•œ ์ ์ด ์žˆ๋‹ค๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด์ „์— ์„ฑ๊ณตํ•œ ์ž‘์—…์„ ํ˜ธ์ถœํ•˜์—ฌ ํ˜„์žฌ ์ž‘์—…์„ ์ง์ ‘ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatDev [18] ๋ฐ MetaGPT [23]๊ณผ ๊ฐ™์€ ํ˜‘์—… ์—์ด์ „ํŠธ์—์„œ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์—์ด์ „ํŠธ๊ฐ€ ์„œ๋กœ ํ†ต์‹ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ๋Œ€ํ™”์˜ ๋Œ€ํ™” ๊ธฐ๋ก์€ ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ์— ๊ธฐ์–ต๋ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๊ฐ€ ์ƒ์„ฑํ•œ ๊ฐ ๋ฐœํ™”๋Š” ๋ฉ”๋ชจ๋ฆฌ์˜ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. (2) _๊ณ„ํš ์ถ”์ข…์„ ํ†ตํ•œ ์•ก์…˜_. ์ด ์ „๋žต์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฏธ๋ฆฌ ์ƒ์„ฑ๋œ ๊ณ„ํš์— ๋”ฐ๋ผ ์•ก์…˜์„ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, DEPS [33]์—์„œ๋Š” ์ฃผ์–ด์ง„ ์ž‘์—…์— ๋Œ€ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € ์ž‘์—… ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํŒจ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์‹ ํ˜ธ๊ฐ€ ์—†์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ์—„๊ฒฉํ•˜๊ฒŒ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜์—ฌ ์ƒ์œ„ ๊ณ„ํš์„ ์„ธ์›๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณ„ํš์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ๊ฐ ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์•ก์…˜์„ ์ทจํ•˜์—ฌ ์ตœ์ข… ์ž‘์—…์„ ์™„๋ฃŒํ•ฉ๋‹ˆ๋‹ค.
Given the need to mitigate forgetting when re-warming and re-decaying, we move on to investigate the effects of replay in our weak and strong-shift continued pre-training scenarios. Specifically, we use compete equivalent replay (see Sec. 4.2 for details) where replay tokens from \(\mathcal{D}_{0}\) are added at the cost of removing the equivalent number of \(\mathcal{D}_{1}\) tokens from the budget. Following the same two dataset settings, the model is pre-trained on \(\mathcal{D}_{0}\) (Pile) for 300B tokens. This is followed by continual pre-training on a SlimPajama (weak shift) or German Common Crawl (strong shift). For more details regarding the setup, please see Section 5.2. Our continued pre-training is conducted for the full size of the respective datasets, which is 300B tokens for SlimPajama (weak shift) and 200B tokens for German Common Crawl (strong shift). We consider 1%, 5%, 10%, and 50% replay for both shifts and add 0.5% and 25% replay runs for the weak and strong distribution shifts respectively. We consider two baselines to put these results into a broader context. The first baseline is a model trained on \(\mathcal{D}_{1}\) without replay. The second baseline model is trained from random initialization on a union of \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) for 600B tokens (SlimPajama) and 500B tokens (German Common Crawl). The latter baseline reflects the practice of fully re-training the model to update it instead of continually pre-training the existing model. All models re-warm and re-decay the learning rate using a cosine decay schedule fit to their token budget with the same \(\eta_{\textit{max}}\) (\(3\cdot 10^{-4}\)) and \(\eta_{\textit{min}}\) (\(3\cdot 10^{-5}\)) values as during pre-training on \(\mathcal{D}_{0}\).
์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์‡  ์‹œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•ด์•ผ ํ•  ํ•„์š”์„ฑ์ด ์ œ๊ธฐ๋จ์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ์•ฝํ•œ shift ๋ฐ ๊ฐ•ํ•œ shift ์ง€์† ์‚ฌ์ „ ํ•™์Šต ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ replay์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathcal{D}_{0}\)์˜ replay ํ† ํฐ์ด ์˜ˆ์‚ฐ์—์„œ ๋™์ผํ•œ ์ˆ˜์˜ \(\mathcal{D}_{1}\) ํ† ํฐ์„ ์ œ๊ฑฐํ•˜๋Š” ๋Œ€๊ฐ€๋กœ ์ถ”๊ฐ€๋˜๋Š” ๋™๋“ฑํ•œ replay(์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์„น์…˜ 4.2 ์ฐธ์กฐ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ค์ •์—์„œ ๋ชจ๋ธ์€ 300B ํ† ํฐ์— ๋Œ€ํ•ด \(\mathcal{D}_{0}\)(ํŒŒ์ผ)์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” SlimPajama(์•ฝํ•œ shift) ๋˜๋Š” German Common Crawl(๊ฐ•ํ•œ shift)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์„ค์ •์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 5.2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๊ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ „์ฒด ํฌ๊ธฐ, ์ฆ‰ SlimPajama(์•ฝํ•œ shift)์˜ ๊ฒฝ์šฐ 300B ํ† ํฐ, German Common Crawl(๊ฐ•ํ•œ shift)์˜ ๊ฒฝ์šฐ 200B ํ† ํฐ์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋‘ shift ๋ชจ๋‘์— ๋Œ€ํ•ด 1%, 5%, 10%, 50%์˜ replay๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์•ฝํ•œ ๋ถ„ํฌ shift์™€ ๊ฐ•ํ•œ ๋ถ„ํฌ shift์— ๋Œ€ํ•ด ๊ฐ๊ฐ 0.5%์™€ 25%์˜ replay๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๋‹ค ๋„“์€ ๋งฅ๋ฝ์— ๋„ฃ๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ baseline์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ baseline์€ replay ์—†์ด \(\mathcal{D}_{1}\)์—์„œ ํ•™์Šตํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ baseline ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์™€ \(\mathcal{D}_{1}\)์˜ ์กฐํ•ฉ์—์„œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋ถ€ํ„ฐ 600B ํ† ํฐ(SlimPajama) ๋ฐ 500B ํ† ํฐ(German Common Crawl)์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž์˜ baseline์€ ๊ธฐ์กด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋Œ€์‹  ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ์™„์ „ํžˆ ์žฌํ•™์Šตํ•˜๋Š” ๊ด€ํ–‰์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘๊ณผ ๋™์ผํ•œ \(\eta_{\textit{max}}\)(3\cdot 10^{-4}) ๋ฐ \(\eta_{\textit{min}}\)(3\cdot 10^{-5}) ๊ฐ’์œผ๋กœ cosine decay ์ผ์ •์„ ์‚ฌ์šฉํ•˜์—ฌ Learning Rate๋ฅผ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•ฉ๋‹ˆ๋‹ค.
์žฌ๊ฐ€์—ด ๋ฐ ์žฌ๊ฐ์‡  ์‹œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ์™„ํ™”ํ•ด์•ผ ํ•  ํ•„์š”์„ฑ์ด ์ œ๊ธฐ๋จ์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ์•ฝํ•œ ์‹œํ”„ํŠธ ๋ฐ ๊ฐ•ํ•œ ์‹œํ”„ํŠธ ์ง€์† ์‚ฌ์ „ ํ•™์Šต ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋ฆฌํ”Œ๋ ˆ์ด์˜ ํšจ๊ณผ๋ฅผ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, \(\mathcal{D}_{0}\)์˜ ๋ฆฌํ”Œ๋ ˆ์ด ํ† ํฐ์ด ์˜ˆ์‚ฐ์—์„œ ๋™์ผํ•œ ์ˆ˜์˜ \(\mathcal{D}_{1}\) ํ† ํฐ์„ ์ œ๊ฑฐํ•˜๋Š” ๋Œ€๊ฐ€๋กœ ์ถ”๊ฐ€๋˜๋Š” ๋™๋“ฑํ•œ ๋ฆฌํ”Œ๋ ˆ์ด(์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์„น์…˜ 4.2 ์ฐธ์กฐ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋™์ผํ•œ ๋‘ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์„ค์ •์—์„œ ๋ชจ๋ธ์€ 300B ํ† ํฐ์— ๋Œ€ํ•ด \(\mathcal{D}_{0}\)(ํŒŒ์ผ)์— ๋Œ€ํ•ด ์‚ฌ์ „ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” ์Šฌ๋ฆผํŒŒ์ž๋งˆ(์•ฝํ•œ ์‹œํ”„ํŠธ) ๋˜๋Š” ๋…์ผ ์ปค๋จผ ํฌ๋กค(๊ฐ•ํ•œ ์‹œํ”„ํŠธ)์— ๋Œ€ํ•œ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์ด ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์„ค์ •์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ์„น์…˜ 5.2๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์€ ๊ฐ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ „์ฒด ํฌ๊ธฐ, ์ฆ‰ ์Šฌ๋ฆผํŒŒ์ž๋งˆ(์•ฝํ•œ ์‹œํ”„ํŠธ)์˜ ๊ฒฝ์šฐ 300B ํ† ํฐ, ๋…์ผ ์ปค๋จผ ํฌ๋กค(๊ฐ•ํ•œ ์‹œํ”„ํŠธ)์˜ ๊ฒฝ์šฐ 200B ํ† ํฐ์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋‘ ์‹œํ”„ํŠธ ๋ชจ๋‘์— ๋Œ€ํ•ด 1%, 5%, 10%, 50%์˜ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ๊ณ ๋ คํ•˜๊ณ  ์•ฝํ•œ ๋ถ„ํฌ ์‹œํ”„ํŠธ์™€ ๊ฐ•ํ•œ ๋ถ„ํฌ ์‹œํ”„ํŠธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ 0.5%์™€ 25%์˜ ๋ฆฌํ”Œ๋ ˆ์ด๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๋‹ค ๋„“์€ ๋งฅ๋ฝ์— ๋„ฃ๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ๋ฒ ์ด์Šค๋ผ์ธ์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋ฒ ์ด์Šค๋ผ์ธ์€ ๋ฆฌํ”Œ๋ ˆ์ด ์—†์ด \(\mathcal{D}_{1}\)์—์„œ ํ•™์Šตํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋ฒ ์ด์Šค๋ผ์ธ ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์™€ \(\mathcal{D}_{1}\)์˜ ์กฐํ•ฉ์—์„œ ๋ฌด์ž‘์œ„ ์ดˆ๊ธฐํ™”๋ถ€ํ„ฐ 600B ํ† ํฐ(SlimPajama) ๋ฐ 500B ํ† ํฐ(German Common Crawl)์— ๋Œ€ํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ›„์ž์˜ ๋ฒ ์ด์Šค๋ผ์ธ์€ ๊ธฐ์กด ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๋Œ€์‹  ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์„ ์™„์ „ํžˆ ์žฌํ•™์Šตํ•˜๋Š” ๊ด€ํ–‰์„ ๋ฐ˜์˜ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์€ \(\mathcal{D}_{0}\)์— ๋Œ€ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘๊ณผ ๋™์ผํ•œ \(\eta_{\textit{max}}\)(3\cdot 10^{-4}) ๋ฐ \(\eta_{\textit{min}}\)(3\cdot 10^{-5}) ๊ฐ’์œผ๋กœ ์ฝ”์‚ฌ์ธ ๊ฐ์‡  ์ผ์ •์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต ์†๋„๋ฅผ ์žฌ๊ฐ€์—ดํ•˜๊ณ  ์žฌ๊ฐ์‡ ํ•ฉ๋‹ˆ๋‹ค.
**Public API of LLMs**. Instead of directly using the model copies, APIs provide a more convenient way for common users to use LLMs, without the need of running the model locally. As a representative interface for using LLMs, the APIs for the GPT-series models [46, 55, 66, 105] have been widely used for both academia and industry19. OpenAI has provided seven major interfaces to the models in GPT-3 series: ada, babbage, curie, davinci (the most powerful version in GPT-3 series), text-ada-001, text-babbbage-001, and text-curie-001. Among them, the first four interfaces can be further fine-tuned on the host server of OpenAI. In particular, babbage, curie, and davinci correspond to the GPT-3 (1B), GPT-3 (6.7B), and GPT-3 (175B) models, respectively [55]. In addition, there are also two APIs related to Codex [105], called code-cushman-001 (a powerful and multilingual version of the Codex (12B [105]) and code-davinci-002. Further, GPT-3.5 series include one base model code-davinci-002 and three enhanced versions, namely text-davinci-002, text-davinci-003, and gpt-3.5-turbo. As more powerful alternatives, in this year, OpenAI has released the model interfaces for GPT-4 series, including gpt-4, gpt-4-32k, gpt-4-1106-preview (_i.e._, GPT-4 Turbo) and gpt-4-vision-preview (_i.e._, GPT-4 Turbo with vision, a multimodal model). It is worth noting that OpenAI has been maintaining and upgrading these model interfaces (gpt-3.5-turbo, gpt-4, gpt-4-32k), so the API name will actually point to the latest version. Currently, ChatGPT can be powered by either GPT-3.5 or GPT-4 models. Overall, one select the suitable model interface based on the specific application scenarios and response requirements. The detailed usage can be found on their project websites20.
**Public API of LLMs**. API๋Š” ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์ผ๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋กœ์ปฌ์—์„œ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•  ํ•„์š” ์—†์ด LLM์„ ๋” ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์„ ์œ„ํ•œ ๋Œ€ํ‘œ์ ์ธ ์ธํ„ฐํŽ˜์ด์Šค์ธ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์šฉ API[46, 55, 66, 105]๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค19. OpenAI๋Š” GPT-3 ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์— ada, babbage, curie, davinci(GPT-3 ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋ฒ„์ „), text-ada-001, text-babbbage-001, text-curie-001 ๋“ฑ 7๊ฐ€์ง€ ์ฃผ์š” ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ฒ˜์Œ ๋„ค ๊ฐ€์ง€ ์ธํ„ฐํŽ˜์ด์Šค๋Š” OpenAI์˜ ํ˜ธ์ŠคํŠธ ์„œ๋ฒ„์—์„œ ์ถ”๊ฐ€๋กœ Fine-tuningํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, babbage, curie, davinci๋Š” ๊ฐ๊ฐ GPT-3(1B), GPT-3(6.7B), GPT-3(175B) ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค[55]. ๋˜ํ•œ Codex[105]์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ API, ์ฆ‰ code-cushman-001(Codex(12B[105])์˜ ๊ฐ•๋ ฅํ•˜๊ณ  ๋‹ค๊ตญ์–ด ๋ฒ„์ „)๊ณผ code-davinci-002๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-3.5 ์‹œ๋ฆฌ์ฆˆ์—๋Š” ๊ธฐ๋ณธ ๋ชจ๋ธ์ธ code-davinci-002์™€ ์„ธ ๊ฐ€์ง€ ํ–ฅ์ƒ๋œ ๋ฒ„์ „์ธ text-davinci-002, text-davinci-003, gpt-3.5-turbo๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์˜ฌํ•ด OpenAI๋Š” gpt-4, gpt-4-32k, gpt-4-1106-preview(_์ฆ‰, GPT-4 Turbo) ๋ฐ gpt-4-vision-preview(_์ฆ‰, ์‹œ๊ฐ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” GPT-4 Turbo, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ) ๋“ฑ GPT-4 ์‹œ๋ฆฌ์ฆˆ์˜ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ์ ์€ OpenAI๊ฐ€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค(gpt-3.5-turbo, gpt-4, gpt-4-32k)๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•˜๊ณ  ์—…๊ทธ๋ ˆ์ด๋“œํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ API ์ด๋ฆ„์€ ์‹ค์ œ๋กœ ์ตœ์‹  ๋ฒ„์ „์„ ๊ฐ€๋ฆฌํ‚ฌ ๊ฒƒ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ChatGPT๋Š” GPT-3.5 ๋˜๋Š” GPT-4 ๋ชจ๋ธ๋กœ ๊ตฌ๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ์‘๋‹ต ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‚ฌ์šฉ๋ฒ•์€ ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ20์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**LLM์˜ ๊ณต์šฉ API**. API๋Š” ๋ชจ๋ธ ๋ณต์‚ฌ๋ณธ์„ ์ง์ ‘ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ์ผ๋ฐ˜ ์‚ฌ์šฉ์ž๊ฐ€ ๋กœ์ปฌ์—์„œ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•  ํ•„์š” ์—†์ด LLM์„ ๋” ํŽธ๋ฆฌํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. LLM ์‚ฌ์šฉ์„ ์œ„ํ•œ ๋Œ€ํ‘œ์ ์ธ ์ธํ„ฐํŽ˜์ด์Šค์ธ GPT ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์šฉ API[46, 55, 66, 105]๋Š” ํ•™๊ณ„์™€ ์—…๊ณ„ ๋ชจ๋‘์—์„œ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค19. OpenAI๋Š” GPT-3 ์‹œ๋ฆฌ์ฆˆ ๋ชจ๋ธ์— ์•„๋‹ค, ๋ฐ”๋น„์ง€, ํ๋ฆฌ, ๋‹ค๋นˆ์น˜(GPT-3 ์‹œ๋ฆฌ์ฆˆ์—์„œ ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ๋ฒ„์ „), ํ…์ŠคํŠธ-์•„๋‹ค-001, ํ…์ŠคํŠธ-๋ฐ”๋น„์ง€-001, ํ…์ŠคํŠธ-ํ๋ฆฌ-001 ๋“ฑ 7๊ฐ€์ง€ ์ฃผ์š” ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘ ์ฒ˜์Œ ๋„ค ๊ฐ€์ง€ ์ธํ„ฐํŽ˜์ด์Šค๋Š” OpenAI์˜ ํ˜ธ์ŠคํŠธ ์„œ๋ฒ„์—์„œ ์ถ”๊ฐ€๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, babbage, curie, davinci๋Š” ๊ฐ๊ฐ GPT-3(1B), GPT-3(6.7B), GPT-3(175B) ๋ชจ๋ธ์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค[55]. ๋˜ํ•œ ์ฝ”๋ฑ์Šค[105]์™€ ๊ด€๋ จ๋œ ๋‘ ๊ฐ€์ง€ API, ์ฆ‰ code-cushman-001(์ฝ”๋ฑ์Šค(12B[105])์˜ ๊ฐ•๋ ฅํ•˜๊ณ  ๋‹ค๊ตญ์–ด ๋ฒ„์ „)๊ณผ code-davinci-002๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-3.5 ์‹œ๋ฆฌ์ฆˆ์—๋Š” ๊ธฐ๋ณธ ๋ชจ๋ธ์ธ code-davinci-002์™€ ์„ธ ๊ฐ€์ง€ ํ–ฅ์ƒ๋œ ๋ฒ„์ „์ธ text-davinci-002, text-davinci-003, gpt-3.5-turbo๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๊ฐ•๋ ฅํ•œ ๋Œ€์•ˆ์œผ๋กœ ์˜ฌํ•ด OpenAI๋Š” gpt-4, gpt-4-32k, gpt-4-1106-preview(_์ฆ‰, GPT-4 ํ„ฐ๋ณด) ๋ฐ gpt-4-vision-preview(_์ฆ‰, ์‹œ๊ฐ ๊ธฐ๋Šฅ์ด ์žˆ๋Š” GPT-4 ํ„ฐ๋ณด, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ) ๋“ฑ GPT-4 ์‹œ๋ฆฌ์ฆˆ์˜ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฃผ๋ชฉํ•  ์ ์€ OpenAI๊ฐ€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค(gpt-3.5-turbo, gpt-4, gpt-4-32k)๋ฅผ ์œ ์ง€ ๊ด€๋ฆฌํ•˜๊ณ  ์—…๊ทธ๋ ˆ์ด๋“œํ•˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ API ์ด๋ฆ„์€ ์‹ค์ œ๋กœ ์ตœ์‹  ๋ฒ„์ „์„ ๊ฐ€๋ฆฌํ‚ฌ ๊ฒƒ์ด๋ผ๋Š” ์ ์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ChatGPT๋Š” GPT-3.5 ๋˜๋Š” GPT-4 ๋ชจ๋ธ๋กœ ๊ตฌ๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ ํŠน์ • ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ์‘๋‹ต ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ๋ชจ๋ธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์„ ํƒํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์‚ฌ์šฉ๋ฒ•์€ ํ”„๋กœ์ ํŠธ ์›น์‚ฌ์ดํŠธ20์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
(1) _Trial-and-error_. In this method, the agent first performs an action, and subsequently, a pre-defined critic is invoked to judge the action. If the action is deemed unsatisfactory, then the agent reacts by incorporating the critic's feedback. In RAH [92], the agent serves as a user assistant in recommender systems. One of the agent's crucial roles is to simulate human behavior and generate responses on behalf of the user. To fulfill this objective, the agent first generates a predicted response and then compares it with the real human feedback. If the predicted response and the real human feedback differ, the critic generates failure information, which is subsequently incorporated into the agent's next action. In DEPS [33], the agent first designs a plan to accomplish a given task. In the plan execution process, if an action fails, the explainer generates specific details explaining the cause of the failure. This information is then incorporated by the agent to redesign the plan. In RoCo [93], the agent first proposes a sub-task plan and a path of 3D waypoints for each robot in a multi-robot collaboration task. The plan and waypoints are then validated by a set of environment checks, such as collision detection and inverse kinematics. If any of the checks fail, the feedback is appended to each agent's prompt and another round of dialog begins. The agents use LLMs to discuss and improve their plan and waypoints until they pass all validations. In PREFER [94], the agent first evaluates its performance on a subset of data. If it fails to solve certain examples, LLMs are leveraged to generate feedback information reflecting on the reasons of the failure. Based on this feedback, the agent improves itself by iteratively refining its actions.
(1) _Trial-and-error_. ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ๋จผ์ € agent๊ฐ€ action์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ ์‚ฌ์ „ ์ •์˜๋œ critic์„ ํ˜ธ์ถœํ•˜์—ฌ action์„ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. action์ด ๋ถˆ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋˜๋ฉด agent๋Š” critic์˜ feedback์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ฐ˜์‘ํ•ฉ๋‹ˆ๋‹ค. RAH [92]์—์„œ agent๋Š” ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ ์‚ฌ์šฉ์ž ๋„์šฐ๋ฏธ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. agent์˜ ์ค‘์š”ํ•œ ์—ญํ•  ์ค‘ ํ•˜๋‚˜๋Š” ์ธ๊ฐ„์˜ ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ณ  ์‚ฌ์šฉ์ž๋ฅผ ๋Œ€์‹ ํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด agent๋Š” ๋จผ์ € ์˜ˆ์ธก๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์‹ค์ œ ์‚ฌ์šฉ์ž feedback๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ ์‘๋‹ต๊ณผ ์‹ค์ œ ์‚ฌ์šฉ์ž feedback์ด ๋‹ค๋ฅด๋ฉด critic์€ ์‹คํŒจ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ด ์ •๋ณด๋Š” agent์˜ ๋‹ค์Œ ์ž‘์—…์— ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. DEPS [33]์—์„œ agent๋Š” ๋จผ์ € ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•œ ๊ณ„ํš์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํ–‰ ๊ณผ์ •์—์„œ ์–ด๋–ค ์ž‘์—…์ด ์‹คํŒจํ•˜๋ฉด explainer๋Š” ์‹คํŒจ ์›์ธ์„ ์„ค๋ช…ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ์ •๋ณด๋Š” agent๊ฐ€ ๊ณ„ํš์„ ์žฌ์„ค๊ณ„ํ•  ๋•Œ ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. RoCo [93]์—์„œ agent๋Š” ๋จผ์ € ๋ฉ€ํ‹ฐ ๋กœ๋ด‡ ํ˜‘์—… ์ž‘์—…์—์„œ ๊ฐ ๋กœ๋ด‡์— ๋Œ€ํ•œ ํ•˜์œ„ ์ž‘์—… ๊ณ„ํš๊ณผ 3D ์›จ์ดํฌ์ธํŠธ ๊ฒฝ๋กœ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ถฉ๋Œ ๊ฐ์ง€ ๋ฐ ์—ญ๋™์—ญํ•™๊ณผ ๊ฐ™์€ ์ผ๋ จ์˜ ํ™˜๊ฒฝ ๊ฒ€์‚ฌ๋ฅผ ํ†ตํ•ด ๊ณ„ํš๊ณผ ์›จ์ดํฌ์ธํŠธ์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์‚ฌ ์ค‘ ํ•˜๋‚˜๋ผ๋„ ์‹คํŒจํ•˜๋ฉด ๊ฐ agent์˜ prompt์— feedback์ด ์ถ”๊ฐ€๋˜๊ณ  ๋‹ค๋ฅธ ๋ผ์šด๋“œ์˜ ๋Œ€ํ™”๊ฐ€ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค. agent๋Š” ๋ชจ๋“  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ๋ฅผ ํ†ต๊ณผํ•  ๋•Œ๊นŒ์ง€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„ํš๊ณผ ์›จ์ดํฌ์ธํŠธ๋ฅผ ๋…ผ์˜ํ•˜๊ณ  ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. PREFER [94]์—์„œ agent๋Š” ๋จผ์ € ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŠน์ • ์˜ˆ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•˜๋ฉด ์‹คํŒจ ์›์ธ์„ ๋ฐ˜์˜ํ•˜๋Š” feedback ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด feedback์„ ๊ธฐ๋ฐ˜์œผ๋กœ agent๋Š” ๋ฐ˜๋ณต์ ์œผ๋กœ ๋™์ž‘์„ ๊ฐœ์„ ํ•˜์—ฌ ์ž์‹ ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
(1) _์‹œํ–‰์ฐฉ์˜ค_. ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ๋จผ์ € ์—์ด์ „ํŠธ๊ฐ€ ์•ก์…˜์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ ์‚ฌ์ „ ์ •์˜๋œ ๋น„ํ‰๊ฐ€๋ฅผ ํ˜ธ์ถœํ•˜์—ฌ ์•ก์…˜์„ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. ์•ก์…˜์ด ๋ถˆ๋งŒ์กฑ์Šค๋Ÿฌ์šด ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋˜๋ฉด ์—์ด์ „ํŠธ๋Š” ๋น„ํ‰๊ฐ€์˜ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ฉํ•˜์—ฌ ๋ฐ˜์‘ํ•ฉ๋‹ˆ๋‹ค. RAH [92]์—์„œ ์—์ด์ „ํŠธ๋Š” ์ถ”์ฒœ ์‹œ์Šคํ…œ์—์„œ ์‚ฌ์šฉ์ž ๋„์šฐ๋ฏธ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ์ค‘์š”ํ•œ ์—ญํ•  ์ค‘ ํ•˜๋‚˜๋Š” ์ธ๊ฐ„์˜ ํ–‰๋™์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ณ  ์‚ฌ์šฉ์ž๋ฅผ ๋Œ€์‹ ํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์˜ˆ์ธก๋œ ์‘๋‹ต์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ์‹ค์ œ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ๊ณผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ์ธก๋œ ์‘๋‹ต๊ณผ ์‹ค์ œ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์ด ๋‹ค๋ฅด๋ฉด ๋น„ํ‰๊ฐ€๋Š” ์‹คํŒจ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์ด ์ •๋ณด๋Š” ์—์ด์ „ํŠธ์˜ ๋‹ค์Œ ์ž‘์—…์— ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. DEPS [33]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์ฃผ์–ด์ง„ ์ž‘์—…์„ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•œ ๊ณ„ํš์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ๊ณ„ํš ์‹คํ–‰ ๊ณผ์ •์—์„œ ์–ด๋–ค ์ž‘์—…์ด ์‹คํŒจํ•˜๋ฉด ์„ค๋ช…๊ธฐ๋Š” ์‹คํŒจ ์›์ธ์„ ์„ค๋ช…ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด ์ •๋ณด๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๊ณ„ํš์„ ์žฌ์„ค๊ณ„ํ•  ๋•Œ ํ†ตํ•ฉ๋ฉ๋‹ˆ๋‹ค. RoCo [93]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฉ€ํ‹ฐ ๋กœ๋ด‡ ํ˜‘์—… ์ž‘์—…์—์„œ ๊ฐ ๋กœ๋ด‡์— ๋Œ€ํ•œ ํ•˜์œ„ ์ž‘์—… ๊ณ„ํš๊ณผ 3D ์›จ์ดํฌ์ธํŠธ ๊ฒฝ๋กœ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ถฉ๋Œ ๊ฐ์ง€ ๋ฐ ์—ญ๋™์—ญํ•™๊ณผ ๊ฐ™์€ ์ผ๋ จ์˜ ํ™˜๊ฒฝ ๊ฒ€์‚ฌ๋ฅผ ํ†ตํ•ด ๊ณ„ํš๊ณผ ์›จ์ดํฌ์ธํŠธ์˜ ์œ ํšจ์„ฑ์„ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์‚ฌ ์ค‘ ํ•˜๋‚˜๋ผ๋„ ์‹คํŒจํ•˜๋ฉด ๊ฐ ์—์ด์ „ํŠธ์˜ ํ”„๋กฌํ”„ํŠธ์— ํ”ผ๋“œ๋ฐฑ์ด ์ถ”๊ฐ€๋˜๊ณ  ๋‹ค๋ฅธ ๋ผ์šด๋“œ์˜ ๋Œ€ํ™”๊ฐ€ ์‹œ์ž‘๋ฉ๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ๋Š” ๋ชจ๋“  ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ๋ฅผ ํ†ต๊ณผํ•  ๋•Œ๊นŒ์ง€ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ณ„ํš๊ณผ ์›จ์ดํฌ์ธํŠธ๋ฅผ ๋…ผ์˜ํ•˜๊ณ  ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. PREFER [94]์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŠน์ • ์˜ˆ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์ง€ ๋ชปํ•˜๋ฉด ์‹คํŒจ ์›์ธ์„ ๋ฐ˜์˜ํ•˜๋Š” ํ”ผ๋“œ๋ฐฑ ์ •๋ณด๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ๋ฐ˜๋ณต์ ์œผ๋กœ ๋™์ž‘์„ ๊ฐœ์„ ํ•˜์—ฌ ์ž์‹ ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
Less than a handful of works begin to provide theoretical explanations for the behavior of models in the "synthetic data age". (Shumailov et al., 2023) attribute model collapse to two mechanisms: a finite sampling bias cutting off low-probability "tails", thus leading to more and more peaked distributions and function approximation errors; they theoretically analyze the (single) Gaussian case and provide empirical evidence for VAEs, Gaussian mixtures and the OPT language model (125M parameters). In the context of vision models, Alemohammad et al. (2023) analyze "_self-consuming loops"_ by introducing a sampling bias that narrows the variance of the data at each generation, and, in addition to empirical demonstration on GANs and denoising diffusion probabilistic models, provide theoretical analysis for the Gaussian model. Finally, let us mention the study of Bertrand et al. (2023) which sheds light on the critical role of data composition in the stability and effectiveness in generative models, applicable to VAEs (Kingma & Welling, 2014), diffusion models and normalizing flows. They explore scenarios involving a mix of clean data, representative of the true distribution, and synthesized data from previous iterations of the generator. Their analysis reveals that if the data mix consists exclusively of synthesized data, the generative process is likely to degenerate over time (_"clueless generator"_). Using fixed-point analysis across iterations, they find that when the proportion of clean data in the mix is sufficiently high, the generator, under certain technical conditions, retains the capability to learn. A recent paper (Fan et al., 2023) empirically observe deteriorated scaling lawswhen training on synthetic data for text-to-image models.
"ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์‹œ๋Œ€"์— ๋ชจ๋ธ์˜ ๋™์ž‘์— ๋Œ€ํ•œ ์ด๋ก ์  ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋Š” ์ž‘์—…์€ ์†Œ์ˆ˜์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. (Shumailov et al., 2023)์€ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ๋‘ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜, ์ฆ‰ ๋‚ฎ์€ ํ™•๋ฅ ์˜ "๊ผฌ๋ฆฌ"๋ฅผ ์ฐจ๋‹จํ•˜์—ฌ ์ ์  ๋” ๋พฐ์กฑํ•œ ๋ถ„ํฌ์™€ ํ•จ์ˆ˜ ๊ทผ์‚ฌ ์˜ค๋ฅ˜๋กœ ์ด์–ด์ง€๋Š” ์œ ํ•œ ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ๊ณผ ๊ด€๋ จ๋œ ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. (๋‹จ์ผ) Gaussian์˜ ๊ฒฝ์šฐ๋ฅผ ์ด๋ก ์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  VAE, Gaussian ํ˜ผํ•ฉ ๋ฐ OPT language model(1์–ต 2,500๋งŒ ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜)์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ ๋ชจ๋ธ์˜ ๋งฅ๋ฝ์—์„œ Alemohammad et al. (2023)์€ ์ƒ์„ฑํ•  ๋•Œ๋งˆ๋‹ค ๋ฐ์ดํ„ฐ์˜ ๋ถ„์‚ฐ์„ ์ขํžˆ๋Š” ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ์„ ๋„์ž…ํ•˜์—ฌ "_์ž๊ธฐ ์†Œ๋น„ ๋ฃจํ”„"_๋ฅผ ๋ถ„์„ํ•˜๊ณ , GAN ๋ฐ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ™•์‚ฐ ํ™•๋ฅ  ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์‹œ์—ฐ ์™ธ์—๋„ Gaussian ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, VAEs(Kingma & Welling, 2014), diffusion models ๋ฐ normalizing flows์— ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์—์„œ ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•˜๋Š” Bertrand et al. (2023)์˜ ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ์–ธ๊ธ‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์‹ค์ œ ๋ถ„ํฌ๋ฅผ ๋Œ€ํ‘œํ•˜๋Š” ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๊ธฐ์˜ ์ด์ „ ๋ฐ˜๋ณต์—์„œ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํƒ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ๋ฏน์Šค๊ฐ€ ์ „์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ๋งŒ ๊ตฌ์„ฑ๋œ ๊ฒฝ์šฐ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ๋ณ€์งˆ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค("_"์ด ์ง€์‹์ด ์—†๋Š” ์ƒ์„ฑ๊ธฐ"_). ๋ฐ˜๋ณต์— ๊ฑธ์ณ ๊ณ ์ •์  ๋ถ„์„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏน์Šค์—์„œ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์˜ ๋น„์œจ์ด ์ถฉ๋ถ„ํžˆ ๋†’์œผ๋ฉด ํŠน์ • ๊ธฐ์ˆ ์  ์กฐ๊ฑด์—์„œ ์ƒ์„ฑ๊ธฐ๊ฐ€ ํ•™์Šต ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋…ผ๋ฌธ(Fan et al., 2023)์—์„œ๋Š” ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•  ๋•Œ scaling laws๊ฐ€ ์•…ํ™”๋˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
"ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์‹œ๋Œ€"์— ๋ชจ๋ธ์˜ ๋™์ž‘์— ๋Œ€ํ•œ ์ด๋ก ์  ์„ค๋ช…์„ ์ œ๊ณตํ•˜๋Š” ์ž‘์—…์€ ์†Œ์ˆ˜์— ๋ถˆ๊ณผํ•ฉ๋‹ˆ๋‹ค. (Shumailov et al., 2023)์€ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ๋‘ ๊ฐ€์ง€ ๋ฉ”์ปค๋‹ˆ์ฆ˜, ์ฆ‰ ๋‚ฎ์€ ํ™•๋ฅ ์˜ "๊ผฌ๋ฆฌ"๋ฅผ ์ฐจ๋‹จํ•˜์—ฌ ์ ์  ๋” ๋พฐ์กฑํ•œ ๋ถ„ํฌ์™€ ํ•จ์ˆ˜ ๊ทผ์‚ฌ ์˜ค๋ฅ˜๋กœ ์ด์–ด์ง€๋Š” ์œ ํ•œ ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ๊ณผ ๊ด€๋ จ๋œ ๊ฒƒ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. (๋‹จ์ผ) ๊ฐ€์šฐ์Šค์˜ ๊ฒฝ์šฐ๋ฅผ ์ด๋ก ์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  VAE, ๊ฐ€์šฐ์Šค ํ˜ผํ•ฉ ๋ฐ OPT ์–ธ์–ด ๋ชจ๋ธ(1์–ต 2,500๋งŒ ๊ฐœ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜)์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์‹œ๊ฐ ๋ชจ๋ธ์˜ ๋งฅ๋ฝ์—์„œ Alemohammad et al. (2023)์€ ์ƒ์„ฑํ•  ๋•Œ๋งˆ๋‹ค ๋ฐ์ดํ„ฐ์˜ ๋ถ„์‚ฐ์„ ์ขํžˆ๋Š” ์ƒ˜ํ”Œ๋ง ํŽธํ–ฅ์„ ๋„์ž…ํ•˜์—ฌ "_์ž๊ธฐ ์†Œ๋น„ ๋ฃจํ”„"_๋ฅผ ๋ถ„์„ํ•˜๊ณ , GAN ๋ฐ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํ™•์‚ฐ ํ™•๋ฅ  ๋ชจ๋ธ์— ๋Œ€ํ•œ ๊ฒฝํ—˜์  ์‹œ์—ฐ ์™ธ์—๋„ ๊ฐ€์šฐ์Šค ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด๋ก ์  ๋ถ„์„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, VAEs(Kingma & Welling, 2014), ํ™•์‚ฐ ๋ชจ๋ธ ๋ฐ ์ •๊ทœํ™” ํ๋ฆ„์— ์ ์šฉ ๊ฐ€๋Šฅํ•œ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์—์„œ ๋ฐ์ดํ„ฐ ๊ตฌ์„ฑ์˜ ์ค‘์š”ํ•œ ์—ญํ• ์„ ์กฐ๋ช…ํ•˜๋Š” Bertrand et al. (2023)์˜ ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ์–ธ๊ธ‰ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์‹ค์ œ ๋ถ„ํฌ๋ฅผ ๋Œ€ํ‘œํ•˜๋Š” ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๊ธฐ์˜ ์ด์ „ ๋ฐ˜๋ณต์—์„œ ํ•ฉ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ํ˜ผํ•ฉ๋œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํƒ์ƒ‰ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์˜ ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด ๋ฐ์ดํ„ฐ ๋ฏน์Šค๊ฐ€ ์ „์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ๋งŒ ๊ตฌ์„ฑ๋œ ๊ฒฝ์šฐ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๊ฐ€ ๋ณ€์งˆ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค("_"์ด ์ง€์‹์ด ์—†๋Š” ์ƒ์„ฑ๊ธฐ"_). ๋ฐ˜๋ณต์— ๊ฑธ์ณ ๊ณ ์ •์  ๋ถ„์„์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏน์Šค์—์„œ ๊นจ๋—ํ•œ ๋ฐ์ดํ„ฐ์˜ ๋น„์œจ์ด ์ถฉ๋ถ„ํžˆ ๋†’์œผ๋ฉด ํŠน์ • ๊ธฐ์ˆ ์  ์กฐ๊ฑด์—์„œ ์ƒ์„ฑ๊ธฐ๊ฐ€ ํ•™์Šต ๊ธฐ๋Šฅ์„ ์œ ์ง€ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋…ผ๋ฌธ(Fan et al., 2023)์—์„œ๋Š” ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•ด ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์ด ์•…ํ™”๋˜๋Š” ๊ฒƒ์„ ๊ฒฝํ—˜์ ์œผ๋กœ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค.
Factuality.AI systems may generate information or responses that are not grounded in factual knowledge or data, leading to the creation of misleading or false content, formally known as _hallucination_(Ji et al., 2023). Factuality evaluation aims to ensure the consistency of the knowledge in the AI system's output with the knowledge provided by its training data and knowledge base (Ji et al., 2023; Zhang et al., 2023c). Early statistical-based hallucination evaluation methods relied on n-grams to directly calculate the overlap of vocabulary between the input and output content (Dhingra et al., 2019; Wang et al., 2020). However, these methods have limitations, as they only consider lexical overlap and do not account for semantics or sentence meaning (Ji et al., 2023), making them unsuitable for evaluating more complex forms of hallucination. Subsequent assurance methods shifted from statistical approaches to model-based methods, which are more robust compared to token-difference-based methods (Honovich et al., 2021). While these model-based evaluation methods are more advanced than their predecessors, they still have limitations. For example, the models can only output the degree of hallucination and may struggle to pinpoint specific errors (Falke et al., 2019). Feng et al. (2023a) propose to combine LLMs generation with random walks on knowledge graphs to generate synthetic evaluation data for factuality, which is aware of entities and relations on the graphs. Wei et al. (2024) created a synthetic dataset called LongFact for long-form factuality evaluation and used Google Search as the grounding source and LLM for the automated judgement, to achieve human-level accuracy but with significantly lower cost (Min et al., 2023).
์‚ฌ์‹ค์„ฑ. AI ์‹œ์Šคํ…œ์€ ์‚ฌ์‹ค์ ์ธ ์ง€์‹์ด๋‚˜ ๋ฐ์ดํ„ฐ์— ๊ทผ๊ฑฐํ•˜์ง€ ์•Š์€ ์ •๋ณด๋‚˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜์—ฌ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๊ฑฐ๋‚˜ ์ž˜๋ชป๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ณต์‹์ ์œผ๋กœ _hallucination_(Ji et al., 2023)์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์‹ค์„ฑ ํ‰๊ฐ€๋Š” AI ์‹œ์Šคํ…œ์˜ ์ถœ๋ ฅ์— ํฌํ•จ๋œ ์ง€์‹์ด ํ•™์Šต ๋ฐ์ดํ„ฐ ๋ฐ ์ง€์‹์ฐฝ๊ณ ์—์„œ ์ œ๊ณตํ•œ ์ง€์‹๊ณผ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(Ji et al., 2023; Zhang et al., 2023c). ์ดˆ๊ธฐ์˜ ํ†ต๊ณ„ ๊ธฐ๋ฐ˜ hallucination ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์€ n-๊ทธ๋žจ์— ์˜์กดํ•˜์—ฌ ์ž…๋ ฅ ์ฝ˜ํ…์ธ ์™€ ์ถœ๋ ฅ ์ฝ˜ํ…์ธ  ๊ฐ„์˜ ์–ดํœ˜ ์ค‘๋ณต์„ ์ง์ ‘ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค(Dhingra et al., 2019; Wang et al., 2020). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์–ดํœ˜ ์ค‘๋ณต๋งŒ ๊ณ ๋ คํ•˜๊ณ  ์˜๋ฏธ๋ก ์ด๋‚˜ ๋ฌธ์žฅ ์˜๋ฏธ๋ฅผ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์—(Ji et al., 2023) ๋ณด๋‹ค ๋ณต์žกํ•œ ํ˜•ํƒœ์˜ hallucination์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ๋Š” ์ ํ•ฉํ•˜์ง€ ์•Š๋‹ค๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๋ณด์ฆ ๋ฐฉ๋ฒ•์€ ํ†ต๊ณ„์  ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ํ† ํฐ ์ฐจ์ด ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์œผ๋กœ ์ „ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค(Honovich et al., 2021). ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์€ ์ด์ „ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๊ณ ๊ธ‰ํ™”๋˜์—ˆ์ง€๋งŒ ์—ฌ์ „ํžˆ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ์€ hallucination์˜ ์ •๋„๋งŒ ์ถœ๋ ฅํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ํŠน์ • ์˜ค๋ฅ˜๋ฅผ ์ •ํ™•ํžˆ ์ฐพ์•„๋‚ด๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Falke et al., 2019). Feng et al. (2023a)์€ LLM ์ƒ์„ฑ๊ณผ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ์˜ ๋žœ๋ค ์›Œํฌ๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ทธ๋ž˜ํ”„์˜ ์—”ํ‹ฐํ‹ฐ์™€ ๊ด€๊ณ„๋ฅผ ์ธ์‹ํ•˜๋Š” ์‚ฌ์‹ค์„ฑ์„ ์œ„ํ•œ ํ•ฉ์„ฑ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Wei et al. (2024)์€ ์žฅ๋ฌธ ์‚ฌ์‹ค์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด LongFact๋ผ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค๊ณ , ๊ทธ๋ผ์šด๋”ฉ ์†Œ์Šค์™€ ์ž๋™ ํŒ๋‹จ์„ ์œ„ํ•œ LLM์œผ๋กœ Google ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ์ˆ˜์ค€์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ–ˆ์ง€๋งŒ ๋น„์šฉ์€ ํ›จ์”ฌ ๋‚ฎ์•˜์Šต๋‹ˆ๋‹ค(Min et al., 2023).
์‚ฌ์‹ค์„ฑ.AI ์‹œ์Šคํ…œ์€ ์‚ฌ์‹ค์ ์ธ ์ง€์‹์ด๋‚˜ ๋ฐ์ดํ„ฐ์— ๊ทผ๊ฑฐํ•˜์ง€ ์•Š์€ ์ •๋ณด๋‚˜ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜์—ฌ ์˜คํ•ด์˜ ์†Œ์ง€๊ฐ€ ์žˆ๊ฑฐ๋‚˜ ์ž˜๋ชป๋œ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๊ณต์‹์ ์œผ๋กœ _ํ™˜๊ฐ_(Ji et al., 2023)์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ์‹ค์„ฑ ํ‰๊ฐ€๋Š” AI ์‹œ์Šคํ…œ์˜ ์ถœ๋ ฅ์— ํฌํ•จ๋œ ์ง€์‹์ด ํ•™์Šต ๋ฐ์ดํ„ฐ ๋ฐ ์ง€์‹์ฐฝ๊ณ ์—์„œ ์ œ๊ณตํ•œ ์ง€์‹๊ณผ ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(Ji et al., 2023; Zhang et al., 2023c). ์ดˆ๊ธฐ์˜ ํ†ต๊ณ„ ๊ธฐ๋ฐ˜ ํ™˜๊ฐ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์€ n-๊ทธ๋žจ์— ์˜์กดํ•˜์—ฌ ์ž…๋ ฅ ์ฝ˜ํ…์ธ ์™€ ์ถœ๋ ฅ ์ฝ˜ํ…์ธ  ๊ฐ„์˜ ์–ดํœ˜ ์ค‘๋ณต์„ ์ง์ ‘ ๊ณ„์‚ฐํ–ˆ์Šต๋‹ˆ๋‹ค(Dhingra et al., 2019; Wang et al., 2020). ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•์€ ์–ดํœ˜ ์ค‘๋ณต๋งŒ ๊ณ ๋ คํ•˜๊ณ  ์˜๋ฏธ๋ก ์ด๋‚˜ ๋ฌธ์žฅ ์˜๋ฏธ๋ฅผ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์—(Ji et al., 2023) ๋ณด๋‹ค ๋ณต์žกํ•œ ํ˜•ํƒœ์˜ ํ™˜๊ฐ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ๋Š” ์ ํ•ฉํ•˜์ง€ ์•Š๋‹ค๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๋ณด์ฆ ๋ฐฉ๋ฒ•์€ ํ†ต๊ณ„์  ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ํ† ํฐ ์ฐจ์ด ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ๋ฐฉ๋ฒ•์œผ๋กœ ์ „ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค(Honovich et al., 2021). ์ด๋Ÿฌํ•œ ๋ชจ๋ธ ๊ธฐ๋ฐ˜ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์€ ์ด์ „ ๋ฐฉ๋ฒ•๋ณด๋‹ค ๊ณ ๊ธ‰ํ™”๋˜์—ˆ์ง€๋งŒ ์—ฌ์ „ํžˆ ํ•œ๊ณ„๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋ชจ๋ธ์€ ํ™˜๊ฐ์˜ ์ •๋„๋งŒ ์ถœ๋ ฅํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ํŠน์ • ์˜ค๋ฅ˜๋ฅผ ์ •ํ™•ํžˆ ์ฐพ์•„๋‚ด๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Falke et al., 2019). Feng et al. (2023a)์€ LLM ์ƒ์„ฑ๊ณผ ์ง€์‹ ๊ทธ๋ž˜ํ”„์—์„œ์˜ ๋žœ๋ค ์›Œํฌ๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ทธ๋ž˜ํ”„์˜ ์—”ํ‹ฐํ‹ฐ์™€ ๊ด€๊ณ„๋ฅผ ์ธ์‹ํ•˜๋Š” ์‚ฌ์‹ค์„ฑ์„ ์œ„ํ•œ ํ•ฉ์„ฑ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Wei et al. (2024)์€ ์žฅ๋ฌธ ์‚ฌ์‹ค์„ฑ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด LongFact๋ผ๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋งŒ๋“ค๊ณ , ๊ทธ๋ผ์šด๋”ฉ ์†Œ์Šค์™€ ์ž๋™ ํŒ๋‹จ์„ ์œ„ํ•œ LLM์œผ๋กœ Google ๊ฒ€์ƒ‰์„ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ์ˆ˜์ค€์˜ ์ •ํ™•๋„๋ฅผ ๋‹ฌ์„ฑํ–ˆ์ง€๋งŒ ๋น„์šฉ์€ ํ›จ์”ฌ ๋‚ฎ์•˜์Šต๋‹ˆ๋‹ค(Min et al., 2023).
\(\bullet\)_Model Feedback._ Apart from the aforementioned environmental and human feedback, which are external signals, researchers have also investigated the utilization of internal feedback from the agents themselves. This type of feedback is usually generated based on pre-trained models. In specific, [62] proposes a self-refine mechanism. This mechanism consists of three crucial components: output, feedback, and refinement. Firstly, the agent generates an output. Then, it utilizes LLMs to provide feedback on the output and offer guidance on how to refine it. At last, the output is improved by the feedback and refinement. This output-feedback-refinement process iterates until reaching some desired conditions. SelfCheck [63] allows agents to examine and evaluate their reasoning steps generated at various stages. They can then correct any errors by comparing the outcomes. InterAct [64] uses different language models (such as ChatGPT and InstructGPT) as auxiliary roles, such as checkers and sorters, to help the main language model avoid erroneous and inefficient actions. ChatCoT [65] utilizes model feedback to improve the quality of its reasoning process. The model feedback is generated by an evaluation module that monitors the agent reasoning steps. Reflexion [12] is developed to enhance the agent's planning capability through detailed verbal feedback. In this model, the agent first produces an action based on its memory, and then, the evaluator generates feedback by taking the agent trajectory as input. In contrast to previous studies, where the feedback is given as a scalar value, this model leverages LLMs to provide more detailed verbal feedback, which can provide more comprehensive supports for the agent plans.
\(\bullet\)_Model Feedback._ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์™ธ๋ถ€ ์‹ ํ˜ธ์ธ ํ™˜๊ฒฝ ๋ฐ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ์™ธ์—๋„ ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ์ž์ฒด์˜ ๋‚ด๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ํ”ผ๋“œ๋ฐฑ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ [62]๋Š” self-refine ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถœ๋ ฅ, ํ”ผ๋“œ๋ฐฑ, ๊ฐœ์„ ์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๊ฐ€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLMs์„ ํ™œ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ  ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์„ ํ†ตํ•ด ์ถœ๋ ฅ์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. ์ด ์ถœ๋ ฅ-ํ”ผ๋“œ๋ฐฑ-๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ์›ํ•˜๋Š” ์กฐ๊ฑด์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณต๋ฉ๋‹ˆ๋‹ค. SelfCheck [63]๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. InterAct [64]๋Š” ๋‹ค๋ฅธ language models (์˜ˆ: ChatGPT ๋ฐ InstructGPT)์„ ๊ฒ€์‚ฌ์ž ๋ฐ ์ •๋ ฌ์ž์™€ ๊ฐ™์€ ๋ณด์กฐ ์—ญํ• ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž˜๋ชป๋œ ๋™์ž‘๊ณผ ๋น„ํšจ์œจ์ ์ธ ๋™์ž‘์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค. ChatCoT [65]๋Š” model feedback์„ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. model feedback์€ ์—์ด์ „ํŠธ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋Š” ํ‰๊ฐ€ ๋ชจ๋“ˆ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. Reflexion [12]์€ ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ์˜ ๊ณ„ํš ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ก์…˜์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ํ‰๊ฐ€๊ธฐ๋Š” ์—์ด์ „ํŠธ ๊ถค์ ์„ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์ด ์Šค์นผ๋ผ ๊ฐ’์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, ์ด ๋ชจ๋ธ์€ LLMs์„ ํ™œ์šฉํ•˜์—ฌ ๋ณด๋‹ค ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ ๊ณ„ํš์— ๋Œ€ํ•œ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ._ ์•ž์„œ ์–ธ๊ธ‰ํ•œ ์™ธ๋ถ€ ์‹ ํ˜ธ์ธ ํ™˜๊ฒฝ ๋ฐ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ์™ธ์—๋„ ์—ฐ๊ตฌ์ž๋“ค์€ ์—์ด์ „ํŠธ ์ž์ฒด์˜ ๋‚ด๋ถ€ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ํ”ผ๋“œ๋ฐฑ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. ํŠนํžˆ [62]๋Š” ์ž์ฒด ๊ฐœ์„  ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฉ”์ปค๋‹ˆ์ฆ˜์€ ์ถœ๋ ฅ, ํ”ผ๋“œ๋ฐฑ, ๊ฐœ์„ ์ด๋ผ๋Š” ์„ธ ๊ฐ€์ง€ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๊ฐ€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์ถœ๋ ฅ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ณ  ๊ฐœ์„  ๋ฐฉ๋ฒ•์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ”ผ๋“œ๋ฐฑ๊ณผ ๊ฐœ์„ ์„ ํ†ตํ•ด ์ถœ๋ ฅ์ด ๊ฐœ์„ ๋ฉ๋‹ˆ๋‹ค. ์ด ์ถœ๋ ฅ-ํ”ผ๋“œ๋ฐฑ-๊ฐœ์„  ํ”„๋กœ์„ธ์Šค๋Š” ์›ํ•˜๋Š” ์กฐ๊ฑด์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณต๋ฉ๋‹ˆ๋‹ค. ์…€ํ”„์ฒดํฌ[63]๋ฅผ ํ†ตํ•ด ์—์ด์ „ํŠธ๋Š” ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์ƒ์„ฑ๋œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•˜์—ฌ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. InterAct [64]๋Š” ๋‹ค๋ฅธ ์–ธ์–ด ๋ชจ๋ธ(์˜ˆ: ChatGPT ๋ฐ InstructGPT)์„ ๊ฒ€์‚ฌ์ž ๋ฐ ์ •๋ ฌ์ž์™€ ๊ฐ™์€ ๋ณด์กฐ ์—ญํ• ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž˜๋ชป๋œ ๋™์ž‘๊ณผ ๋น„ํšจ์œจ์ ์ธ ๋™์ž‘์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋„๋ก ๋•์Šต๋‹ˆ๋‹ค. ChatCoT [65]๋Š” ๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ์„ ํ™œ์šฉํ•˜์—ฌ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค์˜ ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ”ผ๋“œ๋ฐฑ์€ ์—์ด์ „ํŠธ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋Š” ํ‰๊ฐ€ ๋ชจ๋“ˆ์—์„œ ์ƒ์„ฑ๋ฉ๋‹ˆ๋‹ค. Reflexion [12]์€ ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ์—์ด์ „ํŠธ์˜ ๊ณ„ํš ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์—์„œ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์•ก์…˜์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ, ํ‰๊ฐ€๊ธฐ๋Š” ์—์ด์ „ํŠธ ๊ถค์ ์„ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์ด ์Šค์นผ๋ผ ๊ฐ’์œผ๋กœ ์ œ๊ณต๋˜๋Š” ๊ธฐ์กด ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, ์ด ๋ชจ๋ธ์€ LLM์„ ํ™œ์šฉํ•˜์—ฌ ๋ณด๋‹ค ์ƒ์„ธํ•œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•จ์œผ๋กœ์จ ์—์ด์ „ํŠธ ๊ณ„ํš์— ๋Œ€ํ•œ ๋ณด๋‹ค ํฌ๊ด„์ ์ธ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Lazaridou et al. (2021); Jin et al. (2021) suggests implementing prior Continual Learning (CL) methods (Sun et al., 2020; d'Autume et al., 2019) to address this problem. However, it is important to note that there are nontrivial differences between traditional CL and the proposed Continual Knowledge Learning (CKL) formulation which make applying traditional CL methods inadequate. In traditional CL, methods can be largely categorized into _regularization_, _rehearsal_, and _parameter-expansion_ methods. (1) While regularization methods (Kirkpatrick et al., 2017) require identifying important parameters used for previous tasks, exactly how and where the knowledge is stored in the parameters of an LM is currently extremely difficult to identify and localize (Vig et al., 2020; De Cao et al., 2021). (2) While prior rehearsal methods (Lopez-Paz and Ranzato, 2017) consider learning all of the streams of tasks at once (multi-task learning) as the performance upper-bound and replicate such a setting with samples stored in the episodic memory, a few samples from the pretraining corpus cannot represent the overall world knowledge from the corpus. Moreover, if LMs are pretrained on a shuffled concatenation of stream of corpora, there is no guarantee that the LMs will acquire the correct, recent information from the recent corpora, especially in cases where the former corpora are much bigger than the latter ones, which is shown by experiments in Section 5.1. (3) Lastly, prior parameter-expansion methods (Rusu et al., 2016; Yoon et al., 2018) focus on _learning a stream of different tasks via strong supervision_, while in CKL, the focus is _constantly updating world knowledge from a stream of corpora via self-supervision_.
Lazaridou ๋“ฑ(2021); Jin ๋“ฑ(2021)์€ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ Continual Learning (CL) ๋ฐฉ๋ฒ•(Sun ๋“ฑ, 2020; d'Autume ๋“ฑ, 2019)์„ ๊ตฌํ˜„ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด CL๊ณผ ์ œ์•ˆ๋œ Continual Knowledge Learning (CKL) ๊ณต์‹ ์‚ฌ์ด์—๋Š” ์‚ฌ์†Œํ•˜์ง€ ์•Š์€ ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด ๊ธฐ์กด CL ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋ถ€์ ์ ˆํ•˜๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL์—์„œ ๋ฐฉ๋ฒ•์€ ํฌ๊ฒŒ _regularization_, _rehearsal_, _parameter-expansion_ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) Regularization ๋ฐฉ๋ฒ•(Kirkpatrick et al., 2017)์€ ์ด์ „ ์ž‘์—…์— ์‚ฌ์šฉ๋œ ์ค‘์š”ํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์‹๋ณ„ํ•ด์•ผ ํ•˜๋Š”๋ฐ, ํ˜„์žฌ LM์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ง€์‹์ด ์–ด๋–ป๊ฒŒ, ์–ด๋””์— ์ €์žฅ๋˜๋Š”์ง€๋Š” ์‹๋ณ„ ๋ฐ ๋กœ์ปฌ๋ผ์ด์ง•ํ•˜๊ธฐ๊ฐ€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค(Vig et al., 2020; De Cao et al., 2021). (2) ์ด์ „ Rehearsal ๋ฐฉ๋ฒ•(Lopez-Paz and Ranzato, 2017)์€ ๋ชจ๋“  ์ž‘์—… ์ŠคํŠธ๋ฆผ์„ ํ•œ ๋ฒˆ์— ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ์„ฑ๋Šฅ ์ƒํ•œ์œผ๋กœ ๊ฐ„์ฃผํ•˜๊ณ  ์—ํ”ผ์†Œ๋“œ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅ๋œ ์ƒ˜ํ”Œ๋กœ ์ด๋Ÿฌํ•œ ์„ค์ •์„ ๋ณต์ œํ•˜์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ๋ช‡ ๊ฐ€์ง€ ์ƒ˜ํ”Œ์€ ์ฝ”ํผ์Šค์˜ ์ „๋ฐ˜์ ์ธ ์„ธ๊ณ„ ์ง€์‹์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LM์ด ์ŠคํŠธ๋ฆผ์˜ ์ฝ”ํผ์Šค๋ฅผ ์„ž์–ด์„œ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ, ํŠนํžˆ ์ด์ „ ์ฝ”ํผ์Šค๊ฐ€ ํ›„์ž์˜ ์ฝ”ํผ์Šค๋ณด๋‹ค ํ›จ์”ฌ ํฐ ๊ฒฝ์šฐ, ์„น์…˜ 5.1์˜ ์‹คํ—˜์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด LM์ด ์ตœ๊ทผ ์ฝ”ํผ์Šค์—์„œ ์˜ฌ๋ฐ”๋ฅธ ์ตœ๊ทผ ์ •๋ณด๋ฅผ ์Šต๋“ํ•  ๊ฒƒ์ด๋ผ๋Š” ๋ณด์žฅ์ด ์—†์Šต๋‹ˆ๋‹ค. (3) ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด์ „ Parameter-expansion ๋ฐฉ๋ฒ•(Rusu et al., 2016; Yoon et al., 2018)์€ _๊ฐ•๋ ฅํ•œ ๊ฐ๋…์„ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ์ŠคํŠธ๋ฆผ์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๋Š” ๋ฐ˜๋ฉด, CKL์—์„œ๋Š” _self-supervision์„ ํ†ตํ•ด ์ฝ”ํผ์Šค ์ŠคํŠธ๋ฆผ์—์„œ ์„ธ๊ณ„ ์ง€์‹์„ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
Lazaridou ๋“ฑ(2021); Jin ๋“ฑ(2021)์€ ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์ „ ์—ฐ์† ํ•™์Šต(CL) ๋ฐฉ๋ฒ•(Sun ๋“ฑ, 2020; d'Autume ๋“ฑ, 2019)์„ ๊ตฌํ˜„ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด CL๊ณผ ์ œ์•ˆ๋œ ์ง€์†์  ์ง€์‹ ํ•™์Šต(CKL) ๊ณต์‹ ์‚ฌ์ด์—๋Š” ์‚ฌ์†Œํ•˜์ง€ ์•Š์€ ์ฐจ์ด๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋กœ ์ธํ•ด ๊ธฐ์กด CL ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋ถ€์ ์ ˆํ•˜๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด CL์—์„œ ๋ฐฉ๋ฒ•์€ ํฌ๊ฒŒ _์ •๊ทœํ™”_, _๋ฆฌํ—ˆ์„ค_, _๋งค๊ฐœ๋ณ€์ˆ˜ ํ™•์žฅ_ ๋ฐฉ๋ฒ•์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (1) ์ •๊ทœํ™” ๋ฐฉ๋ฒ•(Kirkpatrick et al., 2017)์€ ์ด์ „ ์ž‘์—…์— ์‚ฌ์šฉ๋œ ์ค‘์š”ํ•œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์‹๋ณ„ํ•ด์•ผ ํ•˜๋Š”๋ฐ, ํ˜„์žฌ LM์˜ ๋งค๊ฐœ๋ณ€์ˆ˜์— ์ง€์‹์ด ์–ด๋–ป๊ฒŒ, ์–ด๋””์— ์ €์žฅ๋˜๋Š”์ง€๋Š” ์‹๋ณ„ ๋ฐ ๋กœ์ปฌ๋ผ์ด์ง•ํ•˜๊ธฐ๊ฐ€ ๋งค์šฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค(Vig et al., 2020; De Cao et al., 2021). (2) ์ด์ „ ๋ฆฌํ—ˆ์„ค ๋ฐฉ๋ฒ•(Lopez-Paz and Ranzato, 2017)์€ ๋ชจ๋“  ์ž‘์—… ์ŠคํŠธ๋ฆผ์„ ํ•œ ๋ฒˆ์— ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ์„ฑ๋Šฅ ์ƒํ•œ์œผ๋กœ ๊ฐ„์ฃผํ•˜๊ณ  ์—ํ”ผ์†Œ๋“œ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅ๋œ ์ƒ˜ํ”Œ๋กœ ์ด๋Ÿฌํ•œ ์„ค์ •์„ ๋ณต์ œํ•˜์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ฝ”ํผ์Šค์˜ ๋ช‡ ๊ฐ€์ง€ ์ƒ˜ํ”Œ์€ ์ฝ”ํผ์Šค์˜ ์ „๋ฐ˜์ ์ธ ์„ธ๊ณ„ ์ง€์‹์„ ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ LM์ด ์ŠคํŠธ๋ฆผ์˜ ์ฝ”ํผ์Šค๋ฅผ ์„ž์–ด์„œ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒฝ์šฐ, ํŠนํžˆ ์ด์ „ ์ฝ”ํผ์Šค๊ฐ€ ํ›„์ž์˜ ์ฝ”ํผ์Šค๋ณด๋‹ค ํ›จ์”ฌ ํฐ ๊ฒฝ์šฐ, ์„น์…˜ 5.1์˜ ์‹คํ—˜์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด LM์ด ์ตœ๊ทผ ์ฝ”ํผ์Šค์—์„œ ์˜ฌ๋ฐ”๋ฅธ ์ตœ๊ทผ ์ •๋ณด๋ฅผ ์Šต๋“ํ•  ๊ฒƒ์ด๋ผ๋Š” ๋ณด์žฅ์ด ์—†์Šต๋‹ˆ๋‹ค. (3) ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด์ „ ๋งค๊ฐœ๋ณ€์ˆ˜ ํ™•์žฅ ๋ฐฉ๋ฒ•(Rusu et al., 2016; Yoon et al., 2018)์€ _๊ฐ•๋ ฅํ•œ ๊ฐ๋…์„ ํ†ตํ•ด ๋‹ค์–‘ํ•œ ์ž‘์—…์˜ ์ŠคํŠธ๋ฆผ์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๋Š” ๋ฐ˜๋ฉด, CKL์—์„œ๋Š” _์ž๊ธฐ ๊ฐ๋…์„ ํ†ตํ•ด ์ฝ”ํผ์Šค ์ŠคํŠธ๋ฆผ์—์„œ ์„ธ๊ณ„ ์ง€์‹์„ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค.
RAGAS (Es et al., 2023) employs the WikiEval Dataset to assess the faithfulness, answer relevance, and context relevance of RALMs. Faithfulness is defined as the degree to which responses align with the provided context. Answer relevance refers to the extent to which generated responses address the actual question posed. Context relevance is gauged by the degree to which retrieved context is centralized and devoid of irrelevant information. Additionally, the researchers utilize the prompt gpt-3.5-turbo-16k model to automate the evaluation process. RGB (Chen et al., 2024) developed a bilingual Chinese and English evaluation system employing three evaluation metrics: accuracy, rejection rate, and error detection rate. These metrics were utilized to assess the noise robustness, negative rejection, information integration, and counterfactual robustness of the data sources, which were articles processed by LM and retrieved through Google's API. CRUD-RAG (Lyu et al., 2024) considers the impact of retrieval components and the construction of external knowledge bases that have not been previously considered by researchers. A dataset was generated using a large model to evaluate the Create, Read, Update, and Delete (summarization) capabilities of RALM through four evaluation metrics: ROUGE, BLEU, bertScore, and RAGQuestEval. In addition, ARES (Saad-Falcon et al., 2023) employs datasets generated by the LM, but utilizes a lightweight LM to determine the quality of individual RALM components and utilizes human-labeled data points for prediction-powered inference. The RALM's context is evaluated using the KILT and SuperGLUE benchmarks, with Relevance, Answer Faithfulness, and Answer Relevance being the relevant criteria.
RAGAS(Es et al., 2023)๋Š” RALM์˜ faithfulness, answer relevance ๋ฐ context relevance๋ฅผ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด WikiEval ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. Faithfulness๋Š” ์ œ๊ณต๋œ context์™€ ์‘๋‹ต์ด ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Answer relevance๋Š” ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์ด ์‹ค์ œ ์งˆ๋ฌธ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. Context relevance๋Š” ๊ฒ€์ƒ‰๋œ context๊ฐ€ ์ค‘์•™ ์ง‘์ค‘์‹์ด๊ณ  ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์—†๋Š” ์ •๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด prompt gpt-3.5-turbo-16k ๋ชจ๋ธ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. RGB(Chen et al., 2024)๋Š” accuracy, rejection rate, error detection rate์˜ ์„ธ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์™€ ์˜์–ด์˜ ์ด์ค‘ ์–ธ์–ด ํ‰๊ฐ€ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๋Š” LM์— ์˜ํ•ด ์ฒ˜๋ฆฌ๋œ ๊ธฐ์‚ฌ์™€ Google์˜ API๋ฅผ ํ†ตํ•ด ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ noise robustness, negative rejection, information integration ๋ฐ counterfactual robustness๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CRUD-RAG(Lyu et al., 2024)๋Š” ์ด์ „์— ์—ฐ๊ตฌ์ž๋“ค์ด ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜๋˜ retrieval ๊ตฌ์„ฑ ์š”์†Œ์˜ ์˜ํ–ฅ๊ณผ ์™ธ๋ถ€ ์ง€์‹ ๊ธฐ๋ฐ˜ ๊ตฌ์ถ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ROUGE, BLEU, bertScore, RAGQuestEval์˜ ๋„ค ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด RALM์˜ Create, Read, Update ๋ฐ Delete(์š”์•ฝ) ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ARES(Saad-Falcon et al., 2023)๋Š” LM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, lightweight LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ RALM ๊ตฌ์„ฑ ์š”์†Œ์˜ ํ’ˆ์งˆ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์˜ˆ์ธก ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์œ„ํ•ด ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. Relevance, Answer Faithfulness, Answer Relevance๊ฐ€ ๊ด€๋ จ ๊ธฐ์ค€์ธ KILT ๋ฐ SuperGLUE ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ RALM์˜ context๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
RAGAS(Es et al., 2023)๋Š” RALM์˜ ์ถฉ์‹ค๋„, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ ๋ฐ ๋ฌธ๋งฅ ๊ด€๋ จ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด WikiEval ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ถฉ์‹ค๋„๋Š” ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ์™€ ์‘๋‹ต์ด ์–ผ๋งˆ๋‚˜ ์ผ์น˜ํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ์€ ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์ด ์‹ค์ œ ์งˆ๋ฌธ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ถ€ํ•ฉํ•˜๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋ฌธ๋งฅ ๊ด€๋ จ์„ฑ์€ ๊ฒ€์ƒ‰๋œ ๋ฌธ๋งฅ์ด ์ค‘์•™ ์ง‘์ค‘์‹์ด๊ณ  ๊ด€๋ จ ์—†๋Š” ์ •๋ณด๊ฐ€ ์—†๋Š” ์ •๋„๋กœ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ ํ‰๊ฐ€ ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ gpt-3.5-turbo-16k ๋ชจ๋ธ์„ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. RGB(Chen et al., 2024)๋Š” ์ •ํ™•๋„, ๊ฑฐ๋ถ€์œจ, ์˜ค๋ฅ˜ ๊ฐ์ง€์œจ์˜ ์„ธ ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ค‘๊ตญ์–ด์™€ ์˜์–ด์˜ ์ด์ค‘ ์–ธ์–ด ํ‰๊ฐ€ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง€ํ‘œ๋Š” LM์— ์˜ํ•ด ์ฒ˜๋ฆฌ๋œ ๊ธฐ์‚ฌ์™€ Google์˜ API๋ฅผ ํ†ตํ•ด ๊ฒ€์ƒ‰๋œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ๋…ธ์ด์ฆˆ ๊ฒฌ๊ณ ์„ฑ, ๋ถ€์ • ๊ฑฐ๋ถ€, ์ •๋ณด ํ†ตํ•ฉ ๋ฐ ๋ฐ˜์‚ฌ์‹ค์  ๊ฒฌ๊ณ ์„ฑ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. CRUD-RAG(Lyu et al., 2024)๋Š” ์ด์ „์— ์—ฐ๊ตฌ์ž๋“ค์ด ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜๋˜ ๊ฒ€์ƒ‰ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์˜ํ–ฅ๊ณผ ์™ธ๋ถ€ ์ง€์‹ ๊ธฐ๋ฐ˜ ๊ตฌ์ถ•์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ROUGE, BLEU, bertScore, RAGQuestEval์˜ ๋„ค ๊ฐ€์ง€ ํ‰๊ฐ€ ์ง€ํ‘œ๋ฅผ ํ†ตํ•ด RALM์˜ ์ƒ์„ฑ, ์ฝ๊ธฐ, ์—…๋ฐ์ดํŠธ ๋ฐ ์‚ญ์ œ(์š”์•ฝ) ๊ธฐ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ARES(Saad-Falcon et al., 2023)๋Š” LM์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๊ฒฝ๋Ÿ‰ LM์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐœ๋ณ„ RALM ๊ตฌ์„ฑ ์š”์†Œ์˜ ํ’ˆ์งˆ์„ ๊ฒฐ์ •ํ•˜๊ณ  ์˜ˆ์ธก ๊ธฐ๋ฐ˜ ์ถ”๋ก ์„ ์œ„ํ•ด ์ธ๊ฐ„์ด ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ์„ฑ, ๋‹ต๋ณ€ ์‹ ๋ขฐ์„ฑ, ๋‹ต๋ณ€ ๊ด€๋ จ์„ฑ์ด ๊ด€๋ จ ๊ธฐ์ค€์ธ KILT ๋ฐ SuperGLUE ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ RALM์˜ ์ปจํ…์ŠคํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
**Fairness and Bias.** The evaluation protocol presented in this paper is limited in terms of its ability to capture potential issues related to fairness and bias, which remains an important open question that we will aim to address in subsequent system evaluations. Recent advances in the development of comprehensive frameworks for bias detection in large language models [94, 95] present a promising starting point for establishing such an approach. It should be noted that medical diagnostic dialogue is a particularly challenging use case, due to the complexity of the medical domain, the interactive information gathering nature of the dialogue, and the outcome-driven setting, with the potential of associated harms in case of incorrect diagnosis or incorrect medical advice. Nevertheless, disentangling these issues is an important further research area if LLMs in the domain are to overcome rather than propagate inequities in healthcare. For example, previous studies have found that physicians approach communication with their patients differently, on average, depending on patients' race, resulting in Black patients receiving communication that was less patient-centered, and with a lower positive affect [96]. Other studies have found differences in physicians' communication styles and conversation length based on gender [97]. Effective intercultural communication skills are essential [91]. There is therefore a non-negligible risk that such historical conversational biases may be replicated or amplified in an AI dialogue system, but at the same time there is also an opportunity to work towards designing conversational systems that can be more inclusive, and more personalized to the individual patient's needs.
**Fairness and Bias.** ์ด ๋ฐฑ์„œ์— ์ œ์‹œ๋œ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์€ ๊ณต์ •์„ฑ ๋ฐ ํŽธํ–ฅ์„ฑ๊ณผ ๊ด€๋ จ๋œ ์ž ์žฌ์  ๋ฌธ์ œ๋ฅผ ํฌ์ฐฉํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์ œํ•œ์ ์ด๋ฉฐ, ์ด๋Š” ํ›„์† ์‹œ์Šคํ…œ ํ‰๊ฐ€์—์„œ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ์ค‘์š”ํ•œ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ bias ํƒ์ง€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ๋ฐœ์ „[94, 95]์€ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•๋ฆฝํ•˜๊ธฐ ์œ„ํ•œ ์œ ๋งํ•œ ์ถœ๋ฐœ์ ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ์ง„๋‹จ dialogue๋Š” ์˜๋ฃŒ ์˜์—ญ์˜ ๋ณต์žก์„ฑ, ๋Œ€ํ™”์˜ ์ƒํ˜ธ ์ž‘์šฉ์  ์ •๋ณด ์ˆ˜์ง‘ ํŠน์„ฑ, ๊ฒฐ๊ณผ ์ค‘์‹ฌ์ ์ธ ํ™˜๊ฒฝ์œผ๋กœ ์ธํ•ด ํŠนํžˆ ๊นŒ๋‹ค๋กœ์šด ์‚ฌ์šฉ ์‚ฌ๋ก€์ด๋ฉฐ, ์ž˜๋ชป๋œ ์ง„๋‹จ์ด๋‚˜ ์ž˜๋ชป๋œ ์˜ํ•™์  ์กฐ์–ธ์˜ ๊ฒฝ์šฐ ๊ด€๋ จ ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฃŒ ๋ถ„์•ผ์˜ LLM์ด ์˜๋ฃŒ ์„œ๋น„์Šค์˜ ๋ถˆํ‰๋“ฑ์„ ํ™•์‚ฐ์‹œํ‚ค์ง€ ์•Š๊ณ  ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ค‘์š”ํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์˜์‚ฌ๋Š” ํ™˜์ž์˜ ์ธ์ข…์— ๋”ฐ๋ผ ํ‰๊ท ์ ์œผ๋กœ ํ™˜์ž์™€์˜ ์˜์‚ฌ์†Œํ†ต์— ๋‹ค๋ฅด๊ฒŒ ์ ‘๊ทผํ•˜์—ฌ ํ‘์ธ ํ™˜์ž๋Š” ํ™˜์ž ์ค‘์‹ฌ์ ์ด๊ณ  ๊ธ์ •์ ์ธ ์˜ํ–ฅ์ด ๋‚ฎ์€ ์˜์‚ฌ์†Œํ†ต์„ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[96]. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์„ฑ๋ณ„์— ๋”ฐ๋ผ ์˜์‚ฌ์˜ ์˜์‚ฌ์†Œํ†ต ์Šคํƒ€์ผ๊ณผ ๋Œ€ํ™” ์‹œ๊ฐ„์— ์ฐจ์ด๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[97]. ํšจ๊ณผ์ ์ธ ๋ฌธํ™” ๊ฐ„ ์˜์‚ฌ์†Œํ†ต ๊ธฐ์ˆ ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค[91]. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ์—ญ์‚ฌ์  ๋Œ€ํ™” ํŽธ๊ฒฌ์ด AI dialogue ์‹œ์Šคํ…œ์—์„œ ๋ณต์ œ๋˜๊ฑฐ๋‚˜ ์ฆํญ๋  ์œ„ํ—˜์€ ๋ฌด์‹œํ•  ์ˆ˜ ์—†์ง€๋งŒ, ๋™์‹œ์— ๋ณด๋‹ค ํฌ์šฉ์ ์ด๊ณ  ๊ฐœ๋ณ„ ํ™˜์ž์˜ ํ•„์š”์— ๋งž๋Š” ๋Œ€ํ™” ์‹œ์Šคํ…œ์„ ์„ค๊ณ„ํ•˜๊ธฐ ์œ„ํ•œ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์ผ ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.
**๊ณต์ •์„ฑ ๋ฐ ํŽธํ–ฅ์„ฑ.** ์ด ๋ฐฑ์„œ์— ์ œ์‹œ๋œ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์€ ๊ณต์ •์„ฑ ๋ฐ ํŽธํ–ฅ์„ฑ๊ณผ ๊ด€๋ จ๋œ ์ž ์žฌ์  ๋ฌธ์ œ๋ฅผ ํฌ์ฐฉํ•˜๋Š” ๋Šฅ๋ ฅ์ด ์ œํ•œ์ ์ด๋ฉฐ, ์ด๋Š” ํ›„์† ์‹œ์Šคํ…œ ํ‰๊ฐ€์—์„œ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ์ค‘์š”ํ•œ ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ํŽธํ–ฅ์„ฑ ํƒ์ง€์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ๋ฐœ์ „[94, 95]์€ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ํ™•๋ฆฝํ•˜๊ธฐ ์œ„ํ•œ ์œ ๋งํ•œ ์ถœ๋ฐœ์ ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ์ง„๋‹จ ๋Œ€ํ™”๋Š” ์˜๋ฃŒ ์˜์—ญ์˜ ๋ณต์žก์„ฑ, ๋Œ€ํ™”์˜ ์ƒํ˜ธ ์ž‘์šฉ์  ์ •๋ณด ์ˆ˜์ง‘ ํŠน์„ฑ, ๊ฒฐ๊ณผ ์ค‘์‹ฌ์ ์ธ ํ™˜๊ฒฝ์œผ๋กœ ์ธํ•ด ํŠนํžˆ ๊นŒ๋‹ค๋กœ์šด ์‚ฌ์šฉ ์‚ฌ๋ก€์ด๋ฉฐ, ์ž˜๋ชป๋œ ์ง„๋‹จ์ด๋‚˜ ์ž˜๋ชป๋œ ์˜ํ•™์  ์กฐ์–ธ์˜ ๊ฒฝ์šฐ ๊ด€๋ จ ํ”ผํ•ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์€ ์˜๋ฃŒ ๋ถ„์•ผ์˜ LLM์ด ์˜๋ฃŒ ์„œ๋น„์Šค์˜ ๋ถˆํ‰๋“ฑ์„ ํ™•์‚ฐ์‹œํ‚ค์ง€ ์•Š๊ณ  ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์ค‘์š”ํ•œ ์ถ”๊ฐ€ ์—ฐ๊ตฌ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด์ „ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ์˜์‚ฌ๋Š” ํ™˜์ž์˜ ์ธ์ข…์— ๋”ฐ๋ผ ํ‰๊ท ์ ์œผ๋กœ ํ™˜์ž์™€์˜ ์˜์‚ฌ์†Œํ†ต์— ๋‹ค๋ฅด๊ฒŒ ์ ‘๊ทผํ•˜์—ฌ ํ‘์ธ ํ™˜์ž๋Š” ํ™˜์ž ์ค‘์‹ฌ์ ์ด๊ณ  ๊ธ์ •์ ์ธ ์˜ํ–ฅ์ด ๋‚ฎ์€ ์˜์‚ฌ์†Œํ†ต์„ ๋ฐ›๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[96]. ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์„ฑ๋ณ„์— ๋”ฐ๋ผ ์˜์‚ฌ์˜ ์˜์‚ฌ์†Œํ†ต ์Šคํƒ€์ผ๊ณผ ๋Œ€ํ™” ์‹œ๊ฐ„์— ์ฐจ์ด๊ฐ€ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[97]. ํšจ๊ณผ์ ์ธ ๋ฌธํ™” ๊ฐ„ ์˜์‚ฌ์†Œํ†ต ๊ธฐ์ˆ ์€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค[91]. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ์—ญ์‚ฌ์  ๋Œ€ํ™” ํŽธ๊ฒฌ์ด AI ๋Œ€ํ™” ์‹œ์Šคํ…œ์—์„œ ๋ณต์ œ๋˜๊ฑฐ๋‚˜ ์ฆํญ๋  ์œ„ํ—˜์€ ๋ฌด์‹œํ•  ์ˆ˜ ์—†์ง€๋งŒ, ๋™์‹œ์— ๋ณด๋‹ค ํฌ์šฉ์ ์ด๊ณ  ๊ฐœ๋ณ„ ํ™˜์ž์˜ ํ•„์š”์— ๋งž๋Š” ๋Œ€ํ™” ์‹œ์Šคํ…œ์„ ์„ค๊ณ„ํ•˜๊ธฐ ์œ„ํ•œ ๋…ธ๋ ฅ์„ ๊ธฐ์šธ์ผ ์ˆ˜ ์žˆ๋Š” ๊ธฐํšŒ์ด๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.
speed up training efficiency while ensuring performance. The authors use local attention to extract local information, and then use the state-space model to extract the global information missing from local attention. [48] argues that existing State Space Models, while efficient, lack in performance. The authors believe the reason for this is that too many state transitions make the model lose shallow information. So the authors propose a design that integrates the hidden states in the previous layers into the subsequent layers to retain more shallow information. Eventually, after pre-training on Pile, the experimental results of Zero-shot and four-shot on other datasets are significantly improved. For voice tasks, Du et al. [42] combine the high efficiency of impulse neural networks and the ability to model long distances with the state-space model S4 to obtain an spiking neural network, which has a low number of parameters but comparable performance to that of some artificial neural network (ANN) in deep noise suppression task. In addition, for speech separation task, DPMamba [43] proposed by Jiang et al. uses the selective State Space Model Mamba to replace the traditional transformer architecture. DPMamba simultaneously models the short-term and long-term forward and backward dependencies of speech signals through selective state space, achieving comparable results to the dual-path Transformer model Sepformer [134]. SPAMba [44] proposed by Li et al. uses TF-GridNet [135] as the basic framework and replaces the Transformer module with a bidirectional Mamba module to capture a wider range of language information. Experimental results show that Mamba-based models play an important role in performance.
์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜๋ฉด์„œ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ๋กœ์ปฌ attention์„ ์‚ฌ์šฉํ•˜์—ฌ ๋กœ์ปฌ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ, State Space Model์„ ์‚ฌ์šฉํ•˜์—ฌ ๋กœ์ปฌ attention์—์„œ ๋ˆ„๋ฝ๋œ ๊ธ€๋กœ๋ฒŒ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. [48]์€ ๊ธฐ์กด State Space Models์ด ํšจ์œจ์ ์ด๊ธด ํ•˜์ง€๋งŒ ์„ฑ๋Šฅ์ด ๋ถ€์กฑํ•˜๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ๋„ˆ๋ฌด ๋งŽ์€ ์ƒํƒœ ์ „ํ™˜์œผ๋กœ ์ธํ•ด ๋ชจ๋ธ์ด ์–•์€ ์ •๋ณด๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์ด์œ  ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €์ž๋“ค์€ ์ด์ „ ๋ ˆ์ด์–ด์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ํ›„์† ๋ ˆ์ด์–ด์— ํ†ตํ•ฉํ•˜์—ฌ ๋” ๋งŽ์€ ์–•์€ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•˜๋Š” ์„ค๊ณ„๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ, Pile์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๊ฑฐ์นœ ํ›„ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ zero-shot๊ณผ four-shot์˜ ์‹คํ—˜ ๊ฒฐ๊ณผ๊ฐ€ ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์Œ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ, Du ๋“ฑ [42]์€ ์ž„ํŽ„์Šค ์‹ ๊ฒฝ๋ง์˜ ๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๊ธด ๊ฑฐ๋ฆฌ๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ๋Šฅ๋ ฅ์„ State Space Model S4์™€ ๊ฒฐํ•ฉํ•˜์—ฌ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๋Š” ์ ์ง€๋งŒ ๋”ฅ ๋…ธ์ด์ฆˆ ์–ต์ œ ์ž‘์—…์—์„œ ์ผ๋ถ€ ์ธ๊ณต ์‹ ๊ฒฝ๋ง(ANN)๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ spiking neural network๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์Œ์„ฑ ๋ถ„๋ฆฌ ์ž‘์—…์„ ์œ„ํ•ด Jiang ๋“ฑ์ด ์ œ์•ˆํ•œ DPMamba [43]๋Š” ์„ ํƒ์  State Space Model Mamba๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด์˜ Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. DPMamba๋Š” ์„ ํƒ์  ์ƒํƒœ ๊ณต๊ฐ„์„ ํ†ตํ•ด ์Œ์„ฑ ์‹ ํ˜ธ์˜ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ์ „์ง„ ๋ฐ ํ›„์ง„ ์ข…์†์„ฑ์„ ๋™์‹œ์— ๋ชจ๋ธ๋งํ•˜์—ฌ ์ด์ค‘ ๊ฒฝ๋กœ Transformer ๋ชจ๋ธ Sepformer [134]์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. Li ๋“ฑ์ด ์ œ์•ˆํ•œ SPAMba [44]๋Š” TF-GridNet [135]์„ ๊ธฐ๋ณธ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ ์‚ฌ์šฉํ•˜๊ณ  Transformer ๋ชจ๋“ˆ์„ ์–‘๋ฐฉํ–ฅ Mamba ๋ชจ๋“ˆ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์–ธ์–ด ์ •๋ณด๋ฅผ ์บก์ฒ˜ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, Mamba ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ์„ฑ๋Šฅ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
์„ฑ๋Šฅ์„ ๋ณด์žฅํ•˜๋ฉด์„œ ํ›ˆ๋ จ ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ๋กœ์ปฌ ์–ดํ…์…˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋กœ์ปฌ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•œ ๋‹ค์Œ, ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋กœ์ปฌ ์–ดํ…์…˜์—์„œ ๋ˆ„๋ฝ๋œ ๊ธ€๋กœ๋ฒŒ ์ •๋ณด๋ฅผ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. [48]์€ ๊ธฐ์กด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์ด ํšจ์œจ์ ์ด๊ธด ํ•˜์ง€๋งŒ ์„ฑ๋Šฅ์ด ๋ถ€์กฑํ•˜๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ๋„ˆ๋ฌด ๋งŽ์€ ์ƒํƒœ ์ „ํ™˜์œผ๋กœ ์ธํ•ด ๋ชจ๋ธ์ด ์–•์€ ์ •๋ณด๋ฅผ ์žƒ์–ด๋ฒ„๋ฆฐ๋‹ค๋Š” ์ด์œ  ๋•Œ๋ฌธ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ €์ž๋“ค์€ ์ด์ „ ๋ ˆ์ด์–ด์˜ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ํ›„์† ๋ ˆ์ด์–ด์— ํ†ตํ•ฉํ•˜์—ฌ ๋” ๋งŽ์€ ์–•์€ ์ •๋ณด๋ฅผ ์œ ์ง€ํ•˜๋Š” ์„ค๊ณ„๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ, Pile์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ๊ฑฐ์นœ ํ›„ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ œ๋กœ ์ƒท๊ณผ 4 ์ƒท์˜ ์‹คํ—˜ ๊ฒฐ๊ณผ๊ฐ€ ํฌ๊ฒŒ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์Œ์„ฑ ์ž‘์—…์˜ ๊ฒฝ์šฐ, Du ๋“ฑ [42]์€ ์ž„ํŽ„์Šค ์‹ ๊ฒฝ๋ง์˜ ๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๊ธด ๊ฑฐ๋ฆฌ๋ฅผ ๋ชจ๋ธ๋งํ•˜๋Š” ๋Šฅ๋ ฅ์„ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ S4์™€ ๊ฒฐํ•ฉํ•˜์—ฌ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๋Š” ์ ์ง€๋งŒ ๋”ฅ ๋…ธ์ด์ฆˆ ์–ต์ œ ์ž‘์—…์—์„œ ์ผ๋ถ€ ์ธ๊ณต ์‹ ๊ฒฝ๋ง(ANN)๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ ์ŠคํŒŒ์ดํ‚น ์‹ ๊ฒฝ๋ง์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ์Œ์„ฑ ๋ถ„๋ฆฌ ์ž‘์—…์„ ์œ„ํ•ด Jiang ๋“ฑ์ด ์ œ์•ˆํ•œ DPMamba [43]๋Š” ์„ ํƒ์  ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ Mamba๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ธฐ์กด์˜ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. DPMamba๋Š” ์„ ํƒ์  ์ƒํƒœ ๊ณต๊ฐ„์„ ํ†ตํ•ด ์Œ์„ฑ ์‹ ํ˜ธ์˜ ๋‹จ๊ธฐ ๋ฐ ์žฅ๊ธฐ ์ „์ง„ ๋ฐ ํ›„์ง„ ์ข…์†์„ฑ์„ ๋™์‹œ์— ๋ชจ๋ธ๋งํ•˜์—ฌ ์ด์ค‘ ๊ฒฝ๋กœ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ Sepformer [134]์™€ ๋น„์Šทํ•œ ๊ฒฐ๊ณผ๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. Li ๋“ฑ์ด ์ œ์•ˆํ•œ SPAMba [44]๋Š” TF-GridNet [135]์„ ๊ธฐ๋ณธ ํ”„๋ ˆ์ž„์›Œํฌ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋“ˆ์„ ์–‘๋ฐฉํ–ฅ ๋ง˜๋ฐ” ๋ชจ๋“ˆ๋กœ ๋Œ€์ฒดํ•˜์—ฌ ๋” ๋„“์€ ๋ฒ”์œ„์˜ ์–ธ์–ด ์ •๋ณด๋ฅผ ์บก์ฒ˜ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, ๋ง˜๋ฐ” ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ์„ฑ๋Šฅ์— ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค.
Evaluating different LLMs can be seen also from different perspectives. For example, a LLM with a drastically fewer number of parameters is not completely comparable to one with a larger number of parameters. From this perspective, we will categorize LLMs in four categories as well: **small** (less than or equal to 1 billion parameters), **medium** (between 1 and 10 billion), **large** (between 10 and 100 billion), and **very large** (more than 100 billion). Another classification for the LLMs we use is their primary use case. We consider each LLM to be either: **Foundation** model (pretrained language model with no instruction fine-tuning and chat fine-tuning), **Instruction** model (pretrained language model with only instruction fine-tuning), and **Chat** model (pretrained language model with instruction and chat fine-tuning). Apart from all the categorization described, another category is required to distinguish between original models and tuned ones. **Original** models are those that have been released as a foundation model or a fine-tuned one. **Tuned** models are those that grasped the original model and tuned it with different datasets or even different training approaches. It is also good to note that original models are usually foundation models that have been fine-tuned on specific datasets or even different approaches. Availability of the model weights regardless of the license is another category in our classification. Models that have their weights publicly available (even through request) are noted as **Public** models while others are noted as **Private**. Table III shows all of these definitions and abbreviations used in the rest of the article. Figure 43 illustrate these visually.
๋‹ค๋ฅธ LLM์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋‹ค๋ฅธ ๊ด€์ ์—์„œ๋„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ํ›จ์”ฌ ์ ์€ LLM์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋งŽ์€ LLM๊ณผ ์™„์ „ํžˆ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ ์—์„œ LLM์€ **์†Œํ˜•**(10์–ต ๊ฐœ ์ดํ•˜), **์ค‘ํ˜•**(10์–ต ๊ฐœ์—์„œ 100์–ต ๊ฐœ ์‚ฌ์ด), **๋Œ€ํ˜•**(100์–ต ๊ฐœ์—์„œ 1000์–ต ๊ฐœ ์‚ฌ์ด), **์ดˆ๋Œ€ํ˜•**(1000์–ต ๊ฐœ ์ด์ƒ)์˜ ๋„ค ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” LLM์˜ ๋˜ ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๋Š” ์ฃผ์š” ์‚ฌ์šฉ ์‚ฌ๋ก€์ž…๋‹ˆ๋‹ค. ๊ฐ LLM์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค: **Foundation** ๋ชจ๋ธ(Instruction Fine-tuning ๋ฐ ์ฑ„ํŒ… Fine-tuning์ด ์—†๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ), **Instruction** ๋ชจ๋ธ(Instruction Fine-tuning๋งŒ ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ), **Chat** ๋ชจ๋ธ(Instruction ๋ฐ ์ฑ„ํŒ… Fine-tuning์ด ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ)์ž…๋‹ˆ๋‹ค. ์•ž์„œ ์„ค๋ช…ํ•œ ๋ชจ๋“  ๋ถ„๋ฅ˜ ์™ธ์—๋„ ์›๋ณธ ๋ชจ๋ธ๊ณผ ํŠœ๋‹๋œ ๋ชจ๋ธ์„ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์œ„ํ•ด ๋˜ ๋‹ค๋ฅธ ์นดํ…Œ๊ณ ๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. **Original** ๋ชจ๋ธ์€ Foundation ๋ชจ๋ธ ๋˜๋Š” Fine-tuned ๋ชจ๋ธ๋กœ ์ถœ์‹œ๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. **Tuned** ๋ชจ๋ธ์€ ์›๋ณธ ๋ชจ๋ธ์„ ์บก์ฒ˜ํ•˜์—ฌ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ๋‹ค๋ฅธ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์กฐ์ •ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์›๋ณธ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด Fine-tuned๋œ Foundation ๋ชจ๋ธ์ด๋ผ๋Š” ์ ๋„ ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ผ์ด์„ ์Šค์— ๊ด€๊ณ„์—†์ด ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์˜ ์‚ฌ์šฉ ๊ฐ€๋Šฅ ์—ฌ๋ถ€๋Š” ๋ถ„๋ฅ˜์˜ ๋˜ ๋‹ค๋ฅธ ๋ฒ”์ฃผ์ž…๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜๋ฅผ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ(์š”์ฒญ์„ ํ†ตํ•ด์„œ๋ผ๋„)์€ **Public** ๋ชจ๋ธ๋กœ ํ‘œ์‹œ๋˜๊ณ , ๊ทธ ์™ธ์˜ ๋ชจ๋ธ์€ **Private** ๋ชจ๋ธ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ํ‘œ III๋Š” ๋‚˜๋จธ์ง€ ๋ฌธ์„œ์—์„œ ์‚ฌ์šฉ๋˜๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์ •์˜์™€ ์•ฝ์–ด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 43์€ ์ด๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋‹ค๋ฅธ LLM์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ๋‹ค๋ฅธ ๊ด€์ ์—์„œ๋„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ํ›จ์”ฌ ์ ์€ LLM์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ๋งŽ์€ LLM๊ณผ ์™„์ „ํžˆ ๋น„๊ตํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ ์—์„œ LLM์€ **์†Œํ˜•**(10์–ต ๊ฐœ ์ดํ•˜), **์ค‘ํ˜•**(10์–ต ๊ฐœ์—์„œ 100์–ต ๊ฐœ ์‚ฌ์ด), **๋Œ€ํ˜•**(100์–ต ๊ฐœ์—์„œ 1000์–ต ๊ฐœ ์‚ฌ์ด), **์ดˆ๋Œ€ํ˜•**(1000์–ต ๊ฐœ ์ด์ƒ)์˜ ๋„ค ๊ฐ€์ง€ ๋ฒ”์ฃผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” LLM์˜ ๋˜ ๋‹ค๋ฅธ ๋ถ„๋ฅ˜๋Š” ์ฃผ์š” ์‚ฌ์šฉ ์‚ฌ๋ก€์ž…๋‹ˆ๋‹ค. ๊ฐ LLM์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค: **๊ธฐ๋ฐ˜** ๋ชจ๋ธ(๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ • ๋ฐ ์ฑ„ํŒ… ๋ฏธ์„ธ ์กฐ์ •์ด ์—†๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ), **๋ช…๋ น์–ด** ๋ชจ๋ธ(๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •๋งŒ ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ), **์ฑ„ํŒ…** ๋ชจ๋ธ(๋ช…๋ น์–ด ๋ฐ ์ฑ„ํŒ… ๋ฏธ์„ธ ์กฐ์ •์ด ์žˆ๋Š” ์‚ฌ์ „ ํ•™์Šต ์–ธ์–ด ๋ชจ๋ธ)์ž…๋‹ˆ๋‹ค. ์•ž์„œ ์„ค๋ช…ํ•œ ๋ชจ๋“  ๋ถ„๋ฅ˜ ์™ธ์—๋„ ์›๋ณธ ๋ชจ๋ธ๊ณผ ํŠœ๋‹๋œ ๋ชจ๋ธ์„ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์œ„ํ•ด ๋˜ ๋‹ค๋ฅธ ์นดํ…Œ๊ณ ๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. **์›๋ณธ** ๋ชจ๋ธ์€ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๋˜๋Š” ๋ฏธ์„ธ ์กฐ์ •๋œ ๋ชจ๋ธ๋กœ ์ถœ์‹œ๋œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. **ํŠœ๋‹๋œ** ๋ชจ๋ธ์€ ์›๋ณธ ๋ชจ๋ธ์„ ์บก์ฒ˜ํ•˜์—ฌ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ๋‹ค๋ฅธ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ์กฐ์ •ํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ์›๋ณธ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํŠน์ • ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋˜๋Š” ๋‹ค๋ฅธ ์ ‘๊ทผ ๋ฐฉ์‹์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋œ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด๋ผ๋Š” ์ ๋„ ์œ ์˜ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ผ์ด์„ ์Šค์— ๊ด€๊ณ„์—†์ด ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์˜ ์‚ฌ์šฉ ๊ฐ€๋Šฅ ์—ฌ๋ถ€๋Š” ๋ถ„๋ฅ˜์˜ ๋˜ ๋‹ค๋ฅธ ๋ฒ”์ฃผ์ž…๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜๋ฅผ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ(์š”์ฒญ์„ ํ†ตํ•ด์„œ๋ผ๋„)์€ **๊ณต๊ฐœ** ๋ชจ๋ธ๋กœ ํ‘œ์‹œ๋˜๊ณ , ๊ทธ ์™ธ์˜ ๋ชจ๋ธ์€ **๋น„๊ณต๊ฐœ** ๋ชจ๋ธ๋กœ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค. ํ‘œ III๋Š” ๋‚˜๋จธ์ง€ ๋ฌธ์„œ์—์„œ ์‚ฌ์šฉ๋˜๋Š” ์ด๋Ÿฌํ•œ ๋ชจ๋“  ์ •์˜์™€ ์•ฝ์–ด๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 43์€ ์ด๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
**Model Training.** For pre-training, it is essential to establish a data-centric infrastructure and training procedure for LLM optimization, which can effectively support a systematic process of data collection, data cleaning, data mixture, and data curriculum. Furthermore, it also calls for more flexible mechanisms of hardware support or resource schedule, so as to better organize and utilize the resources in a computing cluster. In practice, it is very challenging to pre-train capable LLMs, due to the huge compute consumption and the sensitivity to data quality and training tricks [78, 93]. Thus, it becomes particularly important to develop systemic, economical pre-training approaches for optimizing LLMs, _e.g._, predictable scaling [46] and proxy model training [59]. More training recipes or principles should be investigated and shared to reduce the potential risk of degradation or failure in large-scale model optimization. Although increasingly more model checkpoints and cleaned datasets have been released, there still lacks reproducible work on pre-training data preparation (_e.g._, detailed cleaning strategies) and data scheduling (_e.g._, data mixture and curriculum). Since it is very costly to pre-train a LLM from scratch, it is important to design suitable mechanisms for continually pre-training or fine-tuning the LLM based on publicly available model checkpoints (_e.g._, LLaMA [57] and Flan-T5 [69]). For this purpose, a number of technical issues have to be resolved, _e.g._, catastrophic forgetting and task specialization. Furthermore, it is also useful to develop effective tuning strategies that effectively inject or edit specific knowledge [672], _e.g._, correcting the outdated facts.
**Model Training**. ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•ด์„œ๋Š” ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  LLM optimization์„ ์œ„ํ•œ ํ›ˆ๋ จ ์ ˆ์ฐจ๋ฅผ ์ˆ˜๋ฆฝํ•ด์•ผ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ๋ฐ์ดํ„ฐ ์ •๋ฆฌ, ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ, ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์˜ ์ฒด๊ณ„์ ์ธ ํ”„๋กœ์„ธ์Šค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•˜๋“œ์›จ์–ด ์ง€์› ๋˜๋Š” ๋ฆฌ์†Œ์Šค ์Šค์ผ€์ค„๋ง์˜ ๋ณด๋‹ค ์œ ์—ฐํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ์ปดํ“จํŒ… ํด๋Ÿฌ์Šคํ„ฐ์˜ ๋ฆฌ์†Œ์Šค๋ฅผ ๋”์šฑ ํšจ์œจ์ ์œผ๋กœ ๊ตฌ์„ฑํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ๋ง‰๋Œ€ํ•œ ์—ฐ์‚ฐ ์†Œ๋น„์™€ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ ๋ฐ ํ›ˆ๋ จ ํŠธ๋ฆญ์— ๋Œ€ํ•œ ๋ฏผ๊ฐ์„ฑ์œผ๋กœ ์ธํ•ด ๋Šฅ๋ ฅ ์žˆ๋Š” LLM์„ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ค์šด ์ผ์ž…๋‹ˆ๋‹ค[78, 93]. ๋”ฐ๋ผ์„œ ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ scaling[46] ๋ฐ proxy model training[59]๊ณผ ๊ฐ™์ด LLM์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ฒด๊ณ„์ ์ด๊ณ  ๊ฒฝ์ œ์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ์ตœ์ ํ™”์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ๋˜๋Š” ์‹คํŒจ ๊ฐ€๋Šฅ์„ฑ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ํ›ˆ๋ จ ๋ ˆ์‹œํ”ผ ๋˜๋Š” ์›์น™์„ ์กฐ์‚ฌํ•˜๊ณ  ๊ณต์œ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ ์  ๋” ๋งŽ์€ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์™€ ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณต๊ฐœ๋˜๊ณ  ์žˆ์ง€๋งŒ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ค€๋น„(_์˜ˆ: ์„ธ๋ถ€์ ์ธ ์ •๋ฆฌ ์ „๋žต) ๋ฐ ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„๋ง(_์˜ˆ: ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฐ ์ปค๋ฆฌํ˜๋Ÿผ)์— ๋Œ€ํ•œ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์ž‘์—…์€ ์—ฌ์ „ํžˆ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ(_์˜ˆ: LLaMA [57] ๋ฐ Flan-T5 [69])๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๊ฑฐ๋‚˜ Fine-tuningํ•˜๊ธฐ ์œ„ํ•œ ์ ์ ˆํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ์น˜๋ช…์ ์ธ ๋ง๊ฐ๊ณผ ์ž‘์—… ์ „๋ฌธํ™” ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ˆ ์  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฃผ์ž…ํ•˜๊ฑฐ๋‚˜ ํŽธ์ง‘ํ•˜๋Š” ํšจ๊ณผ์ ์ธ tuning ์ „๋žต(_์˜ˆ: ์˜ค๋ž˜๋œ ์‚ฌ์‹ค ์ˆ˜์ •)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค[672].
**๋ชจ๋ธ ํ›ˆ๋ จ**. ์‚ฌ์ „ ํ›ˆ๋ จ์„ ์œ„ํ•ด์„œ๋Š” ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  LLM ์ตœ์ ํ™”๋ฅผ ์œ„ํ•œ ํ›ˆ๋ จ ์ ˆ์ฐจ๋ฅผ ์ˆ˜๋ฆฝํ•ด์•ผ ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘, ๋ฐ์ดํ„ฐ ์ •๋ฆฌ, ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ, ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์˜ ์ฒด๊ณ„์ ์ธ ํ”„๋กœ์„ธ์Šค๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ํ•˜๋“œ์›จ์–ด ์ง€์› ๋˜๋Š” ๋ฆฌ์†Œ์Šค ์Šค์ผ€์ค„๋ง์˜ ๋ณด๋‹ค ์œ ์—ฐํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์ด ํ•„์š”ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ์ปดํ“จํŒ… ํด๋Ÿฌ์Šคํ„ฐ์˜ ๋ฆฌ์†Œ์Šค๋ฅผ ๋”์šฑ ํšจ์œจ์ ์œผ๋กœ ๊ตฌ์„ฑํ•˜๊ณ  ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ๋ง‰๋Œ€ํ•œ ์—ฐ์‚ฐ ์†Œ๋น„์™€ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ ๋ฐ ํ›ˆ๋ จ ํŠธ๋ฆญ์— ๋Œ€ํ•œ ๋ฏผ๊ฐ์„ฑ์œผ๋กœ ์ธํ•ด ๋Šฅ๋ ฅ ์žˆ๋Š” LLM์„ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ค์šด ์ผ์ž…๋‹ˆ๋‹ค[78, 93]. ๋”ฐ๋ผ์„œ ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ์Šค์ผ€์ผ๋ง[46] ๋ฐ ํ”„๋ก์‹œ ๋ชจ๋ธ ํ›ˆ๋ จ[59]๊ณผ ๊ฐ™์ด LLM์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ์ฒด๊ณ„์ ์ด๊ณ  ๊ฒฝ์ œ์ ์ธ ์‚ฌ์ „ ํ›ˆ๋ จ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ํŠนํžˆ ์ค‘์š”ํ•ด์กŒ์Šต๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ์ตœ์ ํ™”์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ๋˜๋Š” ์‹คํŒจ ๊ฐ€๋Šฅ์„ฑ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋” ๋งŽ์€ ํ›ˆ๋ จ ๋ ˆ์‹œํ”ผ ๋˜๋Š” ์›์น™์„ ์กฐ์‚ฌํ•˜๊ณ  ๊ณต์œ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ ์  ๋” ๋งŽ์€ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ์™€ ์ •๋ฆฌ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ๊ณต๊ฐœ๋˜๊ณ  ์žˆ์ง€๋งŒ, ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ์ค€๋น„(_์˜ˆ: ์„ธ๋ถ€์ ์ธ ์ •๋ฆฌ ์ „๋žต) ๋ฐ ๋ฐ์ดํ„ฐ ์Šค์ผ€์ค„๋ง(_์˜ˆ: ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ ๋ฐ ์ปค๋ฆฌํ˜๋Ÿผ)์— ๋Œ€ํ•œ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์ž‘์—…์€ ์—ฌ์ „ํžˆ ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. LLM์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ(_์˜ˆ: LLaMA [57] ๋ฐ Flan-T5 [69])๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ›ˆ๋ จํ•˜๊ฑฐ๋‚˜ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ ์ ˆํ•œ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ์น˜๋ช…์ ์ธ ๋ง๊ฐ๊ณผ ์ž‘์—… ์ „๋ฌธํ™” ๋“ฑ ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ˆ ์  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ํŠน์ • ์ง€์‹์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฃผ์ž…ํ•˜๊ฑฐ๋‚˜ ํŽธ์ง‘ํ•˜๋Š” ํšจ๊ณผ์ ์ธ ํŠœ๋‹ ์ „๋žต(_์˜ˆ: ์˜ค๋ž˜๋œ ์‚ฌ์‹ค ์ˆ˜์ •)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ๋„ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค[672].
We introduce the Yi model family, a series of language and multimodal models that demonstrate strong multi-dimensional capabilities. The Yi model family is based on 6B and 34B pretrained language models, then we extend them to chat models, 200K long context models, depth-upscaled models, and vision-language models. Our base models achieve strong performance on a wide range of benchmarks like MMLU, and our finetuned chat models deliver strong human preference rate on major evaluation platforms like AlpacaEval and Chatbot Arena. Building upon our scalable super-computing infrastructure and the classical transformer architecture, we attribute the performance of Yi models primarily to its data quality resulting from our data-engineering efforts. For pretraining, we construct 3.1 trillion tokens of English and Chinese corpora using a cascaded data deduplication and quality filtering pipeline. For finetuning, we polish a small scale (less than 10K) instruction dataset over multiple iterations such that every single instance has been verified directly by our machine learning engineers. For vision-language, we combine the chat language model with a vision transformer encoder and train the model to align visual representations to the semantic space of the language model. We further extend the context length to 200K through lightweight continual pretraining and demonstrate strong needle-in-a-haystack retrieval performance. We show that extending the depth of the pretrained checkpoint through continual pretraining further improves performance. We believe that given our current results, continuing to scale up model parameters using thoroughly optimized data will lead to even stronger frontier models.
๊ฐ•๋ ฅํ•œ ๋‹ค์ฐจ์› ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ผ๋ จ์˜ ์–ธ์–ด ๋ฐ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ชจ๋ธ์ธ Yi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Yi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ 6B ๋ฐ 34B pretrained ์–ธ์–ด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ด๋ฅผ ์ฑ„ํŒ… ๋ชจ๋ธ, 200K long context ๋ชจ๋ธ, depth-upscaled ๋ชจ๋ธ, vision-language ๋ชจ๋ธ๋กœ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์€ MMLU์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ, Fine-tuning๋œ ์ฑ„ํŒ… ๋ชจ๋ธ์€ AlpacaEval ๋ฐ Chatbot Arena์™€ ๊ฐ™์€ ์ฃผ์š” ํ‰๊ฐ€ ํ”Œ๋žซํผ์—์„œ ๋†’์€ ์ธ๊ฐ„ ์„ ํ˜ธ๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์Šˆํผ ์ปดํ“จํŒ… ์ธํ”„๋ผ์™€ ๊ณ ์ „์ ์ธ Transformer ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, Yi ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ๋…ธ๋ ฅ์˜ ๊ฒฐ๊ณผ์ธ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•ฉ๋‹ˆ๋‹ค. Pretraining์„ ์œ„ํ•ด ์บ์Šค์ผ€์ด๋“œ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐ ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ 3.1์กฐ tokens์˜ ์˜์–ด ๋ฐ ์ค‘๊ตญ์–ด ์ฝ”ํผ์Šค๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. Fine-tuning์˜ ๊ฒฝ์šฐ, ์†Œ๊ทœ๋ชจ(10K ๋ฏธ๋งŒ) Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋‹ค๋“ฌ์–ด์„œ ๋ชจ๋“  ๋‹จ์ผ ์ธ์Šคํ„ด์Šค๊ฐ€ ๋จธ์‹  ๋Ÿฌ๋‹ ์—”์ง€๋‹ˆ์–ด๊ฐ€ ์ง์ ‘ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. Vision-language์˜ ๊ฒฝ์šฐ, ์ฑ„ํŒ… ์–ธ์–ด ๋ชจ๋ธ๊ณผ Vision Transformer Encoder๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ์–ธ์–ด ๋ชจ๋ธ์˜ ์‹œ๋งจํ‹ฑ ๊ณต๊ฐ„์— ์‹œ๊ฐ์  ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒฝ๋Ÿ‰ ์—ฐ์† Pretraining์„ ํ†ตํ•ด ๋ฌธ๋งฅ ๊ธธ์ด๋ฅผ 200K๊นŒ์ง€ ํ™•์žฅํ•˜๊ณ  ๊ฐ•๋ ฅํ•œ needle-in-a-haystack retrieval ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ์† Pretraining์„ ํ†ตํ•ด pretrained ์ฒดํฌํฌ์ธํŠธ์˜ ๊นŠ์ด๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์ด ๋”์šฑ ํ–ฅ์ƒ๋จ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์ฒ ์ €ํ•˜๊ฒŒ ์ตœ์ ํ™”๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜๋ฉด ๋”์šฑ ๊ฐ•๋ ฅํ•œ ํ”„๋ก ํ‹ฐ์–ด ๋ชจ๋ธ๋กœ ๋ฐœ์ „ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๋ฏฟ์Šต๋‹ˆ๋‹ค.
๊ฐ•๋ ฅํ•œ ๋‹ค์ฐจ์› ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ๋Š” ์ผ๋ จ์˜ ์–ธ์–ด ๋ฐ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ชจ๋ธ์ธ Yi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. Yi ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ 6B ๋ฐ 34B ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋ฉฐ, ์ด๋ฅผ ์ฑ„ํŒ… ๋ชจ๋ธ, 20๋งŒ ๊ฐœ์˜ ๊ธด ์ปจํ…์ŠคํŠธ ๋ชจ๋ธ, ๊นŠ์ด ์—…์Šค์ผ€์ผ๋ง ๋ชจ๋ธ, ๋น„์ „ ์–ธ์–ด ๋ชจ๋ธ๋กœ ํ™•์žฅํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ๋ณธ ๋ชจ๋ธ์€ MMLU์™€ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋ฉฐ, ํŒŒ์ด๋„ํŠœ๋‹๋œ ์ฑ„ํŒ… ๋ชจ๋ธ์€ AlpacaEval ๋ฐ Chatbot Arena์™€ ๊ฐ™์€ ์ฃผ์š” ํ‰๊ฐ€ ํ”Œ๋žซํผ์—์„œ ๋†’์€ ์ธ๊ฐ„ ์„ ํ˜ธ๋„๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์Šˆํผ ์ปดํ“จํŒ… ์ธํ”„๋ผ์™€ ๊ณ ์ „์ ์ธ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, Yi ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ์ฃผ๋กœ ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ๋…ธ๋ ฅ์˜ ๊ฒฐ๊ณผ์ธ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์— ๊ธฐ์ธํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ์บ์Šค์ผ€์ด๋“œ ๋ฐ์ดํ„ฐ ์ค‘๋ณต ์ œ๊ฑฐ ๋ฐ ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ 3.1์กฐ ํ† ํฐ์˜ ์˜์–ด ๋ฐ ์ค‘๊ตญ์–ด ์ฝ”ํผ์Šค๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŒŒ์ด๋„ํŠœ๋‹์˜ ๊ฒฝ์šฐ, ์†Œ๊ทœ๋ชจ(10K ๋ฏธ๋งŒ) ์ธ์ŠคํŠธ๋Ÿญ์…˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด ๋‹ค๋“ฌ์–ด์„œ ๋ชจ๋“  ๋‹จ์ผ ์ธ์Šคํ„ด์Šค๊ฐ€ ๋จธ์‹  ๋Ÿฌ๋‹ ์—”์ง€๋‹ˆ์–ด๊ฐ€ ์ง์ ‘ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์ „ ์–ธ์–ด์˜ ๊ฒฝ์šฐ, ์ฑ„ํŒ… ์–ธ์–ด ๋ชจ๋ธ๊ณผ ๋น„์ „ ํŠธ๋žœ์Šคํฌ๋จธ ์ธ์ฝ”๋”๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ์–ธ์–ด ๋ชจ๋ธ์˜ ์‹œ๋งจํ‹ฑ ๊ณต๊ฐ„์— ์‹œ๊ฐ์  ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋„๋ก ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒฝ๋Ÿ‰ ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ๋ฌธ๋งฅ ๊ธธ์ด๋ฅผ 200K๊นŒ์ง€ ํ™•์žฅํ•˜๊ณ  ๊ฐ•๋ ฅํ•œ ์นจ-๊ฑด์ดˆ๋”๋ฏธ ๊ฒ€์ƒ‰ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ์† ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด ์‚ฌ์ „ ํ•™์Šต ์ฒดํฌํฌ์ธํŠธ์˜ ๊นŠ์ด๋ฅผ ํ™•์žฅํ•˜๋ฉด ์„ฑ๋Šฅ์ด ๋”์šฑ ํ–ฅ์ƒ๋จ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ๊ฒฐ๊ณผ๋ฅผ ๊ณ ๋ คํ•  ๋•Œ ์ฒ ์ €ํ•˜๊ฒŒ ์ตœ์ ํ™”๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ณ„์† ํ™•์žฅํ•˜๋ฉด ๋”์šฑ ๊ฐ•๋ ฅํ•œ ํ”„๋ก ํ‹ฐ์–ด ๋ชจ๋ธ๋กœ ๋ฐœ์ „ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๋ฏฟ์Šต๋‹ˆ๋‹ค.
Question answering also includes generative and extractive forms. It is a common task in NLP that relies on domain-specific knowledge. RALMs can achieve better results than traditional language models by utilizing externally stored knowledge. Common question answering tasks include domain-specific QA, open-domain QA(ODQA), and multi-hop QA. In the medical field, large language models are commonly used, PKG(Luo et al., 2023) uses the relevant data from the MedMC-QA dataset,using the questions in the training set as input and the medical explanations as background knowledge, and the accuracy of the background knowledge generated by the model as an evaluation metric. HyKGE(Jiang et al., 2023) also targets question answering in the medical field, but uses a knowledge graph-enhanced approach. When targeting ODQA tasks, RAG (Lewis et al., 2020) considers questions and answers as input-output text pairs and trains by minimizing the negative log-likelihood of the answers. In contrast, ICRALM (Ram et al., 2023) exclusively performs the ODQA task using frozen LMs, which have not been enhanced by pre-training, fine-tuning, or retrieval, as well as the associated knowledge documents. Other models (Wang et al., 2023)(Lin et al., 2023)(Shi et al., 2023) were also trained for the ODQA task. In relation to the multi-hop QA task, FILCO (Wang et al., 2023) used their proposed filtered retrieval method to filter multiple documents. They validated their approach using the HotpotQA dataset. RR (He et al., 2022), on the other hand, used the Chain-of-Thought (CoT) approach to address the multi-hop problem. In addition, many other models (Jiang et al., 2023)(Khattab et al., 2022) deal with multi-hop problems.
์งˆ๋ฌธ ๋‹ต๋ณ€์—๋Š” ์ƒ์„ฑํ˜• ๋ฐ ์ถ”์ถœํ˜• ํ˜•ํƒœ๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์— ์˜์กดํ•˜๋Š” NLP์˜ ์ผ๋ฐ˜์ ์ธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์™ธ๋ถ€์— ์ €์žฅ๋œ ์ง€์‹์„ ํ™œ์šฉํ•จ์œผ๋กœ์จ RALM์€ ๊ธฐ์กด ์–ธ์–ด ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์—๋Š” ๋„๋ฉ”์ธ๋ณ„ QA, open-domain QA(ODQA), multi-hop QA ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ๋Š” Large Language Models(LLMs)๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, PKG(Luo et al., 2023)๋Š” MedMC-QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ด€๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ์„ธํŠธ์˜ ์งˆ๋ฌธ์„ ์ž…๋ ฅ์œผ๋กœ, ์˜ํ•™์  ์„ค๋ช…์„ ๋ฐฐ๊ฒฝ ์ง€์‹์œผ๋กœ ์‚ฌ์šฉํ•˜๋ฉฐ, ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐฐ๊ฒฝ ์ง€์‹์˜ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. HyKGE(Jiang et al., 2023)๋„ ์˜๋ฃŒ ๋ถ„์•ผ์˜ ์งˆ๋ฌธ ๋‹ต๋ณ€์„ ๋ชฉํ‘œ๋กœ ํ•˜์ง€๋งŒ ์ง€์‹ ๊ทธ๋ž˜ํ”„ ๊ฐ•ํ™” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ODQA ์ž‘์—…์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ, RAG(Lewis et al., 2020)๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ž…๋ ฅ-์ถœ๋ ฅ ํ…์ŠคํŠธ ์Œ์œผ๋กœ ๊ฐ„์ฃผํ•˜๊ณ  ๋‹ต๋ณ€์˜ ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ์„ ์ตœ์†Œํ™”ํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ICRALM(Ram et al., 2023)์€ ์‚ฌ์ „ ํ•™์Šต, Fine-tuning ๋˜๋Š” retrieval๋กœ ๊ฐ•ํ™”๋˜์ง€ ์•Š์€ ๋™๊ฒฐ๋œ LMs๊ณผ ๊ด€๋ จ ์ง€์‹ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ODQA ์ž‘์—…๋งŒ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ(Wang et al., 2023)(Lin et al., 2023)(Shi et al., 2023)๋„ ODQA ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. multi-hop QA ์ž‘์—…๊ณผ ๊ด€๋ จํ•˜์—ฌ FILCO(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๋ฌธ์„œ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ํ•„ํ„ฐ๋ง retrieval ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ HotpotQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— RR(He et al., 2022)์€ multi-hop ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Chain-of-Thought(CoT) ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ๋งŽ์€ ๋ชจ๋ธ(Jiang et al., 2023)(Khattab et al., 2022)์ด multi-hop ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
์งˆ๋ฌธ ๋‹ต๋ณ€์—๋Š” ์ƒ์„ฑํ˜• ๋ฐ ์ถ”์ถœํ˜• ํ˜•ํƒœ๋„ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์— ์˜์กดํ•˜๋Š” NLP์˜ ์ผ๋ฐ˜์ ์ธ ์ž‘์—…์ž…๋‹ˆ๋‹ค. ์™ธ๋ถ€์— ์ €์žฅ๋œ ์ง€์‹์„ ํ™œ์šฉํ•จ์œผ๋กœ์จ RALM์€ ๊ธฐ์กด ์–ธ์–ด ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ์ž‘์—…์—๋Š” ๋„๋ฉ”์ธ๋ณ„ QA, ์˜คํ”ˆ ๋„๋ฉ”์ธ QA(ODQA), ๋ฉ€ํ‹ฐํ™‰ QA ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์˜๋ฃŒ ๋ถ„์•ผ์—์„œ๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋ฉฐ, PKG(Luo et al., 2023)๋Š” MedMC-QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ด€๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ ์„ธํŠธ์˜ ์งˆ๋ฌธ์„ ์ž…๋ ฅ์œผ๋กœ, ์˜ํ•™์  ์„ค๋ช…์„ ๋ฐฐ๊ฒฝ ์ง€์‹์œผ๋กœ ์‚ฌ์šฉํ•˜๋ฉฐ, ๋ชจ๋ธ์—์„œ ์ƒ์„ฑ๋œ ๋ฐฐ๊ฒฝ ์ง€์‹์˜ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. HyKGE(Jiang et al., 2023)๋„ ์˜๋ฃŒ ๋ถ„์•ผ์˜ ์งˆ๋ฌธ ๋‹ต๋ณ€์„ ๋ชฉํ‘œ๋กœ ํ•˜์ง€๋งŒ ์ง€์‹ ๊ทธ๋ž˜ํ”„ ๊ฐ•ํ™” ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ODQA ์ž‘์—…์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ฒฝ์šฐ, RAG(Lewis et al., 2020)๋Š” ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ž…๋ ฅ-์ถœ๋ ฅ ํ…์ŠคํŠธ ์Œ์œผ๋กœ ๊ฐ„์ฃผํ•˜๊ณ  ๋‹ต๋ณ€์˜ ์Œ์˜ ๋กœ๊ทธ ๊ฐ€๋Šฅ์„ฑ์„ ์ตœ์†Œํ™”ํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ICRALM(Ram et al., 2023)์€ ์‚ฌ์ „ ํ•™์Šต, ๋ฏธ์„ธ ์กฐ์ • ๋˜๋Š” ๊ฒ€์ƒ‰์œผ๋กœ ๊ฐ•ํ™”๋˜์ง€ ์•Š์€ ๋™๊ฒฐ๋œ LM๊ณผ ๊ด€๋ จ ์ง€์‹ ๋ฌธ์„œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ODQA ์ž‘์—…๋งŒ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ(Wang et al., 2023)(Lin et al., 2023)(Shi et al., 2023)๋„ ODQA ์ž‘์—…์„ ์œ„ํ•ด ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ฉ€ํ‹ฐํ™‰ QA ์ž‘์—…๊ณผ ๊ด€๋ จํ•˜์—ฌ FILCO(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๋ฌธ์„œ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆํ•œ ํ•„ํ„ฐ๋ง ๊ฒ€์ƒ‰ ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ HotpotQA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฒ€์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— RR(He et al., 2022)์€ ๋ฉ€ํ‹ฐํ™‰ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด Chain-of-Thought(CoT) ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค๋ฅธ ๋งŽ์€ ๋ชจ๋ธ(Jiang et al., 2023)(Khattab et al., 2022)์ด ๋ฉ€ํ‹ฐํ™‰ ๋ฌธ์ œ๋ฅผ ๋‹ค๋ฃจ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**Scientific research** is another promising field that LLMs can empower the development progress. Prior research demonstrates the effectiveness of LLMs in handling knowledge-intensive scientific tasks (_e.g._, PubMedQA [928], BioASQ [929]), especially for LLMs that are trained on scientific-related corpora [35, 203, 930]. Given the excellent general abilities and broad scientific knowledge, LLMs hold significant potential as helpful assistants across various stages of the scientific research pipeline [931]. First, during the literature survey stage, LLMs can help conduct a comprehensive overview of the progress in a specific research field [932, 933]. Second, during the research idea generation stage, LLMs demonstrate the ability to generate intriguing scientific hypotheses [934]. Third, during the data analysis stage, LLMs can be employed to conduct automatic approaches to analyzing the data characteristics, including data exploration, visualization, and deriving analytical conclusions [935, 936]. Fourth, during the paper writing stage, researchers can also benefit from the assistance of LLMs in scientific writing [937, 938], in which LLMs can offer valuable support for scientific writing through diverse means, such as summarizing the existing content and polishing the writing [939]. In addition, LLMs can aid in the automated paper review process, encompassing tasks such as error detection, checklist verification, and candidate ranking [940]. Despite these advances, there is much room for improving the capacities of LLMs to serve as helpful, trustworthy scientific assistants, to both increase the quality of the generated scientific content and reduce the harmful hallucinations.
**๊ณผํ•™ ์—ฐ๊ตฌ**๋Š” LLM์ด ๊ฐœ๋ฐœ ์ง„์ „์„ ๊ฐ•ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋˜ ๋‹ค๋ฅธ ์œ ๋งํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋Š” ์ง€์‹ ์ง‘์•ฝ์ ์ธ ๊ณผํ•™์  ์ž‘์—…(์˜ˆ: PubMedQA [928], BioASQ [929])์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ LLM์˜ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ํŠนํžˆ ๊ณผํ•™ ๊ด€๋ จ ์ฝ”ํผ๋ผ [35, 203, 930]์— ๋Œ€ํ•ด ํ•™์Šต๋œ LLM์˜ ๊ฒฝ์šฐ ๋”์šฑ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๋›ฐ์–ด๋‚œ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ๊ณผํ•™์  ์ง€์‹์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์€ ๊ณผํ•™ ์—ฐ๊ตฌ ํŒŒ์ดํ”„๋ผ์ธ์˜ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์œ ์šฉํ•œ ๋„์šฐ๋ฏธ๋กœ์„œ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [931]. ์ฒซ์งธ, ๋ฌธํ—Œ ์กฐ์‚ฌ ๋‹จ๊ณ„์—์„œ LLM์€ ํŠน์ • ์—ฐ๊ตฌ ๋ถ„์•ผ์˜ ์ง„ํ–‰ ์ƒํ™ฉ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [932, 933]. ๋‘˜์งธ, ์—ฐ๊ตฌ ์•„์ด๋””์–ด ์ƒ์„ฑ ๋‹จ๊ณ„์—์„œ LLM์€ ํฅ๋ฏธ๋กœ์šด ๊ณผํ•™์  ๊ฐ€์„ค์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค [934]. ์…‹์งธ, ๋ฐ์ดํ„ฐ ๋ถ„์„ ๋‹จ๊ณ„์—์„œ๋Š” ๋ฐ์ดํ„ฐ ํƒ์ƒ‰, ์‹œ๊ฐํ™”, ๋ถ„์„์  ๊ฒฐ๋ก  ๋„์ถœ ๋“ฑ ๋ฐ์ดํ„ฐ ํŠน์„ฑ์„ ๋ถ„์„ํ•˜๋Š” ์ž๋™ํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ LLM์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [935, 936]. ๋„ท์งธ, ๋…ผ๋ฌธ ์ž‘์„ฑ ๋‹จ๊ณ„์—์„œ ์—ฐ๊ตฌ์ž๋“ค์€ ๊ณผํ•™์  ๊ธ€์“ฐ๊ธฐ [937, 938]์— ์žˆ์–ด์„œ๋„ LLM์˜ ๋„์›€์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์€ ๊ธฐ์กด ์ฝ˜ํ…์ธ ๋ฅผ ์š”์•ฝํ•˜๊ณ  ๊ธ€์“ฐ๊ธฐ๋ฅผ ๋‹ค๋“ฌ๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๊ณผํ•™์  ๊ธ€์“ฐ๊ธฐ์— ๊ท€์ค‘ํ•œ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [939]. ๋˜ํ•œ LLM์€ ์˜ค๋ฅ˜ ๊ฐ์ง€, ์ฒดํฌ๋ฆฌ์ŠคํŠธ ํ™•์ธ, ํ›„๋ณด์ž ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๋“ฑ์˜ ์ž‘์—…์„ ํฌํ•จํ•˜๋Š” ์ž๋™ํ™”๋œ ๋…ผ๋ฌธ ๊ฒ€ํ†  ํ”„๋กœ์„ธ์Šค๋ฅผ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [940]. ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๊ณผํ•™ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ์„ ๋†’์ด๊ณ  ์œ ํ•ดํ•œ Hallucination์„ ์ค„์ด๊ธฐ ์œ„ํ•ด LLM์ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ณผํ•™์  ๋„์šฐ๋ฏธ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์—ญ๋Ÿ‰์„ ๊ฐœ์„ ํ•  ์—ฌ์ง€๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค.
**๊ณผํ•™ ์—ฐ๊ตฌ**๋Š” LLM์ด ๊ฐœ๋ฐœ ์ง„์ „์„ ๊ฐ•ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ๋˜ ๋‹ค๋ฅธ ์œ ๋งํ•œ ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ๋Š” ์ง€์‹ ์ง‘์•ฝ์ ์ธ ๊ณผํ•™์  ์ž‘์—…(์˜ˆ: PubMedQA [928], BioASQ [929])์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ LLM์˜ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ํŠนํžˆ ๊ณผํ•™ ๊ด€๋ จ ์ฝ”ํผ๋ผ [35, 203, 930]์— ๋Œ€ํ•ด ํ•™์Šต๋œ LLM์˜ ๊ฒฝ์šฐ ๋”์šฑ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค. ๋›ฐ์–ด๋‚œ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ๊ณผํ•™์  ์ง€์‹์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์€ ๊ณผํ•™ ์—ฐ๊ตฌ ํŒŒ์ดํ”„๋ผ์ธ์˜ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„์—์„œ ์œ ์šฉํ•œ ๋„์šฐ๋ฏธ๋กœ์„œ ์ƒ๋‹นํ•œ ์ž ์žฌ๋ ฅ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค [931]. ์ฒซ์งธ, ๋ฌธํ—Œ ์กฐ์‚ฌ ๋‹จ๊ณ„์—์„œ LLM์€ ํŠน์ • ์—ฐ๊ตฌ ๋ถ„์•ผ์˜ ์ง„ํ–‰ ์ƒํ™ฉ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [932, 933]. ๋‘˜์งธ, ์—ฐ๊ตฌ ์•„์ด๋””์–ด ์ƒ์„ฑ ๋‹จ๊ณ„์—์„œ LLM์€ ํฅ๋ฏธ๋กœ์šด ๊ณผํ•™์  ๊ฐ€์„ค์„ ์ƒ์„ฑํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค [934]. ์…‹์งธ, ๋ฐ์ดํ„ฐ ๋ถ„์„ ๋‹จ๊ณ„์—์„œ๋Š” ๋ฐ์ดํ„ฐ ํƒ์ƒ‰, ์‹œ๊ฐํ™”, ๋ถ„์„์  ๊ฒฐ๋ก  ๋„์ถœ ๋“ฑ ๋ฐ์ดํ„ฐ ํŠน์„ฑ์„ ๋ถ„์„ํ•˜๋Š” ์ž๋™ํ™”๋œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ LLM์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [935, 936]. ๋„ท์งธ, ๋…ผ๋ฌธ ์ž‘์„ฑ ๋‹จ๊ณ„์—์„œ ์—ฐ๊ตฌ์ž๋“ค์€ ๊ณผํ•™์  ๊ธ€์“ฐ๊ธฐ [937, 938]์— ์žˆ์–ด์„œ๋„ LLM์˜ ๋„์›€์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์œผ๋ฉฐ, LLM์€ ๊ธฐ์กด ์ฝ˜ํ…์ธ ๋ฅผ ์š”์•ฝํ•˜๊ณ  ๊ธ€์“ฐ๊ธฐ๋ฅผ ๋‹ค๋“ฌ๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๊ณผํ•™์  ๊ธ€์“ฐ๊ธฐ์— ๊ท€์ค‘ํ•œ ์ง€์›์„ ์ œ๊ณตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [939]. ๋˜ํ•œ LLM์€ ์˜ค๋ฅ˜ ๊ฐ์ง€, ์ฒดํฌ๋ฆฌ์ŠคํŠธ ํ™•์ธ, ํ›„๋ณด์ž ์ˆœ์œ„ ๋งค๊ธฐ๊ธฐ ๋“ฑ์˜ ์ž‘์—…์„ ํฌํ•จํ•˜๋Š” ์ž๋™ํ™”๋œ ๋…ผ๋ฌธ ๊ฒ€ํ†  ํ”„๋กœ์„ธ์Šค๋ฅผ ์ง€์›ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค [940]. ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์„ฑ๋œ ๊ณผํ•™ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ์„ ๋†’์ด๊ณ  ์œ ํ•ดํ•œ ํ™˜๊ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด LLM์ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ณผํ•™์  ๋„์šฐ๋ฏธ ์—ญํ• ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์—ญ๋Ÿ‰์„ ๊ฐœ์„ ํ•  ์—ฌ์ง€๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Memory Reflection_. Memory reflection emulates humans' ability to witness and evaluate their own cognitive, emotional, and behavioral processes. When adapted to agents, the objective is to provide agents with the capability to independently summarize and infer more abstract, complex and high-level information. More specifically, in Generative Agent [20], the agent has the capability to summarize its past experiences stored in memory into broader and more abstract insights. To begin with, the agent generates three key questions based on its recent memories. Then, these questions are used to query the memory to obtain relevant information. Building upon the acquired information, the agent generates five insights, which reflect the agent high-level ideas. For example, the low-level memories "Klaus Mueller is writing a research paper", "Klaus Mueller is engaging with a librarian to further his research", and "Klaus Mueller is conversing with Ayesha Khan about his research" can induce the high-level insight "Klaus Mueller is dedicated to his research". In addition, the reflection process can occur hierarchically, meaning that the insights can be generated based on existing insights. In GITM [16], the actions that successfully accomplish the sub-goals are stored in a list. When the list contains more than five elements, the agent summarizes them into a common and abstract pattern and replaces all the elements. In ExpeL [44], two approaches are introduced for the agent to acquire reflection. Firstly, the agent compares successful or failed trajectories within the same task. Secondly, the agent learns from a collection of successful trajectories to gain experiences.
\(\bullet\)_Memory Reflection_. Memory reflection์€ ์ธ๊ฐ„์ด ์ž์‹ ์˜ ์ธ์ง€, ๊ฐ์ •, ํ–‰๋™ ๊ณผ์ •์„ ๋ชฉ๊ฒฉํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋ชจ๋ฐฉํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์— ์ ์šฉํ•  ๋•Œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ณด๋‹ค ์ถ”์ƒ์ ์ด๊ณ  ๋ณต์žกํ•˜๋ฉฐ ๋†’์€ ์ˆ˜์ค€์˜ ์ •๋ณด๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ์š”์•ฝํ•˜๊ณ  ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ, Generative Agent[20]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅ๋œ ๊ณผ๊ฑฐ ๊ฒฝํ—˜์„ ๋”์šฑ ๊ด‘๋ฒ”์œ„ํ•˜๊ณ  ์ถ”์ƒ์ ์ธ ์ธ์‚ฌ์ดํŠธ๋กœ ์š”์•ฝํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๋Š” ์ตœ๊ทผ ๊ธฐ์–ต์„ ๋ฐ”ํƒ•์œผ๋กœ ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ์งˆ๋ฌธ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ํš๋“ํ•œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ์—์ด์ „ํŠธ์˜ ๋†’์€ ์ˆ˜์ค€์˜ ์•„์ด๋””์–ด๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” 5๊ฐ€์ง€ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ•˜๊ณ  ์žˆ๋‹ค", "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ์‚ฌ์„œ์™€ ๊ต๋ฅ˜ํ•˜๊ณ  ์žˆ๋‹ค", "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์•„์ด์ƒค ์นธ๊ณผ ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ๋Œ€ํ™”ํ•˜๊ณ  ์žˆ๋‹ค"์™€ ๊ฐ™์€ ๋‚ฎ์€ ์ˆ˜์ค€์˜ ๊ธฐ์–ต์€ "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ์— ์ „๋…ํ•˜๊ณ  ์žˆ๋‹ค"๋ผ๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์œ ๋„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฐ˜์„ฑ ๊ณผ์ •์€ ๊ณ„์ธต์ ์œผ๋กœ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ธฐ์กด ์ธ์‚ฌ์ดํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. GITM[16]์—์„œ๋Š” ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ๋‹ฌ์„ฑํ•œ ํ–‰๋™์ด ๋ชฉ๋ก์— ์ €์žฅ๋ฉ๋‹ˆ๋‹ค. ๋ชฉ๋ก์— 5๊ฐœ ์ด์ƒ์˜ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋ฅผ ๊ณตํ†ต์ ์ด๊ณ  ์ถ”์ƒ์ ์ธ ํŒจํ„ด์œผ๋กœ ์š”์•ฝํ•˜๊ณ  ๋ชจ๋“  ์š”์†Œ๋ฅผ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. ExpeL[44]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฐ˜์„ฑ์„ ์Šต๋“ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์—์ด์ „ํŠธ๋Š” ๋™์ผํ•œ ์ž‘์—… ๋‚ด์—์„œ ์„ฑ๊ณตํ–ˆ๊ฑฐ๋‚˜ ์‹คํŒจํ•œ ๊ถค์ ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์—์ด์ „ํŠธ๋Š” ์„ฑ๊ณต์ ์ธ ๊ถค์  ๋ชจ์Œ์—์„œ ํ•™์Šตํ•˜์—ฌ ๊ฒฝํ—˜์„ ํš๋“ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_๋ฉ”๋ชจ๋ฆฌ ๋ฐ˜์‚ฌ_. ๋ฉ”๋ชจ๋ฆฌ ๋ฐ˜์‚ฌ๋Š” ์ธ๊ฐ„์ด ์ž์‹ ์˜ ์ธ์ง€, ๊ฐ์ •, ํ–‰๋™ ๊ณผ์ •์„ ๋ชฉ๊ฒฉํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋Šฅ๋ ฅ์„ ๋ชจ๋ฐฉํ•œ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์— ์ ์šฉํ•  ๋•Œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ณด๋‹ค ์ถ”์ƒ์ ์ด๊ณ  ๋ณต์žกํ•˜๋ฉฐ ๋†’์€ ์ˆ˜์ค€์˜ ์ •๋ณด๋ฅผ ๋…๋ฆฝ์ ์œผ๋กœ ์š”์•ฝํ•˜๊ณ  ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค. ๋ณด๋‹ค ๊ตฌ์ฒด์ ์œผ๋กœ, ์ƒ์„ฑ ์—์ด์ „ํŠธ[20]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅ๋œ ๊ณผ๊ฑฐ ๊ฒฝํ—˜์„ ๋”์šฑ ๊ด‘๋ฒ”์œ„ํ•˜๊ณ  ์ถ”์ƒ์ ์ธ ์ธ์‚ฌ์ดํŠธ๋กœ ์š”์•ฝํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ์—์ด์ „ํŠธ๋Š” ์ตœ๊ทผ ๊ธฐ์–ต์„ ๋ฐ”ํƒ•์œผ๋กœ ์„ธ ๊ฐ€์ง€ ํ•ต์‹ฌ ์งˆ๋ฌธ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด๋Ÿฌํ•œ ์งˆ๋ฌธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ฟผ๋ฆฌํ•˜์—ฌ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. ํš๋“ํ•œ ์ •๋ณด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์—์ด์ „ํŠธ๋Š” ์—์ด์ „ํŠธ์˜ ๋†’์€ ์ˆ˜์ค€์˜ ์•„์ด๋””์–ด๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” 5๊ฐ€์ง€ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ ๋…ผ๋ฌธ์„ ์ž‘์„ฑํ•˜๊ณ  ์žˆ๋‹ค", "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ์‚ฌ์„œ์™€ ๊ต๋ฅ˜ํ•˜๊ณ  ์žˆ๋‹ค", "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์•„์ด์ƒค ์นธ๊ณผ ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ๋Œ€ํ™”ํ•˜๊ณ  ์žˆ๋‹ค"์™€ ๊ฐ™์€ ๋‚ฎ์€ ์ˆ˜์ค€์˜ ๊ธฐ์–ต์€ "ํด๋ผ์šฐ์Šค ๋ฎ๋Ÿฌ๊ฐ€ ์—ฐ๊ตฌ์— ์ „๋…ํ•˜๊ณ  ์žˆ๋‹ค"๋ผ๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์œ ๋„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฐ˜์„ฑ ๊ณผ์ •์€ ๊ณ„์ธต์ ์œผ๋กœ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๊ธฐ์กด ์ธ์‚ฌ์ดํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. GITM [16]์—์„œ๋Š” ํ•˜์œ„ ๋ชฉํ‘œ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ๋‹ฌ์„ฑํ•œ ํ–‰๋™์ด ๋ชฉ๋ก์— ์ €์žฅ๋ฉ๋‹ˆ๋‹ค. ๋ชฉ๋ก์— 5๊ฐœ ์ด์ƒ์˜ ์š”์†Œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ์—์ด์ „ํŠธ๋Š” ์ด๋ฅผ ๊ณตํ†ต์ ์ด๊ณ  ์ถ”์ƒ์ ์ธ ํŒจํ„ด์œผ๋กœ ์š”์•ฝํ•˜๊ณ  ๋ชจ๋“  ์š”์†Œ๋ฅผ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. ExpeL [44]์—์„œ๋Š” ์—์ด์ „ํŠธ๊ฐ€ ๋ฐ˜์„ฑ์„ ์Šต๋“ํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์—์ด์ „ํŠธ๋Š” ๋™์ผํ•œ ์ž‘์—… ๋‚ด์—์„œ ์„ฑ๊ณตํ–ˆ๊ฑฐ๋‚˜ ์‹คํŒจํ•œ ๊ถค์ ์„ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์—์ด์ „ํŠธ๋Š” ์„ฑ๊ณต์ ์ธ ๊ถค์  ๋ชจ์Œ์—์„œ ํ•™์Šตํ•˜์—ฌ ๊ฒฝํ—˜์„ ํš๋“ํ•ฉ๋‹ˆ๋‹ค.
**LLaMA Model Family**. The collection of LLaMA models [57] were introduced by Meta AI in February, 2023, consisting of four sizes (7B, 13B, 30B and 65B). Since released, LLaMA has attracted extensive attention from both research and industry communities. LLaMA models have achieved very excellent performance on various open benchmarks, which have become the most popular open language models thus far. A large number of researchers have extended LLaMA models by either instruction tuning or continual pretraining. In particular, instruction tuning LLaMA has become a major approach to developing customized or specialized models, due to the relatively low computational costs. To effectively adapt LLaMA models in non-English languages, it often needs to extend the original vocabulary (trained mainly on English corpus) or fine-tune it with instructions or data in the target language. Among these extended models, Stanford Alpaca [142] is the first open instruct-following model fine-tuned based on LLaMA (7B). It is trained by 52K instruction-following demonstrations generated via self-instruct [143] using text-davinci-003. The instruction data, named _Alpaca-52K_, and training code have been extensively adopted in subsequent work, such as Alpaca-LoRA [144] (a reproduction of Stanford Alpaca using LoRA [145]), Koala [146], and BELLE [147]. In addition, Vicuna [138] is another popular LLaMA variant, trained upon user-shared conversations collected from ShareGPT [148]. Due to the excellent performance and availability of the LLaMA model family, many multimodal models incorporate them as the base language models, to achieve strong language understanding and generation abilities.
**LLaMA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ**. LLaMA ๋ชจ๋ธ ์ปฌ๋ ‰์…˜ [57]์€ 2023๋…„ 2์›”์— Meta AI์—์„œ ์†Œ๊ฐœํ•œ ๊ฒƒ์œผ๋กœ, 4๊ฐ€์ง€ ํฌ๊ธฐ(7B, 13B, 30B, 65B)๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ถœ์‹œ ์ดํ›„ LLaMA๋Š” ์—ฐ๊ตฌ ๋ฐ ์—…๊ณ„ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋งค์šฐ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์˜คํ”ˆ ์–ธ์–ด ๋ชจ๋ธ์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด Instruction Tuning ๋˜๋Š” ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด LLaMA ๋ชจ๋ธ์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, Instruction Tuning LLaMA๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ธํ•ด ๋งž์ถคํ˜• ๋˜๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ์–ธ์–ด์—์„œ LLaMA ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ์กฐ์ •ํ•˜๋ ค๋ฉด ์›๋ž˜ ์–ดํœ˜(์ฃผ๋กœ ์˜์–ด ์ฝ”ํผ์Šค์—์„œ ํ•™์Šต)๋ฅผ ํ™•์žฅํ•˜๊ฑฐ๋‚˜ ๋Œ€์ƒ ์–ธ์–ด์˜ Instruction ๋˜๋Š” ๋ฐ์ดํ„ฐ๋กœ Fine-tuningํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ™•์žฅ ๋ชจ๋ธ ์ค‘ Stanford Alpaca [142]๋Š” LLaMA(7B)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Fine-tuningํ•œ ์ตœ์ดˆ์˜ ์˜คํ”ˆ Instruction-following ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ํ…์ŠคํŠธ-davinci-003์„ ์‚ฌ์šฉํ•˜์—ฌ self-instruct [143]๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋œ 52,000๊ฐœ์˜ Instruction-following ๋ฐ๋ชจ๋ฅผ ํ†ตํ•ด ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. Alpaca-52K_๋ผ๋Š” Instruction ๋ฐ์ดํ„ฐ์™€ ํ•™์Šต ์ฝ”๋“œ๋Š” Alpaca-LoRA [144](LoRA [145]๋ฅผ ์‚ฌ์šฉํ•œ Stanford Alpaca์˜ ์žฌํ˜„), Koala [146], BELLE [147] ๋“ฑ ํ›„์† ์ž‘์—…์—์„œ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, Vicuna [138]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋Œ€ํ™”๋ฅผ ShareGPT [148]์—์„œ ์ˆ˜์ง‘ํ•˜์—ฌ ํ•™์Šตํ•œ ๋˜ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” LLaMA ๋ณ€ํ˜•์ž…๋‹ˆ๋‹ค. LLaMA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ๊ณผ ๊ฐ€์šฉ์„ฑ์œผ๋กœ ์ธํ•ด ๋งŽ์€ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์ด ์ด๋ฅผ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ๋กœ ํ†ตํ•ฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**LLaMA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ**. LLaMA ๋ชจ๋ธ ์ปฌ๋ ‰์…˜ [57]์€ 2023๋…„ 2์›”์— Meta AI์—์„œ ์†Œ๊ฐœํ•œ ๊ฒƒ์œผ๋กœ, 4๊ฐ€์ง€ ํฌ๊ธฐ(7B, 13B, 30B, 65B)๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ถœ์‹œ ์ดํ›„ LLaMA๋Š” ์—ฐ๊ตฌ ๋ฐ ์—…๊ณ„ ์ปค๋ฎค๋‹ˆํ‹ฐ ๋ชจ๋‘์—์„œ ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. LLaMA ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋งค์šฐ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜์—ฌ ํ˜„์žฌ๊นŒ์ง€ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์˜คํ”ˆ ์–ธ์–ด ๋ชจ๋ธ์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋ช…๋ น์–ด ํŠœ๋‹ ๋˜๋Š” ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ํ†ตํ•ด LLaMA ๋ชจ๋ธ์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ช…๋ น์–ด ํŠœ๋‹ LLaMA๋Š” ์ƒ๋Œ€์ ์œผ๋กœ ๋‚ฎ์€ ๊ณ„์‚ฐ ๋น„์šฉ์œผ๋กœ ์ธํ•ด ๋งž์ถคํ˜• ๋˜๋Š” ํŠน์ˆ˜ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ์ฃผ์š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ์–ธ์–ด์—์„œ LLaMA ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ์กฐ์ •ํ•˜๋ ค๋ฉด ์›๋ž˜ ์–ดํœ˜(์ฃผ๋กœ ์˜์–ด ์ฝ”ํผ์Šค์—์„œ ํ•™์Šต)๋ฅผ ํ™•์žฅํ•˜๊ฑฐ๋‚˜ ๋Œ€์ƒ ์–ธ์–ด์˜ ๋ช…๋ น์–ด ๋˜๋Š” ๋ฐ์ดํ„ฐ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ™•์žฅ ๋ชจ๋ธ ์ค‘ Stanford Alpaca [142]๋Š” LLaMA(7B)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ์ตœ์ดˆ์˜ ์˜คํ”ˆ ๋ช…๋ น์–ด ์ถ”์ข… ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ํ…์ŠคํŠธ-๋‹ค๋นˆ์น˜-003์„ ์‚ฌ์šฉํ•˜์—ฌ ์…€ํ”„-์ธ์ŠคํŠธ๋ŸญํŠธ [143]๋ฅผ ํ†ตํ•ด ์ƒ์„ฑ๋œ 52,000๊ฐœ์˜ ๋ช…๋ น์–ด ์ถ”์ข… ๋ฐ๋ชจ๋ฅผ ํ†ตํ•ด ํ•™์Šต๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด-52K_๋ผ๋Š” ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ์™€ ํ•™์Šต ์ฝ”๋“œ๋Š” Alpaca-LoRA [144](LoRA [145]๋ฅผ ์‚ฌ์šฉํ•œ Stanford Alpaca์˜ ์žฌํ˜„), Koala [146], BELLE [147] ๋“ฑ ํ›„์† ์ž‘์—…์—์„œ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋น„์ฟ ๋‚˜ [138]๋Š” ์‚ฌ์šฉ์ž๊ฐ€ ๊ณต์œ ํ•œ ๋Œ€ํ™”๋ฅผ ShareGPT [148]์—์„œ ์ˆ˜์ง‘ํ•˜์—ฌ ํ•™์Šตํ•œ ๋˜ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” LLaMA ๋ณ€ํ˜•์ž…๋‹ˆ๋‹ค. LLaMA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ๊ณผ ๊ฐ€์šฉ์„ฑ์œผ๋กœ ์ธํ•ด ๋งŽ์€ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์ด ์ด๋ฅผ ๊ธฐ๋ณธ ์–ธ์–ด ๋ชจ๋ธ๋กœ ํ†ตํ•ฉํ•˜์—ฌ ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ดํ•ด ๋ฐ ์ƒ์„ฑ ๊ธฐ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**Synergy-Augmented LLM.** To solve complex tasks (_e.g.,_ multi-hop question answering [656]), it often requires LLMs to query a KG multiple times, following a systematic solution plan. We call such a multi-turn interaction approach to enhancing LLM _synergy-augmented LLM_. To better synergize the LLM and KG in a complementary manner, recent studies propose to decompose the complex task into multiple subgoals and iteratively solve each one by leveraging the necessary knowledge from KG [870, 871, 458]. In this process, the LLM can be regarded as an autonomous agent (detailed in Section 8.1.6), which automatically generates the plan and executes it through interaction with the KG environment [870]. Specially, the mainstream approaches typically start by enumerating the candidates using the available knowledge information at the current step, and then retrieve the most appropriate candidates for the next step according to the question [871, 870]. By iterating the above two steps, LLMs can gradually collect relevant evidence [871, 870], and finally approach the correct solution. Despite the effectiveness, enumeration of the candidates over the KG would lead to a vast search space [872]. To address it, StructGPT [458] proposes a more efficient way to access knowledge information using the specialized interfaces for KGs. Specifically,it carefully designs the specialized interfaces according to the common data operations on KG (_e.g._, relation extraction and triple extraction), to ensure efficient and accurate data extraction. In this way, LLMs can be instructed to better manipulate and process the structural information of KGs, thus achieving improved task performance.
**Synergy-Augmented LLM.** ๋ณต์žกํ•œ ์ž‘์—…(_์˜ˆ:_ multi-hop question answering[656])์„ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด LLM์ด ์ฒด๊ณ„์ ์ธ ์†”๋ฃจ์…˜ ๊ณ„ํš์— ๋”ฐ๋ผ KG๋ฅผ ์—ฌ๋Ÿฌ ๋ฒˆ ์ฟผ๋ฆฌํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹ค์ค‘ ํ„ด ์ƒํ˜ธ ์ž‘์šฉ ๋ฐฉ์‹์„ LLM์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์— _synergy-augmented LLM_์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์™„์ ์ธ ๋ฐฉ์‹์œผ๋กœ LLM๊ณผ KG์„ ๋” ์ž˜ ์‹œ๋„ˆ์ง€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ํ•„์š”ํ•œ ์ง€์‹์„ KG์—์„œ ํ™œ์šฉํ•˜์—ฌ ๊ฐ๊ฐ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค[870, 871, 458]. ์ด ๊ณผ์ •์—์„œ LLM์€ ์ž๋™์œผ๋กœ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  KG ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์‹คํ–‰ํ•˜๋Š” ์ž์œจ ์—์ด์ „ํŠธ(8.1.6์ ˆ์—์„œ ์ž์„ธํžˆ ์„ค๋ช…)๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[870]. ํŠนํžˆ, ์ฃผ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›„๋ณด๋ฅผ ์—ด๊ฑฐํ•œ ๋‹ค์Œ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ๋‹ค์Œ ๋‹จ๊ณ„์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ›„๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค[871, 870]. ์œ„์˜ ๋‘ ๋‹จ๊ณ„๋ฅผ ๋ฐ˜๋ณตํ•˜์—ฌ LLM์€ ๊ด€๋ จ ์ฆ๊ฑฐ๋ฅผ ์ ์ฐจ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[871, 870], ์ตœ์ข…์ ์œผ๋กœ ์ •๋‹ต์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ด์ง€๋งŒ KG์— ๋Œ€ํ•œ ํ›„๋ณด ์—ด๊ฑฐ๋Š” ๋ฐฉ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[872]. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด StructGPT[458]๋Š” KG์— ๋Œ€ํ•œ ์ „๋ฌธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์‹ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ณด๋‹ค ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ KG์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์ž‘์—…(_์˜ˆ: relation extraction ๋ฐ triple extraction)์— ๋”ฐ๋ผ ํŠน์ˆ˜ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„ํ•˜์—ฌ ํšจ์œจ์ ์ด๊ณ  ์ •ํ™•ํ•œ ๋ฐ์ดํ„ฐ ์ถ”์ถœ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ LLM์€ KG์˜ ๊ตฌ์กฐ์  ์ •๋ณด๋ฅผ ๋” ์ž˜ ์กฐ์ž‘ํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋„๋ก ์ง€์‹œํ•˜์—ฌ ์ž‘์—… ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์‹œ๋„ˆ์ง€ ์ฆ๊ฐ• LLM.** ๋ณต์žกํ•œ ์ž‘์—…(_์˜ˆ:_ ๋‹ค์ค‘ ํ™‰ ์งˆ๋ฌธ ๋‹ต๋ณ€[656])์„ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด LLM์ด ์ฒด๊ณ„์ ์ธ ์†”๋ฃจ์…˜ ๊ณ„ํš์— ๋”ฐ๋ผ KG๋ฅผ ์—ฌ๋Ÿฌ ๋ฒˆ ์ฟผ๋ฆฌํ•ด์•ผ ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹ค์ค‘ ํ„ด ์ƒํ˜ธ ์ž‘์šฉ ๋ฐฉ์‹์„ LLM์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์— _์‹œ๋„ˆ์ง€ ์ฆ๊ฐ• LLM_์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์™„์ ์ธ ๋ฐฉ์‹์œผ๋กœ LLM๊ณผ KG์„ ๋” ์ž˜ ์‹œ๋„ˆ์ง€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•˜๊ณ  ํ•„์š”ํ•œ ์ง€์‹์„ KG์—์„œ ํ™œ์šฉํ•˜์—ฌ ๊ฐ๊ฐ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค[870, 871, 458]. ์ด ๊ณผ์ •์—์„œ LLM์€ ์ž๋™์œผ๋กœ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  KG ํ™˜๊ฒฝ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ํ†ตํ•ด ์‹คํ–‰ํ•˜๋Š” ์ž์œจ ์—์ด์ „ํŠธ(8.1.6์ ˆ์—์„œ ์ž์„ธํžˆ ์„ค๋ช…)๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[870]. ํŠนํžˆ, ์ฃผ๋ฅ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ˜„์žฌ ๋‹จ๊ณ„์—์„œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ง€์‹ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›„๋ณด๋ฅผ ์—ด๊ฑฐํ•œ ๋‹ค์Œ ์งˆ๋ฌธ์— ๋”ฐ๋ผ ๋‹ค์Œ ๋‹จ๊ณ„์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ํ›„๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค[871, 870]. ์œ„์˜ ๋‘ ๋‹จ๊ณ„๋ฅผ ๋ฐ˜๋ณตํ•˜์—ฌ LLM์€ ๊ด€๋ จ ์ฆ๊ฑฐ๋ฅผ ์ ์ฐจ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[871, 870], ์ตœ์ข…์ ์œผ๋กœ ์ •๋‹ต์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ด์ง€๋งŒ KG์— ๋Œ€ํ•œ ํ›„๋ณด ์—ด๊ฑฐ๋Š” ๋ฐฉ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ณต๊ฐ„์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[872]. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด StructGPT [458]๋Š” KG์— ๋Œ€ํ•œ ์ „๋ฌธ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ง€์‹ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜๋Š” ๋ณด๋‹ค ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ KG์— ๋Œ€ํ•œ ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์ž‘์—…(_์˜ˆ: ๊ด€๊ณ„ ์ถ”์ถœ ๋ฐ ํŠธ๋ฆฌํ”Œ ์ถ”์ถœ)์— ๋”ฐ๋ผ ํŠน์ˆ˜ ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ค๊ณ„ํ•˜์—ฌ ํšจ์œจ์ ์ด๊ณ  ์ •ํ™•ํ•œ ๋ฐ์ดํ„ฐ ์ถ”์ถœ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์œผ๋กœ LLM์€ KG์˜ ๊ตฌ์กฐ์  ์ •๋ณด๋ฅผ ๋” ์ž˜ ์กฐ์ž‘ํ•˜๊ณ  ์ฒ˜๋ฆฌํ•˜๋„๋ก ์ง€์‹œํ•˜์—ฌ ์ž‘์—… ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**Human Labeler Selection.** In existing work, the dominant method for generating human feedback data is human annotation [66, 367, 116]. This highlights the critical role of selecting appropriate human labelers. To provide high-quality feedback, human labelers are supposed to have a qualified level of education and excellent proficiency in English. For example, Sparrow [116] requires human labelers to be UK-based native English speakers who have obtained at least an undergraduate-level educational qualification. Even then, several studies [367] have found that there still exists a mismatch between the intentions of researchers and human labelers, which may lead to low-quality human feedback and cause LLMs to produce unexpected output. To address this issue, InstructGPT [66] further conducts a screening process to filter labelers by assessing the agreement between human labelers and researchers. Specifically, researchers first label a small amount of data and then measure the agreement between themselves and human labelers. The labelers with the highest agreement will be selected to proceed with the subsequent annotation work. In some other work [370], "super raters" are used to ensure the high quality of human feedback. Researchers evaluate the performance of human labelers and select a group of well-performing human labelers (_e.g._, high agreement) as super raters. The super raters will be given priority to collaborate with the researchers in the subsequent study. When human labelers annotate the output of LLMs, it is helpful to specify detailed instructions and provide instant guidance for human labelers, which can further regulate the annotation of labelers.
**Human Labeler Selection**. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฃผ๋œ ๋ฐฉ๋ฒ•์€ ์ธ๊ฐ„ annotation [66, 367, 116]์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ ์ ˆํ•œ ์ธ๊ฐ„ labeler๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ธ๊ฐ„ labeler๊ฐ€ ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ต์œก์„ ๋ฐ›๊ณ  ์˜์–ด์— ๋Šฅํ†ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Sparrow [116]๋Š” ์ธ๊ฐ„ labeler๊ฐ€ ์ตœ์†Œ ํ•™๋ถ€ ์ˆ˜์ค€์˜ ๊ต์œก ์ž๊ฒฉ์„ ํš๋“ํ•œ ์˜๊ตญ์— ๊ฑฐ์ฃผํ•˜๋Š” ์˜์–ด ์›์–ด๋ฏผ์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ[367]์— ๋”ฐ๋ฅด๋ฉด ์—ฐ๊ตฌ์ž์™€ ์ธ๊ฐ„ labeler ๊ฐ„์˜ ์˜๋„๊ฐ€ ์—ฌ์ „ํžˆ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ ์ด์–ด์ ธ LLM์ด ์˜ˆ๊ธฐ์น˜ ์•Š์€ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด InstructGPT [66]๋Š” ์ธ๊ฐ„ labeler์™€ ์—ฐ๊ตฌ์ž ๊ฐ„์˜ ๋™์˜๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ labeler๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” screening ๊ณผ์ •์„ ์ถ”๊ฐ€๋กœ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์—ฐ๊ตฌ์ž๋Š” ๋จผ์ € ์†Œ๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ์— label์„ ๋ถ™์ธ ๋‹ค์Œ ์ž์‹ ๊ณผ ์ธ๊ฐ„ labeler ๊ฐ„์˜ ์ผ์น˜ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ์น˜๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ labeler๊ฐ€ ์„ ์ •๋˜์–ด ํ›„์† annotation ์ž‘์—…์„ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ผ๋ถ€ ์—ฐ๊ตฌ[370]์—์„œ๋Š” 'super raters'๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„ labeler์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ์ธ๊ฐ„ labeler ๊ทธ๋ฃน(์˜ˆ: ์ผ์น˜๋„๊ฐ€ ๋†’์Œ)์„ super raters๋กœ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค. super raters๋Š” ํ›„์† ์—ฐ๊ตฌ์—์„œ ์—ฐ๊ตฌ์ž์™€ ํ˜‘์—…ํ•  ์ˆ˜ ์žˆ๋Š” ์šฐ์„ ๊ถŒ์„ ๋ถ€์—ฌ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ธ๊ฐ„ labeler๊ฐ€ LLM์˜ ์ถœ๋ ฅ์— annotation์„ ๋‹ฌ ๋•Œ, ์ƒ์„ธํ•œ instructions์„ ๋ช…์‹œํ•˜๊ณ  ์ธ๊ฐ„ labeler์—๊ฒŒ ์ฆ‰๊ฐ์ ์ธ ์•ˆ๋‚ด๋ฅผ ์ œ๊ณตํ•˜๋ฉด labeler์˜ annotation์„ ๋”์šฑ ๊ทœ์ œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ ์„ ํƒ**. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์ฃผ๋œ ๋ฐฉ๋ฒ•์€ ์ธ๊ฐ„ ์ฃผ์„[66, 367, 116]์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ ์ ˆํ•œ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ๋ฅผ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ํ”ผ๋“œ๋ฐฑ์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ๊ฐ€ ์ ์ ˆํ•œ ์ˆ˜์ค€์˜ ๊ต์œก์„ ๋ฐ›๊ณ  ์˜์–ด์— ๋Šฅํ†ตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Sparrow [116]๋Š” ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ๊ฐ€ ์ตœ์†Œ ํ•™๋ถ€ ์ˆ˜์ค€์˜ ๊ต์œก ์ž๊ฒฉ์„ ํš๋“ํ•œ ์˜๊ตญ์— ๊ฑฐ์ฃผํ•˜๋Š” ์˜์–ด ์›์–ด๋ฏผ์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ[367]์— ๋”ฐ๋ฅด๋ฉด ์—ฐ๊ตฌ์ž์™€ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ ๊ฐ„์˜ ์˜๋„๊ฐ€ ์—ฌ์ „ํžˆ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ ์ด์–ด์ ธ LLM์ด ์˜ˆ๊ธฐ์น˜ ์•Š์€ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด InstructGPT [66]๋Š” ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์™€ ์—ฐ๊ตฌ์ž ๊ฐ„์˜ ๋™์˜๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ ๋ผ๋ฒจ๋Ÿฌ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” ์„ ๋ณ„ ๊ณผ์ •์„ ์ถ”๊ฐ€๋กœ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ์—ฐ๊ตฌ์ž๋Š” ๋จผ์ € ์†Œ๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ์— ๋ผ๋ฒจ์„ ๋ถ™์ธ ๋‹ค์Œ ์ž์‹ ๊ณผ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ ๊ฐ„์˜ ์ผ์น˜ ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ์ผ์น˜๋„๊ฐ€ ๊ฐ€์žฅ ๋†’์€ ๋ผ๋ฒจ๋Ÿฌ๊ฐ€ ์„ ์ •๋˜์–ด ํ›„์† ์ฃผ์„ ์ž‘์—…์„ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ผ๋ถ€ ์—ฐ๊ตฌ[370]์—์„œ๋Š” '์Šˆํผ ๋ ˆ์ดํ„ฐ'๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์˜ ํ’ˆ์งˆ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์ž๋“ค์€ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•œ ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ ๊ทธ๋ฃน(์˜ˆ: ์ผ์น˜๋„๊ฐ€ ๋†’์Œ)์„ ์Šˆํผ ๋ ˆ์ดํ„ฐ๋กœ ์„ ์ •ํ•ฉ๋‹ˆ๋‹ค. ์Šˆํผ ๋ ˆ์ดํ„ฐ๋Š” ํ›„์† ์—ฐ๊ตฌ์—์„œ ์—ฐ๊ตฌ์ž์™€ ํ˜‘์—…ํ•  ์ˆ˜ ์žˆ๋Š” ์šฐ์„ ๊ถŒ์„ ๋ถ€์—ฌ๋ฐ›์Šต๋‹ˆ๋‹ค. ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ๊ฐ€ LLM์˜ ์ถœ๋ ฅ์— ์ฃผ์„์„ ๋‹ฌ ๋•Œ, ์ƒ์„ธํ•œ ์ง€์นจ์„ ๋ช…์‹œํ•˜๊ณ  ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์—๊ฒŒ ์ฆ‰๊ฐ์ ์ธ ์•ˆ๋‚ด๋ฅผ ์ œ๊ณตํ•˜๋ฉด ๋ผ๋ฒจ๋Ÿฌ์˜ ์ฃผ์„์„ ๋”์šฑ ๊ทœ์ œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
In designing the Yi model series, we are mostly concerned on the following dimensions regarding _model scale, data scale, and data quality_: (1). when choosing model scale, the desiderata is to have small enough model that is feasible for inference on consumer-grade hardware like the RTX 4090 where the bounding factor is its limited 24G memory, yet still large enough with complex reasoning and emergent abilities. This is why we found 34B gives a nice performance-cost balance; (2). since 34B is smaller than the conventional 70B used by Chinchilla [30] and LLaMA [77], we increase the pretrain data scale to 3.1T tokens to compensate for the decreased compute flops. This makes the model-data scale combination fall into the post Chinchilla optimal regime [64], i.e., we overtrain the model on more tokens (3T) than the compute optimal (around 1T). The benefit is from the inference side, as we achieve stronger performance with reduced serving cost: after int4 [81] quantization, one can serve the 34B chat model on 24G GPU memory with almost no performance drop; (3). our data engineering principle is to promote quality over quantity for both pretraining and finetuning. The pretraining data quality is guaranteed by a sophisticated data cleaning pipeline with cascaded filtering methods and intentionally increased deduplication strength; (4). for finetuning data we heavily emphasize quality by handcrafting less than 10K instructions over multiple iterations based on user feedback. This approach significantly deviates from the quantity-scaling styled instruction tuning works like FLAN [9] and UltraChat [19], but aligns more with the handcrafting styled works like LIMA [94].
Yi ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋ฅผ ์„ค๊ณ„ํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ์ฃผ๋กœ ๋ชจ๋ธ ๊ทœ๋ชจ, ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ ๋ฐ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ_์— ๊ด€ํ•œ ๋‹ค์Œ ์ฐจ์›์— ๊ด€์‹ฌ์„ ๊ธฐ์šธ์˜€์Šต๋‹ˆ๋‹ค: (1). ๋ชจ๋ธ ๊ทœ๋ชจ๋ฅผ ์„ ํƒํ•  ๋•Œ, ๋ฐ”์šด๋”ฉ ์š”์ธ์ด 24G ๋ฉ”๋ชจ๋ฆฌ ์ œํ•œ์ด ์žˆ๋Š” RTX 4090๊ณผ ๊ฐ™์€ ์†Œ๋น„์ž์šฉ ํ•˜๋“œ์›จ์–ด์—์„œ ์ถ”๋ก ์— ์ ํ•ฉํ•œ ์ถฉ๋ถ„ํžˆ ์ž‘์€ ๋ชจ๋ธ์ด๋ฉด์„œ๋„ ๋ณต์žกํ•œ ์ถ”๋ก ๊ณผ ์ฐฝ๋ฐœ์  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ์„ ๊ฐ–์ถ”๋Š” ๊ฒƒ์ด ๋ฐ”๋žŒ์งํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋ฐ”๋กœ 34B๊ฐ€ ํ›Œ๋ฅญํ•œ ์„ฑ๋Šฅ-๋น„์šฉ ๊ท ํ˜•์„ ์ œ๊ณตํ•˜๋Š” ์ด์œ ์ž…๋‹ˆ๋‹ค; (2). 34B๋Š” Chinchilla [30] ๋ฐ LLaMA [77]์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์กด 70B๋ณด๋‹ค ์ž‘๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๋ฅผ 3.1T ํ† ํฐ์œผ๋กœ ๋Š˜๋ ค ๊ณ„์‚ฐ ํ”Œ๋กญ ๊ฐ์†Œ๋ฅผ ๋ณด์™„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋ชจ๋ธ-๋ฐ์ดํ„ฐ ๊ทœ๋ชจ ์กฐํ•ฉ์ด post Chinchilla optimal regime, ์ฆ‰ ๊ณ„์‚ฐ ์ตœ์ (์•ฝ 1T)๋ณด๋‹ค ๋” ๋งŽ์€ ํ† ํฐ(3T)์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณผํ•™ํ•˜๊ฒŒ ํ•™์Šตํ•˜๋Š” post Chinchilla optimal regime์— ์†ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์„œ๋น„์Šค ๋น„์šฉ ์ ˆ๊ฐ: 34B ์ฑ„ํŒ… ๋ชจ๋ธ์„ int4 [81] Quantization ํ›„ ๊ฑฐ์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ์—†์ด 24G GPU ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์„œ๋น„์Šคํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ถ”๋ก  ์ธก๋ฉด์—์„œ ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค; (3). ์‚ฌ์ „ ํ•™์Šต ๋ฐ Fine-tuning ๋ชจ๋‘์— ๋Œ€ํ•ด ์–‘๋ณด๋‹ค ์งˆ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ์›์น™์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์€ ์บ์Šค์ผ€์ด๋“œ ํ•„ํ„ฐ๋ง ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ํŒŒ์ดํ”„๋ผ์ธ๊ณผ ์˜๋„์ ์œผ๋กœ ์ค‘๋ณต ์ œ๊ฑฐ ๊ฐ•๋„๋ฅผ ๋†’์—ฌ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค; (4). Fine-tuning ๋ฐ์ดํ„ฐ์˜ ๊ฒฝ์šฐ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด 10,000๊ฐœ ๋ฏธ๋งŒ์˜ Instruction์„ ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑํ•˜์—ฌ ํ’ˆ์งˆ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ FLAN [9] ๋ฐ UltraChat [19]๊ณผ ๊ฐ™์€ ์–‘์  ํ™•์žฅ ์Šคํƒ€์ผ์˜ Instruction Tuning ์ž‘์—…์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€๋งŒ LIMA [94]์™€ ๊ฐ™์€ ์ˆ˜์ž‘์—… ์Šคํƒ€์ผ์˜ ์ž‘์—…๊ณผ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค.
Yi ๋ชจ๋ธ ์‹œ๋ฆฌ์ฆˆ๋ฅผ ์„ค๊ณ„ํ•  ๋•Œ, ์šฐ๋ฆฌ๋Š” ์ฃผ๋กœ ๋ชจ๋ธ ๊ทœ๋ชจ, ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ ๋ฐ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ_์— ๊ด€ํ•œ ๋‹ค์Œ ์ฐจ์›์— ๊ด€์‹ฌ์„ ๊ธฐ์šธ์˜€์Šต๋‹ˆ๋‹ค: (1). ๋ชจ๋ธ ๊ทœ๋ชจ๋ฅผ ์„ ํƒํ•  ๋•Œ, ๋ฐ”์šด๋”ฉ ์š”์ธ์ด 24G ๋ฉ”๋ชจ๋ฆฌ ์ œํ•œ์ด ์žˆ๋Š” RTX 4090๊ณผ ๊ฐ™์€ ์†Œ๋น„์ž์šฉ ํ•˜๋“œ์›จ์–ด์—์„œ ์ถ”๋ก ์— ์ ํ•ฉํ•œ ์ถฉ๋ถ„ํžˆ ์ž‘์€ ๋ชจ๋ธ์ด๋ฉด์„œ๋„ ๋ณต์žกํ•œ ์ถ”๋ก ๊ณผ ์ฐฝ๋ฐœ์  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด ์ถฉ๋ถ„ํžˆ ํฐ ๋ชจ๋ธ์„ ๊ฐ–์ถ”๋Š” ๊ฒƒ์ด ๋ฐ”๋žŒ์งํ•ฉ๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋ฐ”๋กœ 34B๊ฐ€ ํ›Œ๋ฅญํ•œ ์„ฑ๋Šฅ-๋น„์šฉ ๊ท ํ˜•์„ ์ œ๊ณตํ•˜๋Š” ์ด์œ ์ž…๋‹ˆ๋‹ค; (2). 34B๋Š” Chinchilla [30] ๋ฐ LLaMA [77]์—์„œ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์กด 70B๋ณด๋‹ค ์ž‘๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ๊ทœ๋ชจ๋ฅผ 3.1T ํ† ํฐ์œผ๋กœ ๋Š˜๋ ค ๊ณ„์‚ฐ ํ”Œ๋กญ ๊ฐ์†Œ๋ฅผ ๋ณด์™„ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ๋ชจ๋ธ-๋ฐ์ดํ„ฐ ๊ทœ๋ชจ ์กฐํ•ฉ์ด ์นญ์น ๋ผ ์ตœ์  ๋ ˆ์ง, ์ฆ‰ ๊ณ„์‚ฐ ์ตœ์ (์•ฝ 1T)๋ณด๋‹ค ๋” ๋งŽ์€ ํ† ํฐ(3T)์— ๋Œ€ํ•ด ๋ชจ๋ธ์„ ๊ณผํ•™ํ•˜๊ฒŒ ํ•™์Šตํ•˜๋Š” ํฌ์ŠคํŠธ ์นญ์น ๋ผ ์ตœ์  ๋ ˆ์ง์— ์†ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์„œ๋น„์Šค ๋น„์šฉ ์ ˆ๊ฐ: 34B ์ฑ„ํŒ… ๋ชจ๋ธ์„ int4 [81] ์–‘์žํ™” ํ›„ ๊ฑฐ์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ์—†์ด 24G GPU ๋ฉ”๋ชจ๋ฆฌ์—์„œ ์„œ๋น„์Šคํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ถ”๋ก  ์ธก๋ฉด์—์„œ ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค; (3). ์‚ฌ์ „ ํ•™์Šต ๋ฐ ๋ฏธ์„ธ ์กฐ์ • ๋ชจ๋‘์— ๋Œ€ํ•ด ์–‘๋ณด๋‹ค ์งˆ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด ๋ฐ์ดํ„ฐ ์—”์ง€๋‹ˆ์–ด๋ง ์›์น™์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์€ ์บ์Šค์ผ€์ด๋“œ ํ•„ํ„ฐ๋ง ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ •๊ตํ•œ ๋ฐ์ดํ„ฐ ์ •๋ฆฌ ํŒŒ์ดํ”„๋ผ์ธ๊ณผ ์˜๋„์ ์œผ๋กœ ์ค‘๋ณต ์ œ๊ฑฐ ๊ฐ•๋„๋ฅผ ๋†’์—ฌ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค; (4). ํŒŒ์ธํŠœ๋‹ ๋ฐ์ดํ„ฐ์˜ ๊ฒฝ์šฐ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—ฌ๋Ÿฌ ๋ฐ˜๋ณต์„ ํ†ตํ•ด 10,000๊ฐœ ๋ฏธ๋งŒ์˜ ๋ช…๋ น์–ด๋ฅผ ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑํ•˜์—ฌ ํ’ˆ์งˆ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ FLAN [9] ๋ฐ UltraChat [19]๊ณผ ๊ฐ™์€ ์–‘์  ํ™•์žฅ ์Šคํƒ€์ผ์˜ ๋ช…๋ น์–ด ํŠœ๋‹ ์ž‘์—…์—์„œ ํฌ๊ฒŒ ๋ฒ—์–ด๋‚˜์ง€๋งŒ LIMA [94]์™€ ๊ฐ™์€ ์ˆ˜์ž‘์—… ์Šคํƒ€์ผ์˜ ์ž‘์—…๊ณผ ๋” ์ž˜ ๋ถ€ํ•ฉํ•ฉ๋‹ˆ๋‹ค.
Code.Different from Math, synthetic data for code reasoning can naturally combine the execution results with structured code, as one requirement of correct code is being executable. In coding-enhanced models, CodeRL (Le et al., 2022) presents an actor-critic approach to improve pretrained language models with feedback signals on synthetic code samples. Haluptzok et al. (2022) propose a self-improvement strategy where the models generate their own synthetic puzzle-solution pairs. These pairs are then verified and filtered by a real interpreter before being used to finetune language models. Shypula et al. (2023) further propose a framework that leverages a simulated environment and adaptation strategies like self-improvement synthetic data generation and CoT prompting for code optimization. Yang et al. (2024) developed InterCode, a framework designed to enhance interactive code generation within a reinforcement learning environment, where code serves as actions and execution feedback serves as observations. Reflexion (Shinn et al., 2024) employs external or internally simulated linguistic feedback signals to improve the code reasoning capabilities of language models. Regarding synthetic SFT data, Code Alpaca comprises a dataset of 20K code instructions automatically generated by applying SELF-INSTRUCT (Wang et al., 2022) to ChatGPT across 21 seed tasks. WizardCoder (Luo et al., 2023) introduces Code Evol-Instruct to guide ChatGPT with heuristicprompts to enhance the complexity and diversity of synthetic data. Meanwhile, Magicoder (Wei et al., 2023c) developed OSS-INSTRUCT, which generates 75K diverse synthetic instruction samples from open-source code snippets.
์ฝ”๋“œ. ์ˆ˜ํ•™๊ณผ ๋‹ฌ๋ฆฌ ์ฝ”๋“œ ์ถ”๋ก ์„ ์œ„ํ•œ synthetic data๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ฝ”๋“œ์˜ ํ•œ ๊ฐ€์ง€ ์š”๊ตฌ ์‚ฌํ•ญ์ธ ์‹คํ–‰ ๊ฒฐ๊ณผ๋ฅผ ๊ตฌ์กฐํ™”๋œ ์ฝ”๋“œ์™€ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๊ฒฐํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋”ฉ ๊ฐ•ํ™” ๋ชจ๋ธ์—์„œ CodeRL(Le et al., 2022)์€ synthetic code ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ actor-critic ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. Haluptzok et al.(2022)์€ ๋ชจ๋ธ์ด ์ž์ฒด์ ์œผ๋กœ synthetic puzzle-solution ์Œ์„ ์ƒ์„ฑํ•˜๋Š” self-improvement ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์Œ์€ ์‹ค์ œ ์ธํ„ฐํ”„๋ฆฌํ„ฐ์— ์˜ํ•ด ๊ฒ€์ฆ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ํ›„ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Shypula et al.(2023)์€ ์ฝ”๋“œ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ๊ณผ self-improvement synthetic data generation ๋ฐ CoT prompting๊ณผ ๊ฐ™์€ ์ ์‘ ์ „๋žต์„ ํ™œ์šฉํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋”์šฑ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Yang et al.(2024)์€ reinforcement learning ํ™˜๊ฒฝ ๋‚ด์—์„œ ๋Œ€ํ™”ํ˜• ์ฝ”๋“œ ์ƒ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ InterCode๋ฅผ ๊ฐœ๋ฐœํ–ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฝ”๋“œ๋Š” ์•ก์…˜์œผ๋กœ, ์‹คํ–‰ ํ”ผ๋“œ๋ฐฑ์€ ๊ด€์ฐฐ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Reflexion(Shinn et al., 2024)์€ ์™ธ๋ถ€ ๋˜๋Š” ๋‚ด๋ถ€์ ์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ฝ”๋“œ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. synthetic SFT ๋ฐ์ดํ„ฐ์™€ ๊ด€๋ จํ•˜์—ฌ Code Alpaca๋Š” 21๊ฐœ์˜ ์‹œ๋“œ ์ž‘์—…์— ๊ฑธ์ณ ChatGPT์— SELF-INSTRUCT(Wang et al., 2022)๋ฅผ ์ ์šฉํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑ๋œ 20,000๊ฐœ์˜ ์ฝ”๋“œ ๋ช…๋ น์–ด๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. WizardCoder(Luo et al., 2023)๋Š” synthetic data์˜ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด heuristic prompts๋กœ ChatGPT๋ฅผ ์•ˆ๋‚ดํ•˜๋Š” Code Evol-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, Magicoder(Wei et al., 2023c)๋Š” ์˜คํ”ˆ ์†Œ์Šค ์ฝ”๋“œ ์Šค๋‹ˆํŽซ์—์„œ 75,000๊ฐœ์˜ ๋‹ค์–‘ํ•œ synthetic instruction ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๋Š” OSS-INSTRUCT๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
์ฝ”๋“œ. ์ˆ˜ํ•™๊ณผ ๋‹ฌ๋ฆฌ ์ฝ”๋“œ ์ถ”๋ก ์„ ์œ„ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์˜ฌ๋ฐ”๋ฅธ ์ฝ”๋“œ์˜ ํ•œ ๊ฐ€์ง€ ์š”๊ตฌ ์‚ฌํ•ญ์ธ ์‹คํ–‰ ๊ฒฐ๊ณผ๋ฅผ ๊ตฌ์กฐํ™”๋œ ์ฝ”๋“œ์™€ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ๊ฒฐํ•ฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋”ฉ ๊ฐ•ํ™” ๋ชจ๋ธ์—์„œ CodeRL(Le et al., 2022)์€ ํ•ฉ์„ฑ ์ฝ”๋“œ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ์•กํ„ฐ-๋น„ํ‰๊ฐ€ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. Haluptzok et al.(2022)์€ ๋ชจ๋ธ์ด ์ž์ฒด์ ์œผ๋กœ ํ•ฉ์„ฑ ํผ์ฆ-ํ•ด๊ฒฐ ์Œ์„ ์ƒ์„ฑํ•˜๋Š” ์ž์ฒด ๊ฐœ์„  ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์Œ์€ ์‹ค์ œ ์ธํ„ฐํ”„๋ฆฌํ„ฐ์— ์˜ํ•ด ๊ฒ€์ฆ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ํ›„ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Shypula et al.(2023)์€ ์ฝ”๋“œ ์ตœ์ ํ™”๋ฅผ ์œ„ํ•ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ๊ณผ ์ž์ฒด ๊ฐœ์„  ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ ๋ฐ CoT ํ”„๋กฌํ”„ํŠธ์™€ ๊ฐ™์€ ์ ์‘ ์ „๋žต์„ ํ™œ์šฉํ•˜๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋”์šฑ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Yang et al.(2024)์€ ๊ฐ•ํ™” ํ•™์Šต ํ™˜๊ฒฝ ๋‚ด์—์„œ ๋Œ€ํ™”ํ˜• ์ฝ”๋“œ ์ƒ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ InterCode๋ฅผ ๊ฐœ๋ฐœํ–ˆ์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ ์ฝ”๋“œ๋Š” ์•ก์…˜์œผ๋กœ, ์‹คํ–‰ ํ”ผ๋“œ๋ฐฑ์€ ๊ด€์ฐฐ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. Reflexion(Shinn et al., 2024)์€ ์™ธ๋ถ€ ๋˜๋Š” ๋‚ด๋ถ€์ ์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์–ธ์–ด ํ”ผ๋“œ๋ฐฑ ์‹ ํ˜ธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ฝ”๋“œ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ์™€ ๊ด€๋ จํ•˜์—ฌ Code Alpaca๋Š” 21๊ฐœ์˜ ์‹œ๋“œ ์ž‘์—…์— ๊ฑธ์ณ ChatGPT์— SELF-INSTRUCT(Wang et al., 2022)๋ฅผ ์ ์šฉํ•˜์—ฌ ์ž๋™์œผ๋กœ ์ƒ์„ฑ๋œ 20,000๊ฐœ์˜ ์ฝ”๋“œ ๋ช…๋ น์–ด๋กœ ๊ตฌ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. WizardCoder(Luo et al., 2023)๋Š” ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ํœด๋ฆฌ์Šคํ‹ฑ ํ”„๋กฌํ”„ํŠธ๋กœ ChatGPT๋ฅผ ์•ˆ๋‚ดํ•˜๋Š” Code Evol-Instruct๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ, Magicoder(Wei et al., 2023c)๋Š” ์˜คํ”ˆ ์†Œ์Šค ์ฝ”๋“œ ์Šค๋‹ˆํŽซ์—์„œ 75,000๊ฐœ์˜ ๋‹ค์–‘ํ•œ ํ•ฉ์„ฑ ๋ช…๋ น์–ด ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๋Š” OSS-INSTRUCT๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_GPT-4V, GPT-4 turbo, and beyond_. Based on the work done for GPT-4 [46], OpenAI further released GPT-4V in September 2023, which focused on the safe deployment of the vision capabilities of GPT-4. In the GPT-4V's system card [133], it has extensively discussed the assessment and mitigation of risks related to visually augmented inputs. Specially, GPT-4V exhibited strong vision capacities in various application scenarios, showing the great potential as a powerful multimodal learning system. More recently, in November 2023, OpenAI released an upgraded generation of GPT-4 model at DevDay, named _GPT-4 Turbo_, with a series of technical improvements. GPT-4 Turbo is featured by the improved model capacity (more capable than GPT-4), the extended knowledge source (up to April 2023), long context window (up to 128k tokens), optimized model performance (cheaper price), and other useful functionality updates (function call, reproducible outputs, etc.). At the same time, Assistants API was launched to ease the rapid development of agent-like assistants. With this API, developers can easily create goal-oriented assistants within their applications, by leveraging specific instruction, extra knowledge and tool use. Furthermore, multimodal capacities (see, hear, and speak) were also enhanced in this new release, supported by GPT-4 Turbo with vision, DALL-E 3, Text-to-speech (TTS), and Listen to voice samples. These improvements have greatly extended the capacity scope and enhanced the task performance of GPT models. More importantly, the application ecosystem will be greatly strengthened with the technology upgrade in improved models, APIs, and functionalities.
\(\bullet\)_GPT-4V, GPT-4 Turbo ๋ฐ ๊ทธ ์ด์ƒ_. OpenAI๋Š” GPT-4 [46]์— ๋Œ€ํ•œ ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ 2023๋…„ 9์›”์— GPT-4์˜ ๋น„์ „ ๊ธฐ๋Šฅ์˜ ์•ˆ์ „ํ•œ ๋ฐฐํฌ์— ์ค‘์ ์„ ๋‘” GPT-4V๋ฅผ ์ถ”๊ฐ€๋กœ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4V์˜ ์‹œ์Šคํ…œ ์นด๋“œ [133]์—์„œ๋Š” ์‹œ๊ฐ์ ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ๋œ ์œ„ํ—˜์˜ ํ‰๊ฐ€ ๋ฐ ์™„ํ™”์— ๋Œ€ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ GPT-4V๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๊ฐ•๋ ฅํ•œ ๋น„์ „ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์‹œ์Šคํ…œ์œผ๋กœ์„œ ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์—๋Š” 2023๋…„ 11์›”์— OpenAI๊ฐ€ DevDay์—์„œ ์ผ๋ จ์˜ ๊ธฐ์ˆ  ๊ฐœ์„ ์„ ํ†ตํ•ด _GPT-4 Turbo_๋ผ๋Š” ์ด๋ฆ„์˜ ์—…๊ทธ๋ ˆ์ด๋“œ๋œ GPT-4 ๋ชจ๋ธ ์„ธ๋Œ€๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4 Turbo๋Š” ํ–ฅ์ƒ๋œ ๋ชจ๋ธ ์šฉ๋Ÿ‰(GPT-4๋ณด๋‹ค ๋” ๊ฐ•๋ ฅํ•จ), ํ™•์žฅ๋œ ์ง€์‹ ์†Œ์Šค(์ตœ๋Œ€ 2023๋…„ 4์›”๊นŒ์ง€), ๊ธด context window(์ตœ๋Œ€ 128k tokens), ์ตœ์ ํ™”๋œ ๋ชจ๋ธ ์„ฑ๋Šฅ(๋” ์ €๋ ดํ•œ ๊ฐ€๊ฒฉ), ๊ธฐํƒ€ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ ์—…๋ฐ์ดํŠธ(function call, reproducible outputs ๋“ฑ)๊ฐ€ ํŠน์ง•์ž…๋‹ˆ๋‹ค. ๋™์‹œ์— ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋„๋ก Assistants API๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด API๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ํŠน์ • instruction, ์ถ”๊ฐ€ ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ์„ ํ™œ์šฉํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฒˆ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ(๋ณด๊ธฐ, ๋“ฃ๊ธฐ, ๋งํ•˜๊ธฐ)๋„ ๊ฐ•ํ™”๋˜์–ด ๋น„์ „, DALL-E 3, Text-to-speech (TTS), ์Œ์„ฑ ์ƒ˜ํ”Œ ๋“ฃ๊ธฐ ๋“ฑ์œผ๋กœ GPT-4 Turbo๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์œผ๋กœ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜๊ณ  ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๊ฐœ์„ ๋œ ๋ชจ๋ธ, API ๋ฐ ๊ธฐ๋Šฅ์˜ ๊ธฐ์ˆ  ์—…๊ทธ๋ ˆ์ด๋“œ๋กœ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์ด ํฌ๊ฒŒ ๊ฐ•ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
\(\bullet\)_GPT-4V, GPT-4 ํ„ฐ๋ณด ๋ฐ ๊ทธ ์ด์ƒ_. OpenAI๋Š” GPT-4 [46]์— ๋Œ€ํ•œ ์ž‘์—…์„ ๊ธฐ๋ฐ˜์œผ๋กœ 2023๋…„ 9์›”์— GPT-4์˜ ๋น„์ „ ๊ธฐ๋Šฅ์˜ ์•ˆ์ „ํ•œ ๋ฐฐํฌ์— ์ค‘์ ์„ ๋‘” GPT-4V๋ฅผ ์ถ”๊ฐ€๋กœ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4V์˜ ์‹œ์Šคํ…œ ์นด๋“œ [133]์—์„œ๋Š” ์‹œ๊ฐ์ ์œผ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ๋œ ์œ„ํ—˜์˜ ํ‰๊ฐ€ ๋ฐ ์™„ํ™”์— ๋Œ€ํ•ด ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ๋…ผ์˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ GPT-4V๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๊ฐ•๋ ฅํ•œ ๋น„์ „ ์—ญ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ๋ฉฐ ๊ฐ•๋ ฅํ•œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ํ•™์Šต ์‹œ์Šคํ…œ์œผ๋กœ์„œ ํฐ ์ž ์žฌ๋ ฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์—๋Š” 2023๋…„ 11์›”์— OpenAI๊ฐ€ DevDay์—์„œ ์ผ๋ จ์˜ ๊ธฐ์ˆ  ๊ฐœ์„ ์„ ํ†ตํ•ด _GPT-4 ํ„ฐ๋ณด_๋ผ๋Š” ์ด๋ฆ„์˜ ์—…๊ทธ๋ ˆ์ด๋“œ๋œ GPT-4 ๋ชจ๋ธ ์„ธ๋Œ€๋ฅผ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. GPT-4 ํ„ฐ๋ณด๋Š” ํ–ฅ์ƒ๋œ ๋ชจ๋ธ ์šฉ๋Ÿ‰(GPT-4๋ณด๋‹ค ๋” ๊ฐ•๋ ฅํ•จ), ํ™•์žฅ๋œ ์ง€์‹ ์†Œ์Šค(์ตœ๋Œ€ 2023๋…„ 4์›”๊นŒ์ง€), ๊ธด ์ปจํ…์ŠคํŠธ ์ฐฝ(์ตœ๋Œ€ 128k ํ† ํฐ), ์ตœ์ ํ™”๋œ ๋ชจ๋ธ ์„ฑ๋Šฅ(๋” ์ €๋ ดํ•œ ๊ฐ€๊ฒฉ), ๊ธฐํƒ€ ์œ ์šฉํ•œ ๊ธฐ๋Šฅ ์—…๋ฐ์ดํŠธ(ํ•จ์ˆ˜ ํ˜ธ์ถœ, ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ์ถœ๋ ฅ ๋“ฑ)๊ฐ€ ํŠน์ง•์ž…๋‹ˆ๋‹ค. ๋™์‹œ์— ์—์ด์ „ํŠธ์™€ ๊ฐ™์€ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๋„๋ก ์–ด์‹œ์Šคํ„ดํŠธ API๊ฐ€ ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด API๋ฅผ ํ†ตํ•ด ๊ฐœ๋ฐœ์ž๋Š” ํŠน์ • ์ง€์นจ, ์ถ”๊ฐ€ ์ง€์‹ ๋ฐ ๋„๊ตฌ ์‚ฌ์šฉ์„ ํ™œ์šฉํ•˜์—ฌ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ๋‚ด์—์„œ ๋ชฉํ‘œ ์ง€ํ–ฅ์ ์ธ ์–ด์‹œ์Šคํ„ดํŠธ๋ฅผ ์‰ฝ๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋ฒˆ ๋ฆด๋ฆฌ์Šค์—์„œ๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๊ธฐ๋Šฅ(๋ณด๊ธฐ, ๋“ฃ๊ธฐ, ๋งํ•˜๊ธฐ)๋„ ๊ฐ•ํ™”๋˜์–ด ๋น„์ „, DALL-E 3, ํ…์ŠคํŠธ ์Œ์„ฑ ๋ณ€ํ™˜(TTS), ์Œ์„ฑ ์ƒ˜ํ”Œ ๋“ฃ๊ธฐ ๋“ฑ์œผ๋กœ GPT-4 ํ„ฐ๋ณด๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐœ์„ ์œผ๋กœ ๋ชจ๋ธ์˜ ์šฉ๋Ÿ‰ ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜๊ณ  ์ž‘์—… ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋”์šฑ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๊ฐœ์„ ๋œ ๋ชจ๋ธ, API ๋ฐ ๊ธฐ๋Šฅ์˜ ๊ธฐ์ˆ  ์—…๊ทธ๋ ˆ์ด๋“œ๋กœ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์—์ฝ”์‹œ์Šคํ…œ์ด ํฌ๊ฒŒ ๊ฐ•ํ™”๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
To test this ability, several embodied AI environments and benchmarks can be used for evaluation, described as follows. VirtualHome [606] builds a 3D simulator for household tasks such as cleaning and cooking, in which the agent can execute natural language actions generated by LLMs. ALFRED [608] includes more challenging tasks that require LLMs to accomplish compositional targets. BEHAVIOR [607] focuses on everyday chores in simulation environments and requires LLMs to generate complex solutions, _e.g._, changing the internal status of objects. Apart from restricted environments such as household tasks, a line of research work investigates the proficiency of LLM-based agents to explore open-world environments, such as Minecraft and the Internet [696, 697]. Voyager [697] introduces an automatic curriculum module that enables LLMs to continuously acquire new skills based on feedback from the environment. GITM [696] focuses on solving various challenges in Minecraft based on LLM, through task decomposition, planning, and invocation of interfaces. Based on the generated action plans or task completions, existing work either adopts the regular metrics (_e.g._, executability and correctness of the generated action plans) [694] in the benchmark or directly conducts real-world experiments and measures the success rate [698], to evaluate such ability. It has been shown that LLMs are capable in interacting with the external environment and generating accurate action plans [699]. Recently, several improvement methods have been proposed to enhance the interaction ability of LLMs, _e.g._, designing code-like prompts [530] and providing real-world grounding [698].
# ์ถœ๋ ฅ ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค๋ช…ํ•œ ์—ฌ๋Ÿฌ ๊ตฌํ˜„๋œ AI ํ™˜๊ฒฝ๊ณผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. VirtualHome [606]์€ ์ฒญ์†Œ ๋ฐ ์š”๋ฆฌ์™€ ๊ฐ™์€ ๊ฐ€์ • ๋‚ด ์ž‘์—…์„ ์œ„ํ•œ 3D ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ LLM์—์„œ ์ƒ์„ฑ๋œ ์ž์—ฐ์–ด ๋™์ž‘์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ALFRED [608]์—๋Š” LLM์ด ๊ตฌ์„ฑ์ ์ธ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•ด์•ผ ํ•˜๋Š” ๋”์šฑ ๋„์ „์ ์ธ ์ž‘์—…์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. BEHAVIOR [607]์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ผ์ƒ์ ์ธ ์ง‘์•ˆ์ผ์„ ์ค‘์ ์œผ๋กœ ํ•˜๋ฉฐ, LLM์ด ๋ฌผ์ฒด์˜ ๋‚ด๋ถ€ ์ƒํƒœ๋ฅผ ๋ณ€๊ฒฝํ•˜๋Š” ๋“ฑ ๋ณต์žกํ•œ ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์‚ฌ ์ž‘์—…๊ณผ ๊ฐ™์€ ์ œํ•œ๋œ ํ™˜๊ฒฝ ์™ธ์—๋„ Minecraft ๋ฐ ์ธํ„ฐ๋„ท๊ณผ ๊ฐ™์€ ์˜คํ”ˆ ์›”๋“œ ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ™๋ จ๋„๋ฅผ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ ์ž‘์—…์ด ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[696, 697]. Voyager [697]๋Š” ํ™˜๊ฒฝ์˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์ด ์ง€์†์ ์œผ๋กœ ์ƒˆ๋กœ์šด ๊ธฐ์ˆ ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ ์ปค๋ฆฌํ˜๋Ÿผ ๋ชจ๋“ˆ์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. GITM [696]์€ ์ž‘์—… ๋ถ„ํ•ด, ๊ณ„ํš ๋ฐ ์ธํ„ฐํŽ˜์ด์Šค ํ˜ธ์ถœ์„ ํ†ตํ•ด LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ Minecraft์—์„œ ๋‹ค์–‘ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ์•ก์…˜ ๊ณ„ํš ๋˜๋Š” ์ž‘์—… ์™„๋ฃŒ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ธฐ์กด ์ž‘์—…์€ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ผ๋ฐ˜์ ์ธ ๋ฉ”ํŠธ๋ฆญ(์˜ˆ: ์ƒ์„ฑ๋œ ์•ก์…˜ ๊ณ„ํš์˜ ์‹คํ–‰ ๊ฐ€๋Šฅ์„ฑ ๋ฐ ์ •ํ™•์„ฑ) [694]์„ ์ฑ„ํƒํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹คํ—˜์„ ์ง์ ‘ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๊ณต๋ฅ  [698]์„ ์ธก์ •ํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. LLM์ด ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๊ณ  ์ •ํ™•ํ•œ ์•ก์…˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[699]. ์ตœ๊ทผ์—๋Š” ์ฝ”๋“œ์™€ ๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ ์„ค๊ณ„[530] ๋ฐ ์‹ค์ œ ๊ทธ๋ผ์šด๋”ฉ ์ œ๊ณต[698]๊ณผ ๊ฐ™์ด LLM์˜ ์ƒํ˜ธ ์ž‘์šฉ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค๋ช…ํ•œ ์—ฌ๋Ÿฌ ๊ตฌํ˜„๋œ AI ํ™˜๊ฒฝ๊ณผ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. VirtualHome [606]์€ ์ฒญ์†Œ ๋ฐ ์š”๋ฆฌ์™€ ๊ฐ™์€ ๊ฐ€์ • ๋‚ด ์ž‘์—…์„ ์œ„ํ•œ 3D ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ๋ฅผ ๊ตฌ์ถ•ํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ LLM์—์„œ ์ƒ์„ฑ๋œ ์ž์—ฐ์–ด ๋™์ž‘์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ALFRED [608]์—๋Š” LLM์ด ๊ตฌ์„ฑ์ ์ธ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•ด์•ผ ํ•˜๋Š” ๋”์šฑ ๋„์ „์ ์ธ ์ž‘์—…์ด ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. BEHAVIOR [607]์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ™˜๊ฒฝ์—์„œ ์ผ์ƒ์ ์ธ ์ง‘์•ˆ์ผ์„ ์ค‘์ ์œผ๋กœ ํ•˜๋ฉฐ, LLM์ด ๋ฌผ์ฒด์˜ ๋‚ด๋ถ€ ์ƒํƒœ๋ฅผ ๋ณ€๊ฒฝํ•˜๋Š” ๋“ฑ ๋ณต์žกํ•œ ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•˜๋„๋ก ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๊ฐ€์‚ฌ ์ž‘์—…๊ณผ ๊ฐ™์€ ์ œํ•œ๋œ ํ™˜๊ฒฝ ์™ธ์—๋„ ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ ๋ฐ ์ธํ„ฐ๋„ท๊ณผ ๊ฐ™์€ ์˜คํ”ˆ ์›”๋“œ ํ™˜๊ฒฝ์„ ํƒ์ƒ‰ํ•˜๋Š” LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ์ˆ™๋ จ๋„๋ฅผ ์กฐ์‚ฌํ•˜๋Š” ์—ฐ๊ตฌ ์ž‘์—…์ด ์ง„ํ–‰๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[696, 697]. Voyager [697]๋Š” ํ™˜๊ฒฝ์˜ ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์ด ์ง€์†์ ์œผ๋กœ ์ƒˆ๋กœ์šด ๊ธฐ์ˆ ์„ ์Šต๋“ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ ์ปค๋ฆฌํ˜๋Ÿผ ๋ชจ๋“ˆ์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. GITM [696]์€ ์ž‘์—… ๋ถ„ํ•ด, ๊ณ„ํš ๋ฐ ์ธํ„ฐํŽ˜์ด์Šค ํ˜ธ์ถœ์„ ํ†ตํ•ด LLM์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋งˆ์ธํฌ๋ž˜ํ”„ํŠธ์—์„œ ๋‹ค์–‘ํ•œ ๊ณผ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ์•ก์…˜ ๊ณ„ํš ๋˜๋Š” ์ž‘์—… ์™„๋ฃŒ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ธฐ์กด ์ž‘์—…์€ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ผ๋ฐ˜์ ์ธ ๋ฉ”ํŠธ๋ฆญ(์˜ˆ: ์ƒ์„ฑ๋œ ์•ก์…˜ ๊ณ„ํš์˜ ์‹คํ–‰ ๊ฐ€๋Šฅ์„ฑ ๋ฐ ์ •ํ™•์„ฑ) [694]์„ ์ฑ„ํƒํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‹คํ—˜์„ ์ง์ ‘ ์ˆ˜ํ–‰ํ•˜์—ฌ ์„ฑ๊ณต๋ฅ  [698]์„ ์ธก์ •ํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. LLM์ด ์™ธ๋ถ€ ํ™˜๊ฒฝ๊ณผ ์ƒํ˜ธ ์ž‘์šฉํ•˜๊ณ  ์ •ํ™•ํ•œ ์•ก์…˜ ๊ณ„ํš์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[699]. ์ตœ๊ทผ์—๋Š” ์ฝ”๋“œ์™€ ๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ ์„ค๊ณ„[530] ๋ฐ ์‹ค์ œ ๊ทธ๋ผ์šด๋”ฉ ์ œ๊ณต[698]๊ณผ ๊ฐ™์ด LLM์˜ ์ƒํ˜ธ ์ž‘์šฉ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋ช‡ ๊ฐ€์ง€ ๊ฐœ์„  ๋ฐฉ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
Detailed experimental settings for individual datasets.For OpenQA datasets, we set the maximum new token number to 100 tokens. For closed-set tasks (PubMedHealth and ARC-C), we set the maximum new token length to 50 for all baselines. For Self-Rag inference on PubMed and ARC-C, instead of determining the output with the highest score 4 as in other tasks, we aggregate the scores for each option and select the answer option with the highest score. We found in zero-shot settings of fact checking, some LLMs can generate capitalized class labels (e.g., True) while our gold labels are lower-cased. Therefore, across different LLMs, for fact checking, we lowercase the predictions. In multiple choice tasks, we found some models generate answers in slightly different ways (e.g., (A) instead of A). We slightly modify instructions for each LLM to avoid such format violations, and further conduct string matching between each candidate and model predictions if format violations still remain. After that processing, in closed set tasks, model predictions match one of the gold classes in almost all cases. For ALCE, we found that Llama2-chat tend to generate significantly lower outputs than other models (e.g., on average, their output is nearly 100 token, while ChatGPT generates 40 tokens on average), resulting in inflated str-em scores. We limit the maximum generation length to 100 tokens for all baselines to avoid this issue, rather than the original 300 tokens in the ALCE paper. Consequently, all of the baseline output length is within 30-60 tokens. For FactScore, we set the maximum new token length to 500 for baselines and 200 for Self-Rag at each segment level.
๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์‹คํ—˜ ์„ค์ •. OpenQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ ์ตœ๋Œ€ ์ƒˆ token ์ˆ˜๋ฅผ 100 token์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ์‡„ํ˜• ์„ธํŠธ ์ž‘์—…(PubMedHealth ๋ฐ ARC-C)์˜ ๊ฒฝ์šฐ ๋ชจ๋“  baseline์— ๋Œ€ํ•ด ์ตœ๋Œ€ ์ƒˆ token ๊ธธ์ด๋ฅผ 50์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. PubMed ๋ฐ ARC-C์— ๋Œ€ํ•œ Self-Rag ์ถ”๋ก ์˜ ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์ž‘์—…๊ณผ ๊ฐ™์ด ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜ 4๋ฅผ ๊ฐ€์ง„ ์ถœ๋ ฅ์„ ๊ฒฐ์ •ํ•˜๋Š” ๋Œ€์‹  ๊ฐ ์˜ต์…˜์— ๋Œ€ํ•œ ์ ์ˆ˜๋ฅผ ์ง‘๊ณ„ํ•˜๊ณ  ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋‹ต๋ณ€ ์˜ต์…˜์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ํŒฉํŠธ ์ฒดํฌ์˜ zero-shot ์„ค์ •์—์„œ ์ผ๋ถ€ LLM์€ ๋Œ€๋ฌธ์ž๋กœ ๋œ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”(์˜ˆ: True)์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ˜๋ฉด gold ๋ ˆ์ด๋ธ”์€ ์†Œ๋ฌธ์ž๋กœ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ฌ๋Ÿฌ LLM์— ๊ฑธ์ณ ํŒฉํŠธ ์ฒดํฌ์˜ ๊ฒฝ์šฐ ์˜ˆ์ธก์„ ์†Œ๋ฌธ์ž๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ๊ฐ๊ด€์‹ ๋ฌธ์ œ์—์„œ ์ผ๋ถ€ ๋ชจ๋ธ์ด ์•ฝ๊ฐ„ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๋‹ต์•ˆ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ(์˜ˆ: A ๋Œ€์‹  (A))๊ฐ€ ๋ฐœ๊ฒฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜•์‹ ์œ„๋ฐ˜์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ LLM์— ๋Œ€ํ•œ instruction์„ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๊ณ , ํ˜•์‹ ์œ„๋ฐ˜์ด ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ๋Š” ๊ฒฝ์šฐ ๊ฐ ํ›„๋ณด์™€ ๋ชจ๋ธ ์˜ˆ์ธก ๊ฐ„์˜ ๋ฌธ์ž์—ด ์ผ์น˜์„ฑ์„ ์ถ”๊ฐ€๋กœ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฒ˜๋ฆฌ ํ›„ ํ์‡„ํ˜• ์„ธํŠธ ์ž‘์—…์—์„œ๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ๊ฒฝ์šฐ์— ๋ชจ๋ธ ์˜ˆ์ธก์ด gold ํด๋ž˜์Šค ์ค‘ ํ•˜๋‚˜์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ALCE์˜ ๊ฒฝ์šฐ, Llama2-chat์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ‰๊ท ์ ์œผ๋กœ ์ถœ๋ ฅ์ด ๊ฑฐ์˜ 100 token์ธ ๋ฐ˜๋ฉด ChatGPT๋Š” ํ‰๊ท ์ ์œผ๋กœ 40 token์„ ์ƒ์„ฑํ•จ). ์ด ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“  baseline์— ๋Œ€ํ•ด ์ตœ๋Œ€ ์ƒ์„ฑ ๊ธธ์ด๋ฅผ 100 token์œผ๋กœ ์ œํ•œํ–ˆ์œผ๋ฉฐ, ALCE ๋…ผ๋ฌธ์˜ ์›๋ž˜ 300 token์ด ์•„๋‹Œ 100 token์œผ๋กœ ์ œํ•œํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋ชจ๋“  baseline ์ถœ๋ ฅ ๊ธธ์ด๋Š” 30~60 token ์ด๋‚ด์ž…๋‹ˆ๋‹ค. FactScore์˜ ๊ฒฝ์šฐ, ๊ฐ ์„ธ๊ทธ๋จผํŠธ ์ˆ˜์ค€์—์„œ baseline์€ ์ตœ๋Œ€ ์ƒˆ token ๊ธธ์ด๋ฅผ 500์œผ๋กœ, Self-Rag๋Š” 200์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
๊ฐœ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์‹คํ—˜ ์„ค์ •.OpenQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฒฝ์šฐ ์ตœ๋Œ€ ์ƒˆ ํ† ํฐ ์ˆ˜๋ฅผ 100 ํ† ํฐ์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ์‡„ํ˜• ์„ธํŠธ ์ž‘์—…(PubMedHealth ๋ฐ ARC-C)์˜ ๊ฒฝ์šฐ ๋ชจ๋“  ๋ฒ ์ด์Šค๋ผ์ธ์— ๋Œ€ํ•ด ์ตœ๋Œ€ ์ƒˆ ํ† ํฐ ๊ธธ์ด๋ฅผ 50์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค. PubMed ๋ฐ ARC-C์— ๋Œ€ํ•œ Self-Rag ์ถ”๋ก ์˜ ๊ฒฝ์šฐ, ๋‹ค๋ฅธ ์ž‘์—…๊ณผ ๊ฐ™์ด ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜ 4๋ฅผ ๊ฐ€์ง„ ์ถœ๋ ฅ์„ ๊ฒฐ์ •ํ•˜๋Š” ๋Œ€์‹  ๊ฐ ์˜ต์…˜์— ๋Œ€ํ•œ ์ ์ˆ˜๋ฅผ ์ง‘๊ณ„ํ•˜๊ณ  ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋‹ต๋ณ€ ์˜ต์…˜์„ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ํŒฉํŠธ ์ฒดํฌ์˜ ์ œ๋กœ ์ƒท ์„ค์ •์—์„œ ์ผ๋ถ€ LLM์€ ๋Œ€๋ฌธ์ž๋กœ ๋œ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”(์˜ˆ: True)์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ˜๋ฉด ๊ณจ๋“œ ๋ ˆ์ด๋ธ”์€ ์†Œ๋ฌธ์ž๋กœ ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ฌ๋Ÿฌ LLM์— ๊ฑธ์ณ ํŒฉํŠธ ์ฒดํฌ์˜ ๊ฒฝ์šฐ ์˜ˆ์ธก์„ ์†Œ๋ฌธ์ž๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ๊ฐ๊ด€์‹ ๋ฌธ์ œ์—์„œ ์ผ๋ถ€ ๋ชจ๋ธ์ด ์•ฝ๊ฐ„ ๋‹ค๋ฅธ ๋ฐฉ์‹์œผ๋กœ ๋‹ต์•ˆ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝ์šฐ(์˜ˆ: A ๋Œ€์‹  (A))๊ฐ€ ๋ฐœ๊ฒฌ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ํ˜•์‹ ์œ„๋ฐ˜์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ LLM์— ๋Œ€ํ•œ ์ง€์นจ์„ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๊ณ , ํ˜•์‹ ์œ„๋ฐ˜์ด ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ๋Š” ๊ฒฝ์šฐ ๊ฐ ํ›„๋ณด์™€ ๋ชจ๋ธ ์˜ˆ์ธก ๊ฐ„์˜ ๋ฌธ์ž์—ด ์ผ์น˜์„ฑ์„ ์ถ”๊ฐ€๋กœ ๊ฒ€์‚ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฒ˜๋ฆฌ ํ›„ ํ์‡„ํ˜• ์„ธํŠธ ์ž‘์—…์—์„œ๋Š” ๊ฑฐ์˜ ๋ชจ๋“  ๊ฒฝ์šฐ์— ๋ชจ๋ธ ์˜ˆ์ธก์ด ๊ณจ๋“œ ํด๋ž˜์Šค ์ค‘ ํ•˜๋‚˜์™€ ์ผ์น˜ํ•ฉ๋‹ˆ๋‹ค. ALCE์˜ ๊ฒฝ์šฐ, Llama2-chat์€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ์€ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ‰๊ท ์ ์œผ๋กœ ์ถœ๋ ฅ์ด ๊ฑฐ์˜ 100 ํ† ํฐ์ธ ๋ฐ˜๋ฉด ChatGPT๋Š” ํ‰๊ท ์ ์œผ๋กœ 40 ํ† ํฐ์„ ์ƒ์„ฑํ•จ). ์ด ๋ฌธ์ œ๋ฅผ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ๊ธฐ์ค€์— ๋Œ€ํ•ด ์ตœ๋Œ€ ์ƒ์„ฑ ๊ธธ์ด๋ฅผ 100 ํ† ํฐ์œผ๋กœ ์ œํ•œํ–ˆ์œผ๋ฉฐ, ALCE ๋…ผ๋ฌธ์˜ ์›๋ž˜ 300 ํ† ํฐ์ด ์•„๋‹Œ 100 ํ† ํฐ์œผ๋กœ ์ œํ•œํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ๋ชจ๋“  ๊ธฐ์ค€ ์ถœ๋ ฅ ๊ธธ์ด๋Š” 30~60 ํ† ํฐ ์ด๋‚ด์ž…๋‹ˆ๋‹ค. ํŒฉํŠธ์Šค์ฝ”์–ด์˜ ๊ฒฝ์šฐ, ๊ฐ ์„ธ๊ทธ๋จผํŠธ ์ˆ˜์ค€์—์„œ ๊ธฐ์ค€์€ ์ตœ๋Œ€ ์ƒˆ ํ† ํฐ ๊ธธ์ด๋ฅผ 500์œผ๋กœ, ์…€ํ”„๋ž˜๊ทธ๋Š” 200์œผ๋กœ ์„ค์ •ํ–ˆ์Šต๋‹ˆ๋‹ค.
Uncovering early-stage metrics that reflect final model performance is one core principle for large-scale pretraining. The existing scaling law demonstrates the power-law correlation between pretraining loss and training flops, which serves as an important indicator of the current training state for large language models. However, this principle only focuses on the model's compression properties on the training data, resulting in an inconsistency with the ability improvements on the downstream tasks. Some follow-up works attempted to extend the scaling-law to more complex metrics (such as hyperparameters), but still lacked a comprehensive analysis of the dynamic differences among various capabilities during pretraining. To address the aforementioned limitations, this paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. Through this analysis, we confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes, up to 67 billion parameters. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints. This initiative offers valuable resources to the research community and facilitates the verification and exploration of LLM pretraining by open-source researchers. Besides, we provide empirical summaries, including performance comparisons of different models and capabilities, and tuition of key metrics for different training phases. Based on these findings, we provide a more user-friendly strategy for evaluating the optimization state, offering guidance for establishing a stable pretraining process.
์ตœ์ข… ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•˜๋Š” ์ดˆ๊ธฐ ๋‹จ๊ณ„ metrics๋ฅผ ๋ฐœ๊ฒฌํ•˜๋Š” ๊ฒƒ์€ ๋Œ€๊ทœ๋ชจ pretraining์˜ ํ•ต์‹ฌ ์›์น™ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ scaling law๋Š” pretraining loss์™€ training flops ์‚ฌ์ด์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™์  ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” Large Language Models (LLMs)์˜ ํ˜„์žฌ training ์ƒํƒœ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ค‘์š”ํ•œ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์›์น™์€ training ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์••์ถ• ํŠน์„ฑ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ธฐ ๋•Œ๋ฌธ์— downstream ์ž‘์—…์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ ํ–ฅ์ƒ๊ณผ ์ผ๊ด€์„ฑ์ด ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ํ›„์† ์—ฐ๊ตฌ์—์„œ๋Š” scaling law๋ฅผ ๋” ๋ณต์žกํ•œ metrics (์˜ˆ: hyperparameters)๋กœ ํ™•์žฅํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์ง€๋งŒ, pretraining ์ค‘ ๋‹ค์–‘ํ•œ ์—ญ๋Ÿ‰ ๊ฐ„์˜ ๋™์  ์ฐจ์ด์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๋ถ„์„์ด ์—ฌ์ „ํžˆ ๋ถ€์กฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค์–‘ํ•œ pretraining ์ค‘๊ฐ„ checkpoints์—์„œ ๋ชจ๋ธ ์—ญ๋Ÿ‰์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ„์„์„ ํ†ตํ•ด ํŠน์ • downstream metrics๊ฐ€ ์ตœ๋Œ€ 670์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ ํฌ๊ธฐ์˜ ๋ชจ๋ธ์—์„œ ์œ ์‚ฌํ•œ training dynamics๋ฅผ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•ต์‹ฌ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ ์™ธ์—๋„ Amber์™€ OpenLLaMA๋ฅผ ์žฌํ˜„ํ•˜์—ฌ ์ค‘๊ฐ„ checkpoints๋ฅผ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋Š” ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜คํ”ˆ ์†Œ์Šค ์—ฐ๊ตฌ์ž๋“ค์ด LLM pretraining์„ ๊ฒ€์ฆํ•˜๊ณ  ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ๊ณผ ๊ธฐ๋Šฅ์˜ ์„ฑ๋Šฅ ๋น„๊ต, ๋‹ค์–‘ํ•œ training ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ฃผ์š” metrics ๊ต์œก ๋“ฑ์˜ ๊ฒฝํ—˜์  ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ optimization ์ƒํƒœ๋ฅผ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ์ „๋žต์„ ์ œ๊ณตํ•˜์—ฌ ์•ˆ์ •์ ์ธ pretraining ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
์ตœ์ข… ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋ฐ˜์˜ํ•˜๋Š” ์ดˆ๊ธฐ ๋‹จ๊ณ„ ๋ฉ”ํŠธ๋ฆญ์„ ๋ฐœ๊ฒฌํ•˜๋Š” ๊ฒƒ์€ ๋Œ€๊ทœ๋ชจ ์‚ฌ์ „ ํ•™์Šต์˜ ํ•ต์‹ฌ ์›์น™ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์€ ์‚ฌ์ „ ํ•™์Šต ์†์‹ค๊ณผ ํ•™์Šต ํ”Œ๋กญ ์‚ฌ์ด์˜ ๊ฑฐ๋“ญ์ œ๊ณฑ ๋ฒ•์น™์  ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์ด๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ํ˜„์žฌ ํ•™์Šต ์ƒํƒœ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ค‘์š”ํ•œ ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์›์น™์€ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์••์ถ• ํŠน์„ฑ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ”๊ธฐ ๋•Œ๋ฌธ์— ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋Œ€ํ•œ ๋Šฅ๋ ฅ ํ–ฅ์ƒ๊ณผ ์ผ๊ด€์„ฑ์ด ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ํ›„์† ์—ฐ๊ตฌ์—์„œ๋Š” ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ๋” ๋ณต์žกํ•œ ๋ฉ”ํŠธ๋ฆญ(์˜ˆ: ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ)์œผ๋กœ ํ™•์žฅํ•˜๋ ค๊ณ  ์‹œ๋„ํ–ˆ์ง€๋งŒ, ์‚ฌ์ „ ํ•™์Šต ์ค‘ ๋‹ค์–‘ํ•œ ์—ญ๋Ÿ‰ ๊ฐ„์˜ ๋™์  ์ฐจ์ด์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๋ถ„์„์ด ์—ฌ์ „ํžˆ ๋ถ€์กฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ํ•œ๊ณ„๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์‚ฌ์ „ ํ•™์Šต ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์—์„œ ๋ชจ๋ธ ์—ญ๋Ÿ‰์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ„์„์„ ํ†ตํ•ด ํŠน์ • ๋‹ค์šด์ŠคํŠธ๋ฆผ ๋ฉ”ํŠธ๋ฆญ์ด ์ตœ๋Œ€ 670์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋‹ค์–‘ํ•œ ํฌ๊ธฐ์˜ ๋ชจ๋ธ์—์„œ ์œ ์‚ฌํ•œ ํ›ˆ๋ จ ๋™์—ญํ•™์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•ต์‹ฌ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ ์™ธ์—๋„ Amber์™€ OpenLLaMA๋ฅผ ์žฌํ˜„ํ•˜์—ฌ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ด๋‹ˆ์…”ํ‹ฐ๋ธŒ๋Š” ์—ฐ๊ตฌ ์ปค๋ฎค๋‹ˆํ‹ฐ์— ๊ท€์ค‘ํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์˜คํ”ˆ ์†Œ์Šค ์—ฐ๊ตฌ์ž๋“ค์ด LLM ์‚ฌ์ „ ํ•™์Šต์„ ๊ฒ€์ฆํ•˜๊ณ  ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ๊ณผ ๊ธฐ๋Šฅ์˜ ์„ฑ๋Šฅ ๋น„๊ต, ๋‹ค์–‘ํ•œ ํ›ˆ๋ จ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ฃผ์š” ๋ฉ”ํŠธ๋ฆญ ๊ต์œก ๋“ฑ์˜ ๊ฒฝํ—˜์  ์š”์•ฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ตœ์ ํ™” ์ƒํƒœ๋ฅผ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋ณด๋‹ค ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ์ „๋žต์„ ์ œ๊ณตํ•˜์—ฌ ์•ˆ์ •์ ์ธ ์‚ฌ์ „ ํ•™์Šต ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•œ ์ง€์นจ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Memory Writing_. The purpose of memory writing is to store information about the perceived environment in memory. Storing valuable information in memory provides a foundation for retrieving informative memories in the future, enabling the agent to act more efficiently and rationally. During the memory writing process, there are two potential problems that should be carefully addressed. On one hand, it is crucial to address how to store information that is similar to existing memories (_i.e._, memory duplicated). On the other hand, it is important to consider how to remove information when the memory reaches its storage limit (_i.e._, memory overflow). In the following, we discuss these problems more in detail. (1) _Memory Duplicated_. To incorporate similar information, people have developed various methods for integrating new and previous records. For instance, in [7], the successful action sequences related to the same sub-goal are stored in a list. Once the size of the list reaches N(=5), all the sequences in it are condensed into a unified plan solution using LLMs. The original sequences in the memory are replaced with the newly generated one. Augmented LLM [43] aggregates duplicate information via count accumulation, avoiding redundant storage. (2) _Memory Overflow_. In order to write information into the memory when it is full, people design different methods to delete existing information to continue the memorizing process. For example, in ChatDB [40], memories can be explicitly deleted based on user commands. RET-LLM [42] uses a fixed-size buffer for memory, overwriting the oldest entries in a first-in-first-out (FIFO) manner.
\(\bullet\)_Memory Writing_. ๋ฉ”๋ชจ๋ฆฌ ์“ฐ๊ธฐ์˜ ๋ชฉ์ ์€ ์ธ์ง€๋œ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ€์น˜ ์žˆ๋Š” ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๋ฉด ๋‚˜์ค‘์— ์œ ์ตํ•œ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์ด ๋งˆ๋ จ๋˜์–ด ์—์ด์ „ํŠธ๊ฐ€ ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํ•ฉ๋ฆฌ์ ์œผ๋กœ ํ–‰๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์“ฐ๊ธฐ ๊ณผ์ •์—์„œ ์‹ ์ค‘ํ•˜๊ฒŒ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋‘ ๊ฐ€์ง€ ์ž ์žฌ์  ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ๋Š” ๊ธฐ์กด ๋ฉ”๋ชจ๋ฆฌ์™€ ์œ ์‚ฌํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•(์˜ˆ: memory duplicated)์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ๋Š” ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ์ €์žฅ ํ•œ๋„์— ๋„๋‹ฌํ–ˆ์„ ๋•Œ ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ฐฉ๋ฒ•(์˜ˆ: memory overflow)์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ข€ ๋” ์ž์„ธํžˆ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. (1) _Memory Duplicated_. ์œ ์‚ฌํ•œ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ๋“ค์€ ์ƒˆ๋กœ์šด ๊ธฐ๋ก๊ณผ ์ด์ „ ๊ธฐ๋ก์„ ํ†ตํ•ฉํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [7]์—์„œ๋Š” ๋™์ผํ•œ ํ•˜์œ„ ๋ชฉํ‘œ์™€ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ํ–‰๋™ ์‹œํ€€์Šค๋ฅผ ๋ชฉ๋ก์— ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ๋ชฉ๋ก์˜ ํฌ๊ธฐ๊ฐ€ N(=5)์— ๋„๋‹ฌํ•˜๋ฉด ๋ชฉ๋ก์— ์žˆ๋Š” ๋ชจ๋“  ์‹œํ€€์Šค๋ฅผ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํ†ตํ•ฉ๋œ ๊ณ„ํš ์†”๋ฃจ์…˜์œผ๋กœ ์••์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ์˜ ์›๋ž˜ ์‹œํ€€์Šค๋Š” ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์‹œํ€€์Šค๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. Augmented LLM [43]์€ ์ค‘๋ณต ์ €์žฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์นด์šดํŠธ ๋ˆ„์ ์„ ํ†ตํ•ด ์ค‘๋ณต ์ •๋ณด๋ฅผ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค. (2) _Memory Overflow_. ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ๊ฐ€๋“ ์ฐผ์„ ๋•Œ ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์“ฐ๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ •๋ณด๋ฅผ ์‚ญ์ œํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌํ™” ๊ณผ์ •์„ ๊ณ„์†ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatDB [40]์—์„œ๋Š” ์‚ฌ์šฉ์ž ๋ช…๋ น์— ๋”ฐ๋ผ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์‚ญ์ œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. RET-LLM [42]๋Š” ๋ฉ”๋ชจ๋ฆฌ์— ๋Œ€ํ•ด ๊ณ ์ • ํฌ๊ธฐ ๋ฒ„ํผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ ์ž… ์„ ์ถœ(FIFO) ๋ฐฉ์‹์œผ๋กœ ๊ฐ€์žฅ ์˜ค๋ž˜๋œ ํ•ญ๋ชฉ์„ ๋ฎ์–ด์”๋‹ˆ๋‹ค.
\(\bullet\)_๋ฉ”๋ชจ๋ฆฌ ์“ฐ๊ธฐ_. ๋ฉ”๋ชจ๋ฆฌ ์“ฐ๊ธฐ์˜ ๋ชฉ์ ์€ ์ธ์ง€๋œ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ€์น˜ ์žˆ๋Š” ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์ €์žฅํ•˜๋ฉด ๋‚˜์ค‘์— ์œ ์ตํ•œ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๊ฒ€์ƒ‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์ด ๋งˆ๋ จ๋˜์–ด ์—์ด์ „ํŠธ๊ฐ€ ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํ•ฉ๋ฆฌ์ ์œผ๋กœ ํ–‰๋™ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ ์“ฐ๊ธฐ ๊ณผ์ •์—์„œ ์‹ ์ค‘ํ•˜๊ฒŒ ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋‘ ๊ฐ€์ง€ ์ž ์žฌ์  ๋ฌธ์ œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ๋Š” ๊ธฐ์กด ๋ฉ”๋ชจ๋ฆฌ์™€ ์œ ์‚ฌํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๋Š” ๋ฐฉ๋ฒ•(์˜ˆ: ๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต)์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ๋Š” ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ์ €์žฅ ํ•œ๋„์— ๋„๋‹ฌํ–ˆ์„ ๋•Œ ์ •๋ณด๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ฐฉ๋ฒ•(์˜ˆ: ๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ”Œ๋กœ)์„ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์— ๋Œ€ํ•ด ์ข€ ๋” ์ž์„ธํžˆ ๋…ผ์˜ํ•ฉ๋‹ˆ๋‹ค. (1) _๋ฉ”๋ชจ๋ฆฌ ์ค‘๋ณต_. ์œ ์‚ฌํ•œ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ๋žŒ๋“ค์€ ์ƒˆ๋กœ์šด ๊ธฐ๋ก๊ณผ ์ด์ „ ๊ธฐ๋ก์„ ํ†ตํ•ฉํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [7]์—์„œ๋Š” ๋™์ผํ•œ ํ•˜์œ„ ๋ชฉํ‘œ์™€ ๊ด€๋ จ๋œ ์„ฑ๊ณต์ ์ธ ํ–‰๋™ ์‹œํ€€์Šค๋ฅผ ๋ชฉ๋ก์— ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ๋ชฉ๋ก์˜ ํฌ๊ธฐ๊ฐ€ N(=5)์— ๋„๋‹ฌํ•˜๋ฉด ๋ชฉ๋ก์— ์žˆ๋Š” ๋ชจ๋“  ์‹œํ€€์Šค๋ฅผ LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํ†ตํ•ฉ๋œ ๊ณ„ํš ์†”๋ฃจ์…˜์œผ๋กœ ์••์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ์˜ ์›๋ž˜ ์‹œํ€€์Šค๋Š” ์ƒˆ๋กœ ์ƒ์„ฑ๋œ ์‹œํ€€์Šค๋กœ ๋Œ€์ฒด๋ฉ๋‹ˆ๋‹ค. ์ฆ๊ฐ• LLM [43]์€ ์ค‘๋ณต ์ €์žฅ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ์นด์šดํŠธ ๋ˆ„์ ์„ ํ†ตํ•ด ์ค‘๋ณต ์ •๋ณด๋ฅผ ์ง‘๊ณ„ํ•ฉ๋‹ˆ๋‹ค. (2) _๋ฉ”๋ชจ๋ฆฌ ์˜ค๋ฒ„ํ”Œ๋กœ_. ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ๊ฐ€๋“ ์ฐผ์„ ๋•Œ ์ •๋ณด๋ฅผ ๋ฉ”๋ชจ๋ฆฌ์— ์“ฐ๊ธฐ ์œ„ํ•ด ๊ธฐ์กด ์ •๋ณด๋ฅผ ์‚ญ์ œํ•˜์—ฌ ๋ฉ”๋ชจ๋ฆฌํ™” ๊ณผ์ •์„ ๊ณ„์†ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatDB [40]์—์„œ๋Š” ์‚ฌ์šฉ์ž ๋ช…๋ น์— ๋”ฐ๋ผ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์‚ญ์ œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. RET-LLM [42]์€ ๋ฉ”๋ชจ๋ฆฌ์— ๋Œ€ํ•ด ๊ณ ์ • ํฌ๊ธฐ ๋ฒ„ํผ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ ์ž… ์„ ์ถœ(FIFO) ๋ฐฉ์‹์œผ๋กœ ๊ฐ€์žฅ ์˜ค๋ž˜๋œ ํ•ญ๋ชฉ์„ ๋ฎ์–ด์”๋‹ˆ๋‹ค.
**Basics and Principles.** Instead of training on specific task goals, LLMs learn from unsupervised pre-training on large-scale text data. This is quite different from previous multi-task learning approaches, which aim to extend the training tasks as possible to achieve sufficient generalization. Thus, it is essential to reveal the basic principles or elements that establish the foundation of the abilities of LLMs. Although the basic idea of language models is intuitive, it is still challenging to formally explain why LLMs trained by simple language modeling objectives (_e.g._, next token prediction) can become capable of solving various real-world tasks. To investigate this problem, a promising approach is to study the capacity learning (or selection) mechanism based on unsupervised pre-training, since the model capacity of LLMs strongly depends on pre-training data. In addition, _scaling_ plays an important role in improving the capacity of LLMs [31, 55, 64], and it is very useful to conduct more theoretical analysis about how the behaviors of large models relate to those of small models, _e.g._, what behaviors of large models can be inferred from small models and what can't be predicted indeed. Another research direction is to explore more deep analysis on model generalization for LLMs, since increasing concerns have been raised about whether LLMs can generalize beyond the knowledge encoded by pre-training data. Furthermore, data contamination has become a severe issue for fairly assessing the performance of LLMs [738], and thus setting appropriate evaluation protocol will be the basis to investigate and analyze the model capacity of LLMs.
**๊ธฐ๋ณธ ๋ฐ ์›๋ฆฌ**. LLM์€ ํŠน์ • ์ž‘์—… ๋ชฉํ‘œ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ๋Œ€์‹  ๋Œ€๊ทœ๋ชจ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋น„์ง€๋„ pre-training์„ ํ†ตํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถฉ๋ถ„ํ•œ ์ผ๋ฐ˜ํ™”๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ํ•œ ํ›ˆ๋ จ ์ž‘์—…์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ธฐ์กด์˜ multi-task learning ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ์ƒ๋‹นํžˆ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์˜ ๋Šฅ๋ ฅ์˜ ๊ธฐ๋ฐ˜์„ ๋‹ค์ง€๋Š” ๊ธฐ๋ณธ ์›๋ฆฌ ๋˜๋Š” ์š”์†Œ๋ฅผ ๋ฐํžˆ๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ์ง๊ด€์ ์ด์ง€๋งŒ, ๋‹จ์ˆœํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ชฉํ‘œ(์˜ˆ: ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก)์— ์˜ํ•ด ํ›ˆ๋ จ๋œ LLM์ด ๋‹ค์–‘ํ•œ ์‹ค์ œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์ด์œ ๋ฅผ ๊ณต์‹์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด pre-training ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„์ง€๋„ pre-training์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์šฉ๋Ÿ‰ ํ•™์Šต(๋˜๋Š” ์„ ํƒ) ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, scaling์€ LLM์˜ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ–‰๋™์ด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ–‰๋™๊ณผ ์–ด๋–ค ๊ด€๋ จ์ด ์žˆ๋Š”์ง€, ์ฆ‰ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์–ด๋–ค ํ–‰๋™์ด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์—์„œ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€, ์‹ค์ œ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ํ–‰๋™์€ ๋ฌด์—‡์ธ์ง€์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ด๋ก ์ ์ธ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ LLM์— ๋Œ€ํ•œ ๋ชจ๋ธ ์ผ๋ฐ˜ํ™”์— ๋Œ€ํ•œ ๋” ๊นŠ์€ ๋ถ„์„์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ธ๋ฐ, ์ด๋Š” LLM์ด pre-training ๋ฐ์ดํ„ฐ์— ์ธ์ฝ”๋”ฉ๋œ ์ง€์‹์„ ๋„˜์–ด ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ปค์ง€๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฐ์ดํ„ฐ ์˜ค์—ผ์€ LLM์˜ ์„ฑ๋Šฅ์„ ๊ณต์ •ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ๊ฐ€ ๋˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ ์ ์ ˆํ•œ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์„ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์€ LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ์กฐ์‚ฌํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๊ธฐ์ดˆ๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
**๊ธฐ๋ณธ ๋ฐ ์›๋ฆฌ**. LLM์€ ํŠน์ • ์ž‘์—… ๋ชฉํ‘œ์— ๋Œ€ํ•œ ํ›ˆ๋ จ ๋Œ€์‹  ๋Œ€๊ทœ๋ชจ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋น„์ง€๋„ ์‚ฌ์ „ ํ›ˆ๋ จ์„ ํ†ตํ•ด ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์ถฉ๋ถ„ํ•œ ์ผ๋ฐ˜ํ™”๋ฅผ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ํ•œ ํ›ˆ๋ จ ์ž‘์—…์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋Š” ๊ธฐ์กด์˜ ๋‹ค์ค‘ ์ž‘์—… ํ•™์Šต ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ๋Š” ์ƒ๋‹นํžˆ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ LLM์˜ ๋Šฅ๋ ฅ์˜ ๊ธฐ๋ฐ˜์„ ๋‹ค์ง€๋Š” ๊ธฐ๋ณธ ์›๋ฆฌ ๋˜๋Š” ์š”์†Œ๋ฅผ ๋ฐํžˆ๋Š” ๊ฒƒ์ด ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ธฐ๋ณธ ์•„์ด๋””์–ด๋Š” ์ง๊ด€์ ์ด์ง€๋งŒ, ๋‹จ์ˆœํ•œ ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ชฉํ‘œ(์˜ˆ: ๋‹ค์Œ ํ† ํฐ ์˜ˆ์ธก)์— ์˜ํ•ด ํ›ˆ๋ จ๋œ LLM์ด ๋‹ค์–‘ํ•œ ์‹ค์ œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์ด์œ ๋ฅผ ๊ณต์‹์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌ๊ฒŒ ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„์ง€๋„ ์‚ฌ์ „ ํ•™์Šต์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์šฉ๋Ÿ‰ ํ•™์Šต(๋˜๋Š” ์„ ํƒ) ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ์Šค์ผ€์ผ๋ง_์€ LLM์˜ ์šฉ๋Ÿ‰์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ–‰๋™์ด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์˜ ํ–‰๋™๊ณผ ์–ด๋–ค ๊ด€๋ จ์ด ์žˆ๋Š”์ง€, ์ฆ‰ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์–ด๋–ค ํ–‰๋™์ด ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ์—์„œ ์œ ์ถ”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€, ์‹ค์ œ๋กœ ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ํ–‰๋™์€ ๋ฌด์—‡์ธ์ง€์— ๋Œ€ํ•ด ๋ณด๋‹ค ์ด๋ก ์ ์ธ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ LLM์— ๋Œ€ํ•œ ๋ชจ๋ธ ์ผ๋ฐ˜ํ™”์— ๋Œ€ํ•œ ๋” ๊นŠ์€ ๋ถ„์„์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ธ๋ฐ, ์ด๋Š” LLM์ด ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์ธ์ฝ”๋”ฉ๋œ ์ง€์‹์„ ๋„˜์–ด ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ปค์ง€๊ณ  ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋ฐ์ดํ„ฐ ์˜ค์—ผ์€ LLM์˜ ์„ฑ๋Šฅ์„ ๊ณต์ •ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‹ฌ๊ฐํ•œ ๋ฌธ์ œ๊ฐ€ ๋˜๊ณ  ์žˆ์œผ๋ฏ€๋กœ ์ ์ ˆํ•œ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์„ ์„ค์ •ํ•˜๋Š” ๊ฒƒ์€ LLM์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ์กฐ์‚ฌํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๊ธฐ์ดˆ๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
tonomously, and then metrics like task success rate and human similarity are leveraged to evaluate the capability of the agents based on their trajectories and completed objectives [164, 166, 167, 33, 38, 59, 85, 138, 164, 170]. This method is expected to evaluate the agents' practical capabilities in real-world scenarios. (2) _Social evaluation_: This method utilizes metrics to assess social intelligence based on the agent interactions in simulated societies. Various approaches have been adopted, such as collaborative tasks to evaluate teamwork skills, debates to analyze argumentative reasoning, and human studies to measure social aptitude [165, 173, 34, 79, 102]. These approaches analyze qualities such as coherence, theory of mind, and social IQ to assess agents' capabilities in areas including cooperation, communication, empathy, and mimicking human social behavior. By subjecting agents to complex interactive settings, social evaluation provides valuable insights into agents' higher-level social cognition. (3) _Multi-task evaluation:_ In this method, people use a set of diverse tasks from different domains to evaluate the agent, which can effectively measure the agent generalization capability in open-domain environments [165, 172, 173, 166, 85, 29, 85]. (4) _Software testing:_ In this method, researchers evaluate the agents by letting them conduct tasks such as software testing tasks, such as generating test cases, reproducing bugs, debugging code, and interacting with developers and external tools [162, 169, 163, 169]. Then, one can use metrics like test coverage and bug detection rate to measure the effectiveness of LLM-based agents.
ํ† ๋…ธ๋ฏธํ•˜๊ฒŒ ํ‰๊ฐ€ํ•œ ๋‹ค์Œ, ์ž‘์—… ์„ฑ๊ณต๋ฅ  ๋ฐ ์‚ฌ๋žŒ๊ณผ์˜ ์œ ์‚ฌ์„ฑ๊ณผ ๊ฐ™์€ metrics๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ์˜ ๊ถค์ ๊ณผ ์™„๋ฃŒ๋œ ๋ชฉํ‘œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[164, 166, 167, 33, 38, 59, 85, 138, 164, 170]. ์ด ๋ฐฉ๋ฒ•์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์—์ด์ „ํŠธ์˜ ์‹ค์ œ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. (2) _์‚ฌํšŒ์  ํ‰๊ฐ€_: ์ด ๋ฐฉ๋ฒ•์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์‚ฌํšŒ์—์„œ ์—์ด์ „ํŠธ์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ metrics๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‚ฌํšŒ์  ์ง€๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŒ€์›Œํฌ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํ˜‘์—… ๊ณผ์ œ, ๋…ผ์ฆ์  ์ถ”๋ก ์„ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•œ ํ† ๋ก , ์‚ฌํšŒ์  ์ ์„ฑ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ฐ„ ์—ฐ๊ตฌ ๋“ฑ ๋‹ค์–‘ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[165, 173, 34, 79, 102]. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๊ด€์„ฑ, ๋งˆ์Œ์˜ ์ด๋ก , ์‚ฌํšŒ์  IQ์™€ ๊ฐ™์€ ์ž์งˆ์„ ๋ถ„์„ํ•˜์—ฌ ํ˜‘๋ ฅ, ์˜์‚ฌ์†Œํ†ต, ๊ณต๊ฐ, ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™ ๋ชจ๋ฐฉ ๋“ฑ์˜ ์˜์—ญ์—์„œ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌํšŒ์  ํ‰๊ฐ€๋Š” ์—์ด์ „ํŠธ๋ฅผ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ ํ™˜๊ฒฝ์— ๋…ธ์ถœ์‹œํ‚ด์œผ๋กœ์จ ์—์ด์ „ํŠธ์˜ ์ƒ์œ„ ์ˆ˜์ค€์˜ ์‚ฌํšŒ์  ์ธ์ง€์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (3) _Multi-task ํ‰๊ฐ€:_ ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์˜ ๋‹ค์–‘ํ•œ ํƒœ์Šคํฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š”๋ฐ, ์ด๋Š” ์˜คํ”ˆ ๋„๋ฉ”์ธ ํ™˜๊ฒฝ์—์„œ ์—์ด์ „ํŠธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[165, 172, 173, 166, 85, 29, 85]. (4) _์†Œํ”„ํŠธ์›จ์–ด ํ…Œ์ŠคํŠธ:_ ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์—ฐ๊ตฌ์ž๋“ค์ด ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค ์ƒ์„ฑ, ๋ฒ„๊ทธ ์žฌํ˜„, ์ฝ”๋“œ ๋””๋ฒ„๊น…, ๊ฐœ๋ฐœ์ž ๋ฐ ์™ธ๋ถ€ ๋„๊ตฌ์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ ๋“ฑ์˜ ์†Œํ”„ํŠธ์›จ์–ด ํ…Œ์ŠคํŠธ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[162, 169, 163, 169]. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ…Œ์ŠคํŠธ ์ปค๋ฒ„๋ฆฌ์ง€ ๋ฐ ๋ฒ„๊ทธ ํƒ์ง€์œจ๊ณผ ๊ฐ™์€ metrics๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
ํ† ๋…ธ๋ฏธํ•˜๊ฒŒ ํ‰๊ฐ€ํ•œ ๋‹ค์Œ, ์ž‘์—… ์„ฑ๊ณต๋ฅ  ๋ฐ ์‚ฌ๋žŒ๊ณผ์˜ ์œ ์‚ฌ์„ฑ๊ณผ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ํ™œ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ์˜ ๊ถค์ ๊ณผ ์™„๋ฃŒ๋œ ๋ชฉํ‘œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[164, 166, 167, 33, 38, 59, 85, 138, 164, 170]. ์ด ๋ฐฉ๋ฒ•์€ ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์—์ด์ „ํŠธ์˜ ์‹ค์ œ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค. (2) _์‚ฌํšŒ์  ํ‰๊ฐ€_: ์ด ๋ฐฉ๋ฒ•์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ์‚ฌํšŒ์—์„œ ์—์ด์ „ํŠธ์˜ ์ƒํ˜ธ ์ž‘์šฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฉ”ํŠธ๋ฆญ์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฌํšŒ์  ์ง€๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ํŒ€์›Œํฌ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํ˜‘์—… ๊ณผ์ œ, ๋…ผ์ฆ์  ์ถ”๋ก ์„ ๋ถ„์„ํ•˜๊ธฐ ์œ„ํ•œ ํ† ๋ก , ์‚ฌํšŒ์  ์ ์„ฑ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•œ ์ธ๊ฐ„ ์—ฐ๊ตฌ ๋“ฑ ๋‹ค์–‘ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[165, 173, 34, 79, 102]. ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ผ๊ด€์„ฑ, ๋งˆ์Œ์˜ ์ด๋ก , ์‚ฌํšŒ์  IQ์™€ ๊ฐ™์€ ์ž์งˆ์„ ๋ถ„์„ํ•˜์—ฌ ํ˜‘๋ ฅ, ์˜์‚ฌ์†Œํ†ต, ๊ณต๊ฐ, ์ธ๊ฐ„์˜ ์‚ฌํšŒ์  ํ–‰๋™ ๋ชจ๋ฐฉ ๋“ฑ์˜ ์˜์—ญ์—์„œ ์—์ด์ „ํŠธ์˜ ์—ญ๋Ÿ‰์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌํšŒ์  ํ‰๊ฐ€๋Š” ์—์ด์ „ํŠธ๋ฅผ ๋ณต์žกํ•œ ์ƒํ˜ธ ์ž‘์šฉ ํ™˜๊ฒฝ์— ๋…ธ์ถœ์‹œํ‚ด์œผ๋กœ์จ ์—์ด์ „ํŠธ์˜ ์ƒ์œ„ ์ˆ˜์ค€์˜ ์‚ฌํšŒ์  ์ธ์ง€์— ๋Œ€ํ•œ ๊ท€์ค‘ํ•œ ์ธ์‚ฌ์ดํŠธ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. (3) _๋ฉ€ํ‹ฐํƒœ์Šคํฌ ํ‰๊ฐ€:_ ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์˜ ๋‹ค์–‘ํ•œ ํƒœ์Šคํฌ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•˜๋Š”๋ฐ, ์ด๋Š” ์˜คํ”ˆ ๋„๋ฉ”์ธ ํ™˜๊ฒฝ์—์„œ ์—์ด์ „ํŠธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํšจ๊ณผ์ ์œผ๋กœ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[165, 172, 173, 166, 85, 29, 85]. (4) _์†Œํ”„ํŠธ์›จ์–ด ํ…Œ์ŠคํŠธ:_ ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ์—ฐ๊ตฌ์ž๋“ค์ด ํ…Œ์ŠคํŠธ ์ผ€์ด์Šค ์ƒ์„ฑ, ๋ฒ„๊ทธ ์žฌํ˜„, ์ฝ”๋“œ ๋””๋ฒ„๊น…, ๊ฐœ๋ฐœ์ž ๋ฐ ์™ธ๋ถ€ ๋„๊ตฌ์™€์˜ ์ƒํ˜ธ ์ž‘์šฉ ๋“ฑ์˜ ์†Œํ”„ํŠธ์›จ์–ด ํ…Œ์ŠคํŠธ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ์—์ด์ „ํŠธ๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค[162, 169, 163, 169]. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํ…Œ์ŠคํŠธ ์ปค๋ฒ„๋ฆฌ์ง€ ๋ฐ ๋ฒ„๊ทธ ํƒ์ง€์œจ๊ณผ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•˜์—ฌ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์˜ ํšจ๊ณผ๋ฅผ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The integration and deployment of large language model (LLM)-based intelligent agents have been fraught with challenges that compromise their efficiency and efficacy. Among these issues are sub-optimal scheduling and resource allocation of agent requests over the LLM, the difficulties in maintaining context during interactions between agent and LLM, and the complexities inherent in integrating heterogeneous agents with different capabilities and specializations. The rapid increase of agent quantity and complexity further exacerbates these issues, often leading to bottlenecks and sub-optimal utilization of resources. Inspired by these challenges, this paper presents AIOS, an LLM agent operating system, which embeds large language model into operating systems (OS) as the brain of the OS, enabling an operating system "with soul"--an important step towards AGI. Specifically, AIOS is designed to optimize resource allocation, facilitate context switch across agents, enable concurrent execution of agents, provide tool service for agents, and maintain access control for agents. We present the architecture of such an operating system, outline the core challenges it aims to resolve, and provide the basic design and implementation of the AIOS. Our experiments on concurrent execution of multiple agents demonstrate the reliability and efficiency of our AIOS modules. Through this, we aim to not only improve the performance and efficiency of LLM agents but also to pioneer for better development and deployment of the AIOS ecosystem in the future. The project is open-source at [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS).
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์ง€๋Šฅํ˜• ์—์ด์ „ํŠธ์˜ ํ†ตํ•ฉ ๋ฐ ๋ฐฐํฌ์—๋Š” ํšจ์œจ์„ฑ๊ณผ ํšจ๊ณผ์„ฑ์„ ์ €ํ•ดํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์ˆ˜๋ฐ˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ ์ค‘์—๋Š” LLM์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์ตœ์ ์ด ์•„๋‹Œ ์Šค์ผ€์ค„๋ง ๋ฐ ๋ฆฌ์†Œ์Šค ํ• ๋‹น, ์—์ด์ „ํŠธ์™€ LLM ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘ ์ปจํ…์ŠคํŠธ ์œ ์ง€์˜ ์–ด๋ ค์›€, ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ๊ณผ ์ „๋ฌธ์„ฑ์„ ๊ฐ€์ง„ ์ด๊ธฐ์ข… ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋‚ด์žฌ๋œ ๋ณต์žก์„ฑ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ์ˆ˜์™€ ๋ณต์žก์„ฑ์ด ๊ธ‰๊ฒฉํžˆ ์ฆ๊ฐ€ํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋˜์–ด ๋ณ‘๋ชฉ ํ˜„์ƒ๊ณผ ๋ฆฌ์†Œ์Šค์˜ ์ตœ์  ํ™œ์šฉ์ด ์ €ํ•ด๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณผ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM ์—์ด์ „ํŠธ ์šด์˜ ์ฒด์ œ์ธ AIOS๋ฅผ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ์šด์˜ ์ฒด์ œ(OS)์— Large Language Model์„ ๋‚ด์žฅํ•˜์—ฌ OS์˜ ๋‘๋‡Œ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ AGI๋ฅผ ํ–ฅํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ธ '์˜ํ˜ผ์„ ๊ฐ€์ง„' ์šด์˜ ์ฒด์ œ๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ AIOS๋Š” ๋ฆฌ์†Œ์Šค ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๊ณ , ์—์ด์ „ํŠธ ๊ฐ„์— ์ปจํ…์ŠคํŠธ ์ „ํ™˜์„ ์ด‰์ง„ํ•˜๊ณ , ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ณ , ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ๋„๊ตฌ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ์ œ์–ด๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์šด์˜ ์ฒด์ œ์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์‹œํ•˜๊ณ , ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ํ•ต์‹ฌ ๊ณผ์ œ๋ฅผ ๊ฐœ๊ด„์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋ฉฐ, AIOS์˜ ๊ธฐ๋ณธ ์„ค๊ณ„ ๋ฐ ๊ตฌํ˜„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด AIOS ๋ชจ๋“ˆ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM ์—์ด์ „ํŠธ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ–ฅํ›„ AIOS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋” ๋‚˜์€ ๊ฐœ๋ฐœ๊ณผ ๋ฐฐํฌ๋ฅผ ๊ฐœ์ฒ™ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS)์—์„œ ์˜คํ”ˆ ์†Œ์Šค๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM) ๊ธฐ๋ฐ˜ ์ง€๋Šฅํ˜• ์—์ด์ „ํŠธ์˜ ํ†ตํ•ฉ ๋ฐ ๋ฐฐํฌ์—๋Š” ํšจ์œจ์„ฑ๊ณผ ํšจ๊ณผ์„ฑ์„ ์ €ํ•ดํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๋ฌธ์ œ๊ฐ€ ์ˆ˜๋ฐ˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ ์ค‘์—๋Š” LLM์— ๋Œ€ํ•œ ์—์ด์ „ํŠธ ์š”์ฒญ์˜ ์ตœ์ ์ด ์•„๋‹Œ ์Šค์ผ€์ค„๋ง ๋ฐ ๋ฆฌ์†Œ์Šค ํ• ๋‹น, ์—์ด์ „ํŠธ์™€ LLM ๊ฐ„์˜ ์ƒํ˜ธ ์ž‘์šฉ ์ค‘ ์ปจํ…์ŠคํŠธ ์œ ์ง€์˜ ์–ด๋ ค์›€, ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ๊ณผ ์ „๋ฌธ์„ฑ์„ ๊ฐ€์ง„ ์ด๊ธฐ์ข… ์—์ด์ „ํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” ๋ฐ ๋‚ด์žฌ๋œ ๋ณต์žก์„ฑ ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—์ด์ „ํŠธ์˜ ์ˆ˜์™€ ๋ณต์žก์„ฑ์ด ๊ธ‰๊ฒฉํžˆ ์ฆ๊ฐ€ํ•˜๋ฉด์„œ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋˜์–ด ๋ณ‘๋ชฉ ํ˜„์ƒ๊ณผ ๋ฆฌ์†Œ์Šค์˜ ์ตœ์  ํ™œ์šฉ์ด ์ €ํ•ด๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ณผ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ด ๋ฐฑ์„œ์—์„œ๋Š” LLM ์—์ด์ „ํŠธ ์šด์˜ ์ฒด์ œ์ธ AIOS๋ฅผ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ์šด์˜ ์ฒด์ œ(OS)์— ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‚ด์žฅํ•˜์—ฌ OS์˜ ๋‘๋‡Œ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ•จ์œผ๋กœ์จ AGI๋ฅผ ํ–ฅํ•œ ์ค‘์š”ํ•œ ๋‹จ๊ณ„์ธ '์˜ํ˜ผ์„ ๊ฐ€์ง„' ์šด์˜ ์ฒด์ œ๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. ํŠนํžˆ AIOS๋Š” ๋ฆฌ์†Œ์Šค ํ• ๋‹น์„ ์ตœ์ ํ™”ํ•˜๊ณ , ์—์ด์ „ํŠธ ๊ฐ„์— ์ปจํ…์ŠคํŠธ ์ „ํ™˜์„ ์ด‰์ง„ํ•˜๊ณ , ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๊ณ , ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ๋„๊ตฌ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์—์ด์ „ํŠธ์— ๋Œ€ํ•œ ์•ก์„ธ์Šค ์ œ์–ด๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์šด์˜ ์ฒด์ œ์˜ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์ œ์‹œํ•˜๊ณ , ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ํ•ต์‹ฌ ๊ณผ์ œ๋ฅผ ๊ฐœ๊ด„์ ์œผ๋กœ ์„ค๋ช…ํ•˜๋ฉฐ, AIOS์˜ ๊ธฐ๋ณธ ์„ค๊ณ„ ๋ฐ ๊ตฌํ˜„์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ์˜ ๋™์‹œ ์‹คํ–‰์— ๋Œ€ํ•œ ์‹คํ—˜์„ ํ†ตํ•ด AIOS ๋ชจ๋“ˆ์˜ ์•ˆ์ •์„ฑ๊ณผ ํšจ์œจ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด LLM ์—์ด์ „ํŠธ์˜ ์„ฑ๋Šฅ๊ณผ ํšจ์œจ์„ฑ์„ ๊ฐœ์„ ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ํ–ฅํ›„ AIOS ์—์ฝ”์‹œ์Šคํ…œ์˜ ๋” ๋‚˜์€ ๊ฐœ๋ฐœ๊ณผ ๋ฐฐํฌ๋ฅผ ๊ฐœ์ฒ™ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” [https://github.com/agiresearch/AIOS](https://github.com/agiresearch/AIOS)์—์„œ ์˜คํ”ˆ ์†Œ์Šค๋กœ ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค.
**Computer Science & Software Engineering**: In the field of computer science and software engineering, LLM-based agents offer potential for automating coding, testing, debugging, and documentation generation [14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18] proposes an end-to-end framework, where multiple agent roles communicate and collaborate through natural language conversations to complete the software development life cycle. This framework demonstrates efficient and cost-effective generation of executable software systems. ToolBench [14] can be used for tasks such as code auto-completion and code recommendation. MetaGPT [23] abstracts multiple roles, such as product managers, architects, project managers, and engineers, to supervise code generation process and enhance the quality of the final output code. This enables low-cost software development. [24] presents a self-collaboration framework for code generation using LLMs. In this framework, multiple LLMs are assumed to be distinct "experts" for specific sub-tasks. They collaborate and interact according to specified instructions, forming a virtual team that facilitates each other's work. Ultimately, the virtual team collaboratively addresses code generation tasks without requiring human intervention. LLIFT [139] employs LLMs to assist in conducting static analysis, specifically for identifying potential code vulnerabilities. This approach effectively manages the trade-off between accuracy and scalability. ChatEDA [123] is an agent developed for electronic design automation (EDA) to streamline the design process by integrating task planning, script generation, and execution.
**์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง**: ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ฝ”๋”ฉ, ํ…Œ์ŠคํŠธ, ๋””๋ฒ„๊น… ๋ฐ ๋ฌธ์„œ ์ƒ์„ฑ ์ž๋™ํ™”์— ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ˆ˜๋ช… ์ฃผ๊ธฐ๋ฅผ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ์—ญํ• ์ด ์ž์—ฐ์–ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘์—…ํ•˜๋Š” end-to-end ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํšจ์œจ์ ์ด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ธ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์†Œํ”„ํŠธ์›จ์–ด ์‹œ์Šคํ…œ ์ƒ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ToolBench [14]๋Š” ์ฝ”๋“œ ์ž๋™ ์™„์„ฑ ๋ฐ ์ฝ”๋“œ ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MetaGPT [23]๋Š” ์ œํ’ˆ ๊ด€๋ฆฌ์ž, ์•„ํ‚คํ…ํŠธ, ํ”„๋กœ์ ํŠธ ๊ด€๋ฆฌ์ž, ์—”์ง€๋‹ˆ์–ด ๋“ฑ ์—ฌ๋Ÿฌ ์—ญํ• ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ๋…ํ•˜๊ณ  ์ตœ์ข… ์ถœ๋ ฅ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ €๋น„์šฉ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. [24]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•œ ์ž์ฒด ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ LLM์ด ํŠน์ • ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ณ„๊ฐœ์˜ "์ „๋ฌธ๊ฐ€"๋กœ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ง€์ •๋œ instruction์— ๋”ฐ๋ผ ํ˜‘์—…ํ•˜๊ณ  ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์„œ๋กœ์˜ ์ž‘์—…์„ ์ด‰์ง„ํ•˜๋Š” ๊ฐ€์ƒ ํŒ€์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๊ฐ€์ƒ ํŒ€์€ ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์—†์ด๋„ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ํ˜‘๋ ฅ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. LLIFT [139]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠนํžˆ ์ž ์žฌ์ ์ธ ์ฝ”๋“œ ์ทจ์•ฝ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ •์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •ํ™•์„ฑ๊ณผ ํ™•์žฅ์„ฑ ๊ฐ„์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatEDA [123]๋Š” ์ž‘์—… ๊ณ„ํš, ์Šคํฌ๋ฆฝํŠธ ์ƒ์„ฑ ๋ฐ ์‹คํ–‰์„ ํ†ตํ•ฉํ•˜์—ฌ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ž ์„ค๊ณ„ ์ž๋™ํ™”(EDA)๋ฅผ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.
**์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง**: ์ปดํ“จํ„ฐ ๊ณผํ•™ ๋ฐ ์†Œํ”„ํŠธ์›จ์–ด ์—”์ง€๋‹ˆ์–ด๋ง ๋ถ„์•ผ์—์„œ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ฝ”๋”ฉ, ํ…Œ์ŠคํŠธ, ๋””๋ฒ„๊น… ๋ฐ ๋ฌธ์„œ ์ƒ์„ฑ ์ž๋™ํ™”์— ์ž ์žฌ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค[14; 18; 23; 24; 126; 127; 128]. Chat-Dev [18]์€ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ ์ˆ˜๋ช… ์ฃผ๊ธฐ๋ฅผ ์™„๋ฃŒํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ ์—ญํ• ์ด ์ž์—ฐ์–ด ๋Œ€ํ™”๋ฅผ ํ†ตํ•ด ์†Œํ†ตํ•˜๊ณ  ํ˜‘์—…ํ•˜๋Š” ์—”๋“œํˆฌ์—”๋“œ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ๋Š” ํšจ์œจ์ ์ด๊ณ  ๋น„์šฉ ํšจ์œจ์ ์ธ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ์†Œํ”„ํŠธ์›จ์–ด ์‹œ์Šคํ…œ ์ƒ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ToolBench [14]๋Š” ์ฝ”๋“œ ์ž๋™ ์™„์„ฑ ๋ฐ ์ฝ”๋“œ ์ถ”์ฒœ๊ณผ ๊ฐ™์€ ์ž‘์—…์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MetaGPT [23]๋Š” ์ œํ’ˆ ๊ด€๋ฆฌ์ž, ์•„ํ‚คํ…ํŠธ, ํ”„๋กœ์ ํŠธ ๊ด€๋ฆฌ์ž, ์—”์ง€๋‹ˆ์–ด ๋“ฑ ์—ฌ๋Ÿฌ ์—ญํ• ์„ ์ถ”์ƒํ™”ํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ๋…ํ•˜๊ณ  ์ตœ์ข… ์ถœ๋ ฅ ์ฝ”๋“œ์˜ ํ’ˆ์งˆ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ €๋น„์šฉ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. [24]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฝ”๋“œ ์ƒ์„ฑ์„ ์œ„ํ•œ ์ž์ฒด ํ˜‘์—… ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ๋Š” ์—ฌ๋Ÿฌ LLM์ด ํŠน์ • ํ•˜์œ„ ์ž‘์—…์— ๋Œ€ํ•œ ๋ณ„๊ฐœ์˜ "์ „๋ฌธ๊ฐ€"๋กœ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ง€์ •๋œ ์ง€์นจ์— ๋”ฐ๋ผ ํ˜‘์—…ํ•˜๊ณ  ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ์„œ๋กœ์˜ ์ž‘์—…์„ ์ด‰์ง„ํ•˜๋Š” ๊ฐ€์ƒ ํŒ€์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ถ๊ทน์ ์œผ๋กœ ๊ฐ€์ƒ ํŒ€์€ ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์—†์ด๋„ ์ฝ”๋“œ ์ƒ์„ฑ ์ž‘์—…์„ ํ˜‘๋ ฅ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค. LLIFT [139]๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ํŠนํžˆ ์ž ์žฌ์ ์ธ ์ฝ”๋“œ ์ทจ์•ฝ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ •์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ๋„์›€์„ ์ค๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •ํ™•์„ฑ๊ณผ ํ™•์žฅ์„ฑ ๊ฐ„์˜ ๊ท ํ˜•์„ ํšจ๊ณผ์ ์œผ๋กœ ๊ด€๋ฆฌํ•ฉ๋‹ˆ๋‹ค. ChatEDA [123]๋Š” ์ž‘์—… ๊ณ„ํš, ์Šคํฌ๋ฆฝํŠธ ์ƒ์„ฑ ๋ฐ ์‹คํ–‰์„ ํ†ตํ•ฉํ•˜์—ฌ ์„ค๊ณ„ ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ„์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ „์ž ์„ค๊ณ„ ์ž๋™ํ™”(EDA)๋ฅผ ์œ„ํ•ด ๊ฐœ๋ฐœ๋œ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค.
Large Language Models For Instruction Fine-Tuning.The initial endeavors in instruction-following training work primarily focused on enhancing the language model's capacity for generalization across diverse tasks. This often involves the process of fine-tuning across substantially available Natural Language Processing datasets, and evaluates on the different NLP tasks. T5 [86] undertake the earliest attempts to train a range of NLP tasks, including Question and Answer, Document Summarization, and Sentiment Classification, by employing a consistent prompt format across all the data. Subsequently, instruction fine-tuning work such as FLAN [87], ExT5 [88], T0 [89], UnifiedQA [90], ZeroPrompt [91], and FLAN-T5 [92] emerged to adapt for a large number of downstream tasks. To address the challenge of misalignment between model outputs and human requirements, OpenAI manually annotates the instruction library to construct a diverse range of tasks. Simultaneously, Reinforcement Learning from Human Feedback technology is employed, which facilitate the rapid development of LLMs such as InstructGPT [2], ChatGPT\({}^{\ast}\), GPT-4 [3]. To reduce manual involvement, self-instruct [93] improves instruction-following through self-generated instructions. Alpaca [22] used a dataset of 50k instructions generated from a limited (e.g., 175 samples) seed set of manually-written instructions. Vicuna [23] used 70k user-shared conversations with ChatGPT collected from ShareGPT.com. Meanwhile, WizardLM [24] introduces the evol-instruct approach, which seeks to refine the existing instruction data by enhancing both its complexity and diversity.
๋ช…๋ น์–ด Fine-tuning์„ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ. ๋ช…๋ น์–ด ํ›ˆ๋ จ ์ž‘์—…์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์€ ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑธ์ณ Fine-tuningํ•˜๋Š” ๊ณผ์ •์ด ํฌํ•จ๋˜๋ฉฐ, ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. T5 [86]๋Š” ๋ชจ๋“  ๋ฐ์ดํ„ฐ์— ์ผ๊ด€๋œ ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ ๋ฐ ๋‹ต๋ณ€, ๋ฌธ์„œ ์š”์•ฝ, ๊ฐ์ • ๋ถ„๋ฅ˜ ๋“ฑ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ›ˆ๋ จํ•˜๋Š” ์ตœ์ดˆ์˜ ์‹œ๋„๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ FLAN [87], ExT5 [88], T0 [89], UnifiedQA [90], ZeroPrompt [91], FLAN-T5 [92]์™€ ๊ฐ™์€ Instruction Fine-tuning ์ž‘์—…์ด ๋“ฑ์žฅํ•˜์—ฌ ๋‹ค์ˆ˜์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์ถœ๋ ฅ๊ณผ ์‚ฌ๋žŒ์˜ ์š”๊ตฌ ์‚ฌํ•ญ ๊ฐ„์˜ ์ •ํ•ฉ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด OpenAI๋Š” ์ˆ˜๋™์œผ๋กœ Instruction ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์ฃผ์„ ์ฒ˜๋ฆฌํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— Reinforcement Learning from Human Feedback ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ InstructGPT [2], ChatGPT\({}^{\ast}\), GPT-4 [3]์™€ ๊ฐ™์€ LLM์„ ๋น ๋ฅด๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜๋™ ์ž‘์—…์„ ์ค„์ด๊ธฐ ์œ„ํ•ด self-instruct [93]๋Š” ์ž์ฒด ์ƒ์„ฑ๋œ Instruction์„ ํ†ตํ•ด Instruction ํŒ”๋กœ์ž‰์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. Alpaca [22]๋Š” ์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑ๋œ ์ œํ•œ๋œ(์˜ˆ: 175๊ฐœ ์ƒ˜ํ”Œ) ์‹œ๋“œ ์„ธํŠธ์—์„œ ์ƒ์„ฑ๋œ 5๋งŒ ๊ฐœ์˜ Instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna [23]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ChatGPT์™€์˜ 7๋งŒ ๊ฑด์˜ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ, WizardLM [24]์€ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œ์ผœ ๊ธฐ์กด Instruction ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ์„ ํ•˜๋Š” evol-instruct ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ •์„ ์œ„ํ•œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ.๋ช…๋ น์–ด ํ›ˆ๋ จ ์ž‘์—…์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์€ ์ฃผ๋กœ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๊ฑธ์ณ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ณผ์ •์ด ํฌํ•จ๋˜๋ฉฐ, ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์— ๋Œ€ํ•ด ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. T5 [86]๋Š” ๋ชจ๋“  ๋ฐ์ดํ„ฐ์— ์ผ๊ด€๋œ ํ”„๋กฌํ”„ํŠธ ํ˜•์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ ๋ฐ ๋‹ต๋ณ€, ๋ฌธ์„œ ์š”์•ฝ, ๊ฐ์ • ๋ถ„๋ฅ˜ ๋“ฑ ๋‹ค์–‘ํ•œ NLP ์ž‘์—…์„ ํ›ˆ๋ จํ•˜๋Š” ์ตœ์ดˆ์˜ ์‹œ๋„๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ FLAN [87], ExT5 [88], T0 [89], UnifiedQA [90], ZeroPrompt [91], FLAN-T5 [92]์™€ ๊ฐ™์€ ๋ช…๋ น์–ด ๋ฏธ์„ธ ์กฐ์ • ์ž‘์—…์ด ๋“ฑ์žฅํ•˜์—ฌ ๋‹ค์ˆ˜์˜ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ์ ์‘ํ•˜๊ธฐ ์‹œ์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ์ถœ๋ ฅ๊ณผ ์‚ฌ๋žŒ์˜ ์š”๊ตฌ ์‚ฌํ•ญ ๊ฐ„์˜ ์ •ํ•ฉ์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด OpenAI๋Š” ์ˆ˜๋™์œผ๋กœ ๋ช…๋ น ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์ฃผ์„ ์ฒ˜๋ฆฌํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ์œผ๋กœ๋ถ€ํ„ฐ์˜ ๊ฐ•ํ™” ํ•™์Šต ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ InstructGPT [2], ChatGPT\({}^{\ast}\), GPT-4 [3]์™€ ๊ฐ™์€ LLM์„ ๋น ๋ฅด๊ฒŒ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜๋™ ์ž‘์—…์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์…€ํ”„ ์ธ์ŠคํŠธ๋ŸญํŠธ [93]๋Š” ์ž์ฒด ์ƒ์„ฑ๋œ ์ธ์ŠคํŠธ๋Ÿญ์…˜์„ ํ†ตํ•ด ์ธ์ŠคํŠธ๋Ÿญ์…˜ ํŒ”๋กœ์ž‰์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค. ์•ŒํŒŒ์นด [22]๋Š” ์ˆ˜๋™์œผ๋กœ ์ž‘์„ฑ๋œ ์ œํ•œ๋œ(์˜ˆ: 175๊ฐœ ์ƒ˜ํ”Œ) ์‹œ๋“œ ์„ธํŠธ์—์„œ ์ƒ์„ฑ๋œ 5๋งŒ ๊ฐœ์˜ ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. Vicuna [23]๋Š” ShareGPT.com์—์„œ ์ˆ˜์ง‘ํ•œ ChatGPT์™€์˜ 7๋งŒ ๊ฑด์˜ ์‚ฌ์šฉ์ž ๊ณต์œ  ๋Œ€ํ™”๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•œํŽธ, WizardLM [24]์€ ๋ณต์žก์„ฑ๊ณผ ๋‹ค์–‘์„ฑ์„ ๋ชจ๋‘ ํ–ฅ์ƒ์‹œ์ผœ ๊ธฐ์กด ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐœ์„ ํ•˜๋Š” evol-instruct ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค.
Large language models (LLMs) are routinely pre-trained on billions of tokens, only to start the process over again once new data becomes available. A much more efficient solution is to continually pre-train these models - saving significant compute compared to re-training. However, the distribution shift induced by new data typically results in degraded performance on previous data or poor adaptation to the new data. In this work, we show that a simple and scalable combination of learning rate (LR) re-warming, LR re-decaying, and replay of previous data is sufficient to match the performance of fully re-training from scratch on all available data, as measured by the final loss and the average score on several language model (LM) evaluation benchmarks. Specifically, we show this for a weak but realistic distribution shift between two commonly used LLM pre-training datasets (English\(\rightarrow\)English) and a stronger distribution shift (English\(\rightarrow\)German) at the 405M parameter model scale with large dataset sizes (hundreds of billions of tokens). Selecting the weak but realistic shift for larger-scale experiments, we also find that our continual learning strategies match the re-training baseline for a 10B parameter LLM. Our results demonstrate that LLMs can be successfully updated via simple and scalable continual learning strategies, matching the re-training baseline using only a fraction of the compute. Finally, inspired by previous work, we propose alternatives to the cosine learning rate schedule that help circumvent forgetting induced by LR re-warming and that are not bound to a fixed token budget.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ์ผ์ƒ์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ์ด ๊ณผ์ •์„ ๋‹ค์‹œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ํ›จ์”ฌ ๋” ํšจ์œจ์ ์ธ ์†”๋ฃจ์…˜์€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ธ๋ฐ, ์žฌํ•™์Šต์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์ปดํ“จํŒ… ๋น„์šฉ์„ ์ ˆ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ์˜ํ•ด ์œ ๋ฐœ๋˜๋Š” ๋ถ„ํฌ ๋ณ€ํ™”๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ด์ „ ๋ฐ์ดํ„ฐ์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ๋˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ ์‘๋ ฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” Learning Rate (LR) ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์†Œ, ์ด์ „ ๋ฐ์ดํ„ฐ ์žฌ์ƒ์˜ ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์กฐํ•ฉ๋งŒ์œผ๋กœ๋„ ์ตœ์ข… Loss์™€ ์—ฌ๋Ÿฌ ์–ธ์–ด ๋ชจ๋ธ(LM) ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ์˜ ํ‰๊ท  ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ๋ชจ๋“  ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ฒ˜์Œ๋ถ€ํ„ฐ ์™„์ „ํžˆ ์žฌํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ํŠนํžˆ, ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐœ์˜ LLM ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ(English\(\rightarrow\)English) ์‚ฌ์ด์˜ ์•ฝํ•˜์ง€๋งŒ ํ˜„์‹ค์ ์ธ ๋ถ„ํฌ ์ด๋™๊ณผ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ(์ˆ˜์ฒœ์–ต ๊ฐœ์˜ ํ† ํฐ)๋ฅผ ๊ฐ€์ง„ 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ๊ทœ๋ชจ์˜ ๋” ๊ฐ•๋ ฅํ•œ ๋ถ„ํฌ ์ด๋™(English\(\rightarrow\)German)์— ๋Œ€ํ•ด ์ด๋ฅผ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ๋” ํฐ ๊ทœ๋ชจ์˜ ์‹คํ—˜์„ ์œ„ํ•ด ์•ฝํ•˜์ง€๋งŒ ํ˜„์‹ค์ ์ธ ์‹œํ”„ํŠธ๋ฅผ ์„ ํƒํ•œ ๊ฒฐ๊ณผ, ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์šฐ๋ฆฌ์˜ ์ง€์† ํ•™์Šต ์ „๋žต์ด 100์–ต ํŒŒ๋ผ๋ฏธํ„ฐ LLM์˜ ๊ฒฝ์šฐ ์žฌํ•™์Šต ๊ธฐ์ค€๊ณผ ์ผ์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์ง€์† ํ•™์Šต ์ „๋žต์„ ํ†ตํ•ด LLM์„ ์„ฑ๊ณต์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ปดํ“จํŒ…์˜ ์ผ๋ถ€๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ์žฌํ•™์Šต ๊ธฐ์ค€์— ๋ถ€ํ•ฉํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด์ „ ์ž‘์—…์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ฝ”์‚ฌ์ธ Learning Rate ์ผ์ •์— ๋Œ€ํ•œ ๋Œ€์•ˆ์„ ์ œ์•ˆํ•˜์—ฌ LR ์žฌ๊ฐ€์—ด๋กœ ์ธํ•œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ํ”ผํ•˜๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์— ์–ฝ๋งค์ด์ง€ ์•Š๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ˆ˜์‹ญ์–ต ๊ฐœ์˜ ํ† ํฐ์— ๋Œ€ํ•ด ์ผ์ƒ์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์นœ ํ›„ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ์ด ๊ณผ์ •์„ ๋‹ค์‹œ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ํ›จ์”ฌ ๋” ํšจ์œจ์ ์ธ ์†”๋ฃจ์…˜์€ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ์ง€์†์ ์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ธ๋ฐ, ์žฌํ•™์Šต์— ๋น„ํ•ด ์ƒ๋‹นํ•œ ์ปดํ“จํŒ… ๋น„์šฉ์„ ์ ˆ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ์˜ํ•ด ์œ ๋ฐœ๋˜๋Š” ๋ถ„ํฌ ๋ณ€ํ™”๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ด์ „ ๋ฐ์ดํ„ฐ์˜ ์„ฑ๋Šฅ ์ €ํ•˜ ๋˜๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์ ์‘๋ ฅ ์ €ํ•˜๋กœ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ํ•™์Šต ์†๋„(LR) ์žฌ๊ฐ€์—ด, LR ์žฌ๊ฐ์†Œ, ์ด์ „ ๋ฐ์ดํ„ฐ ์žฌ์ƒ์˜ ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์กฐํ•ฉ๋งŒ์œผ๋กœ๋„ ์ตœ์ข… ์†์‹ค๊ณผ ์—ฌ๋Ÿฌ ์–ธ์–ด ๋ชจ๋ธ(LM) ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ์˜ ํ‰๊ท  ์ ์ˆ˜๋กœ ์ธก์ •ํ•œ ๋ชจ๋“  ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ฒ˜์Œ๋ถ€ํ„ฐ ์™„์ „ํžˆ ์žฌํ•™์Šตํ•˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ํŠนํžˆ, ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ๋‘ ๊ฐœ์˜ LLM ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ(English\(\rightarrow\)English) ์‚ฌ์ด์˜ ์•ฝํ•˜์ง€๋งŒ ํ˜„์‹ค์ ์ธ ๋ถ„ํฌ ์ด๋™๊ณผ ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํฌ๊ธฐ(์ˆ˜์ฒœ์–ต ๊ฐœ์˜ ํ† ํฐ)๋ฅผ ๊ฐ€์ง„ 405M ํŒŒ๋ผ๋ฏธํ„ฐ ๋ชจ๋ธ ๊ทœ๋ชจ์˜ ๋” ๊ฐ•๋ ฅํ•œ ๋ถ„ํฌ ์ด๋™(English\(\rightarrow\)German)์— ๋Œ€ํ•ด ์ด๋ฅผ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. ๋” ํฐ ๊ทœ๋ชจ์˜ ์‹คํ—˜์„ ์œ„ํ•ด ์•ฝํ•˜์ง€๋งŒ ํ˜„์‹ค์ ์ธ ์‹œํ”„ํŠธ๋ฅผ ์„ ํƒํ•œ ๊ฒฐ๊ณผ, ์šฐ๋ฆฌ๋Š” ๋˜ํ•œ ์šฐ๋ฆฌ์˜ ์ง€์† ํ•™์Šต ์ „๋žต์ด 100์–ต ํŒŒ๋ผ๋ฏธํ„ฐ LLM์˜ ๊ฒฝ์šฐ ์žฌํ•™์Šต ๊ธฐ์ค€๊ณผ ์ผ์น˜ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฐ๊ณผ๋Š” ๊ฐ„๋‹จํ•˜๊ณ  ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์ง€์† ํ•™์Šต ์ „๋žต์„ ํ†ตํ•ด LLM์„ ์„ฑ๊ณต์ ์œผ๋กœ ์—…๋ฐ์ดํŠธํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ปดํ“จํŒ…์˜ ์ผ๋ถ€๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ์žฌํ•™์Šต ๊ธฐ์ค€์— ๋ถ€ํ•ฉํ•œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ์ด์ „ ์ž‘์—…์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์ฝ”์‚ฌ์ธ ํ•™์Šต ์†๋„ ์ผ์ •์— ๋Œ€ํ•œ ๋Œ€์•ˆ์„ ์ œ์•ˆํ•˜์—ฌ LR ์žฌ๊ฐ€์—ด๋กœ ์ธํ•œ ์žŠ์–ด๋ฒ„๋ฆผ์„ ํ”ผํ•˜๊ณ  ๊ณ ์ •๋œ ํ† ํฐ ์˜ˆ์‚ฐ์— ์–ฝ๋งค์ด์ง€ ์•Š๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
FLAN(Chung et al., 2022) is one of the most systematic and comprehensive approaches among the many studies on instruction tuning. This approach fine-tunes the language model on the instruction-optimised dataset, scales the number of tasks and model size, and incorporates chain-of-thought data in the fine-tuning. Although the authors did not consider a specific approach to tuning instructions in RALM architecture, their work provides a valuable reference for future research. In the instruction fine-tuning of RALM, Lin et al. (2023) integrated in-context retrieval augmentation. This greatly reduces the likelihood of the language model being misled by irrelevant retrieval content. SAIL(Luo et al., 2023) builds language generation and instruction tracking capabilities on complex search results generated by internal and external search engines. Using a corpus of instruction tuning, they collect search results for each training case from different search APIs and domains, and construct a search-based training set containing a triplet of (instruction, grounding information, response). In contrast to training on instruction-tuned datasets, Madaan et al. (2022) and Lazaridou et al. (2022) propose to prompt large models directly from retrieved knowledge. Madaan et al. (2022) used GPT-3 to clarify memory pairings of recorded cases where the model misinterpreted the user's intention, as well as user feedback. This ensures that their system can generate enhanced prompts for each new query based on user feedback. In contrast, Lazaridou et al. (2022) uses few-shot prompts and answer reordering to improve inference computation.
FLAN(Chung et al., 2022)์€ Instruction Tuning์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ ์ค‘ ๊ฐ€์žฅ ์ฒด๊ณ„์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ Instruction-optimised dataset์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•˜๊ณ , ์ž‘์—… ์ˆ˜์™€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋ฉฐ, Fine-tuning์— chain-of-thought ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ RALM ์•„ํ‚คํ…์ฒ˜์—์„œ Instruction์„ ์กฐ์ •ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ์ด ์ž‘์—…์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๊ท€์ค‘ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Lin et al. (2023)์€ RALM์˜ Instruction Fine-tuning์—์„œ In-Context Retrieval Augmentation์„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์˜ํ•ด ์˜ค๋„๋  ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SAIL(Luo et al., 2023)์€ ๋‚ด๋ถ€ ๋ฐ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ์—”์ง„์—์„œ ์ƒ์„ฑ๋œ ๋ณต์žกํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์–ธ์–ด ์ƒ์„ฑ ๋ฐ Instruction ์ถ”์  ๊ธฐ๋Šฅ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. Instruction Tuning ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ํ›ˆ๋ จ ์‚ฌ๋ก€์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ API ๋ฐ ๋„๋ฉ”์ธ์—์„œ ์ˆ˜์ง‘ํ•˜๊ณ , (Instruction, grounding information, response)์˜ ํŠธ๋ฆฌํ”Œ๋ ›์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. Instruction-Tuned ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, Madaan et al. (2022)๊ณผ Lazaridou et al. (2022)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์—์„œ ์ง์ ‘ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ”„๋กฌํ”„ํŠธํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Madaan et al. (2022)์€ GPT-3๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•œ ๊ธฐ๋ก๋œ ์‚ฌ๋ก€์˜ ๋ฉ”๋ชจ๋ฆฌ ํŽ˜์–ด๋ง๊ณผ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๋ช…ํ™•ํžˆ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ด ์‹œ์Šคํ…œ์€ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ํ–ฅ์ƒ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Lazaridou et al. (2022)์€ Few-shot ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€ ์žฌ์ •๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๊ณ„์‚ฐ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
FLAN(Chung et al., 2022)์€ ์ง€์‹œ ํŠœ๋‹์— ๋Œ€ํ•œ ๋งŽ์€ ์—ฐ๊ตฌ ์ค‘ ๊ฐ€์žฅ ์ฒด๊ณ„์ ์ด๊ณ  ํฌ๊ด„์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹ ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ง€์‹œ ์ตœ์ ํ™” ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๊ณ , ์ž‘์—… ์ˆ˜์™€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ํ™•์žฅํ•˜๋ฉฐ, ๋ฏธ์„ธ ์กฐ์ •์— ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ RALM ์•„ํ‚คํ…์ฒ˜์—์„œ ์ง€์นจ์„ ์กฐ์ •ํ•˜๋Š” ๊ตฌ์ฒด์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ณ ๋ คํ•˜์ง€ ์•Š์•˜์ง€๋งŒ, ์ด ์ž‘์—…์€ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ๊ท€์ค‘ํ•œ ์ฐธ๊ณ  ์ž๋ฃŒ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Lin et al. (2023)์€ RALM์˜ ์ง€์นจ ๋ฏธ์„ธ ์กฐ์ •์—์„œ ์ปจํ…์ŠคํŠธ ๊ฒ€์ƒ‰ ์ฆ๊ฐ•์„ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ํ•˜๋ฉด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ด€๋ จ ์—†๋Š” ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์˜ํ•ด ์˜ค๋„๋  ๊ฐ€๋Šฅ์„ฑ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SAIL(Luo et al., 2023)์€ ๋‚ด๋ถ€ ๋ฐ ์™ธ๋ถ€ ๊ฒ€์ƒ‰ ์—”์ง„์—์„œ ์ƒ์„ฑ๋œ ๋ณต์žกํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ์— ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ์ง€์นจ ์ถ”์  ๊ธฐ๋Šฅ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ํŠœ๋‹ ์ฝ”ํผ์Šค๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ ํ›ˆ๋ จ ์‚ฌ๋ก€์— ๋Œ€ํ•œ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์–‘ํ•œ ๊ฒ€์ƒ‰ API ๋ฐ ๋„๋ฉ”์ธ์—์„œ ์ˆ˜์ง‘ํ•˜๊ณ , (์ง€์นจ, ๊ทธ๋ผ์šด๋”ฉ ์ •๋ณด, ์‘๋‹ต)์˜ ํŠธ๋ฆฌํ”Œ๋ ›์„ ํฌํ•จํ•˜๋Š” ๊ฒ€์ƒ‰ ๊ธฐ๋ฐ˜ ํ›ˆ๋ จ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํ›ˆ๋ จ๊ณผ๋Š” ๋‹ฌ๋ฆฌ, Madaan et al. (2022)๊ณผ Lazaridou et al. (2022)์€ ๊ฒ€์ƒ‰๋œ ์ง€์‹์—์„œ ์ง์ ‘ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ”„๋กฌํ”„ํŠธํ•˜๋„๋ก ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. Madaan et al. (2022)์€ GPT-3๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์‚ฌ์šฉ์ž์˜ ์˜๋„๋ฅผ ์ž˜๋ชป ํ•ด์„ํ•œ ๊ธฐ๋ก๋œ ์‚ฌ๋ก€์˜ ๋ฉ”๋ชจ๋ฆฌ ํŽ˜์–ด๋ง๊ณผ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๋ช…ํ™•ํžˆ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ด ์‹œ์Šคํ…œ์€ ์‚ฌ์šฉ์ž ํ”ผ๋“œ๋ฐฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ๊ฐ์˜ ์ƒˆ๋กœ์šด ์ฟผ๋ฆฌ์— ๋Œ€ํ•ด ํ–ฅ์ƒ๋œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— Lazaridou et al. (2022)์€ ์†Œ๋Ÿ‰ ํ”„๋กฌํ”„ํŠธ์™€ ๋‹ต๋ณ€ ์žฌ์ •๋ ฌ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก  ๊ณ„์‚ฐ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
1. **Automated Model Composition**: We introduce **Evolutionary Model Merge**, a general evolutionary method to automatically discover optimal combinations of diverse open-source models for creating new foundation models with user-specified capabilities. This approach harnesses the collective intelligence of existing open models, enabling the creation of powerful models without the need for extensive training data or compute. 2. **Cross-Domain Merging**: We demonstrate that our method can discover novel ways to merge models from disparate domains (e.g., non-English language and Math, non-English language and Vision), potentially exceeding the capabilities achievable through conventional human design strategies. 3. **State-of-the-Art Performance**: We showcase the effectiveness of our method by automatically generating a Japanese LLM with Math reasoning capability and a Japanese Vision-Language Model (VLM). Notably, both models achieve state-of-the-art performance on various benchmarks, even without explicit optimization for those tasks. 4. **High Efficiency and Surprising Generalizability**: We observe that our 7B parameter LLM surpasses the performance of some previous 70B parameter Japanese LLMs on benchmark datasets, highlighting the high efficiency and surprising generalization capability of our approach. We believe this model can serve as a strong general-purpose Japanese LLM. 5. **Culturally-Aware VLM**: The generated Japanese VLM achieves top results when tested on a domestically-sourced dataset of Japanese image-description pairs, demonstrating its ability to handle Japanese culture-specific content.
1. **์ž๋™ํ™”๋œ ๋ชจ๋ธ ๊ตฌ์„ฑ**: ์‚ฌ์šฉ์ž๊ฐ€ ์ง€์ •ํ•œ ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์ƒˆ๋กœ์šด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ์ตœ์  ์กฐํ•ฉ์„ ์ž๋™์œผ๋กœ ๊ฒ€์ƒ‰ํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ง„ํ™” ๋ฐฉ๋ฒ•์ธ **Evolutionary Model Merge**๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์กด ์˜คํ”ˆ ๋ชจ๋ธ์˜ ์ง‘๋‹จ ์ง€๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋ฐฉ๋Œ€ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋‚˜ ์ปดํ“จํŒ… ์—†์ด๋„ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. 2. **ํฌ๋กœ์Šค ๋„๋ฉ”์ธ ๋ณ‘ํ•ฉ**: ์ด ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๊ธฐ์กด์˜ ์ธ๊ฐ„ ์„ค๊ณ„ ์ „๋žต์„ ํ†ตํ•ด ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋›ฐ์–ด๋„˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ์‹์œผ๋กœ ์„œ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ(์˜ˆ: ๋น„์˜์–ด ์–ธ์–ด์™€ ์ˆ˜ํ•™, ๋น„์˜์–ด ์–ธ์–ด์™€ Vision)์˜ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ**: ์ˆ˜ํ•™ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์ผ๋ณธ์–ด LLM๊ณผ ์ผ๋ณธ์–ด Vision-Language Model (VLM)์„ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ด ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํŠนํžˆ ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ํ•ด๋‹น ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ์ตœ์ ํ™” ์—†์ด๋„ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. 4. **๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๋†€๋ผ์šด ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ**: 70์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ LLM์ด ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด์ „ 700์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ์ผ๋ถ€ ์ผ๋ณธ์–ด LLM์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•˜์—ฌ ๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๋†€๋ผ์šด ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๊ฐ•๋ ฅํ•œ ๋ฒ”์šฉ ์ผ๋ณธ์–ด LLM์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. 5. **๋ฌธํ™” ์ธ์‹ VLM**: ์ƒ์„ฑ๋œ ์ผ๋ณธ์–ด VLM์€ ๊ตญ๋‚ด์—์„œ ์กฐ์‚ฌํ•œ ์ผ๋ณธ์–ด ์ด๋ฏธ์ง€ ์„ค๋ช… ์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธํ–ˆ์„ ๋•Œ ์ตœ๊ณ ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ์ผ๋ณธ ๋ฌธํ™” ๊ด€๋ จ ์ฝ˜ํ…์ธ ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
1. **์ž๋™ํ™”๋œ ๋ชจ๋ธ ๊ตฌ์„ฑ**: ์‚ฌ์šฉ์ž๊ฐ€ ์ง€์ •ํ•œ ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ์ƒˆ๋กœ์šด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ์˜ ์ตœ์  ์กฐํ•ฉ์„ ์ž๋™์œผ๋กœ ๊ฒ€์ƒ‰ํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ์ง„ํ™” ๋ฐฉ๋ฒ•์ธ **์ง„ํ™”ํ˜• ๋ชจ๋ธ ๋ณ‘ํ•ฉ**์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ธฐ์กด ์˜คํ”ˆ ๋ชจ๋ธ์˜ ์ง‘๋‹จ ์ง€๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๋ฐฉ๋Œ€ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ๋‚˜ ์ปดํ“จํŒ… ์—†์ด๋„ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค๋‹ˆ๋‹ค. 2. **ํฌ๋กœ์Šค ๋„๋ฉ”์ธ ๋ณ‘ํ•ฉ**: ์ด ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ๊ธฐ์กด์˜ ์ธ๊ฐ„ ์„ค๊ณ„ ์ „๋žต์„ ํ†ตํ•ด ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋›ฐ์–ด๋„˜๋Š” ์ƒˆ๋กœ์šด ๋ฐฉ์‹์œผ๋กœ ์„œ๋กœ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ(์˜ˆ: ๋น„์˜์–ด ์–ธ์–ด์™€ ์ˆ˜ํ•™, ๋น„์˜์–ด ์–ธ์–ด์™€ ๋น„์ „)์˜ ๋ชจ๋ธ์„ ๋ณ‘ํ•ฉํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ**: ์ˆ˜ํ•™ ์ถ”๋ก  ๊ธฐ๋Šฅ์„ ๊ฐ–์ถ˜ ์ผ๋ณธ์–ด LLM๊ณผ ์ผ๋ณธ์–ด ๋น„์ „-์–ธ์–ด ๋ชจ๋ธ(VLM)์„ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜์—ฌ ์ด ๋ฐฉ๋ฒ•์˜ ํšจ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ํŠนํžˆ ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ํ•ด๋‹น ์ž‘์—…์— ๋Œ€ํ•œ ๋ช…์‹œ์ ์ธ ์ตœ์ ํ™” ์—†์ด๋„ ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ์ฒจ๋‹จ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. 4. **๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๋†€๋ผ์šด ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ์„ฑ**: 70์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ LLM์ด ๋ฒค์น˜๋งˆํฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด์ „ 700์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ์ผ๋ถ€ ์ผ๋ณธ์–ด LLM์˜ ์„ฑ๋Šฅ์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•˜์—ฌ ๋†’์€ ํšจ์œจ์„ฑ๊ณผ ๋†€๋ผ์šด ์ผ๋ฐ˜ํ™” ๊ธฐ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๊ฐ•๋ ฅํ•œ ๋ฒ”์šฉ ์ผ๋ณธ์–ด LLM์œผ๋กœ ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. 5. **๋ฌธํ™” ์ธ์‹ VLM**: ์ƒ์„ฑ๋œ ์ผ๋ณธ์–ด VLM์€ ๊ตญ๋‚ด์—์„œ ์กฐ์‚ฌํ•œ ์ผ๋ณธ์–ด ์ด๋ฏธ์ง€ ์„ค๋ช… ์Œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ…Œ์ŠคํŠธํ–ˆ์„ ๋•Œ ์ตœ๊ณ ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜์—ฌ ์ผ๋ณธ ๋ฌธํ™” ๊ด€๋ จ ์ฝ˜ํ…์ธ ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
**Model Utilization.** Based on the natural language interface, _prompting_ has become the prominent approach for using LLMs to solving various tasks. By combining task descriptions and demonstration examples into prompts, in-context learning (ICL) endows LLMs with the ability to perform well on new tasks, even outperforming full-data fine-tuned models in some cases. To enhance the ability of complex reasoning, advanced prompting techniques have been proposed, exemplified by the chain-of-thought (CoT) strategy, which includes the intermediate reasoning steps into prompts. Furthermore, planning is a promising approach for solving complex tasks, which iteratively invokes LLMs by leveraging tool use capacities. Despite these efforts, several basic problems related to prompting are still under-explored: why a good prompt can elicit the correct answer but a bad prompt cannot, how to reveal the working principles of advanced prompting methods (_e.g._, ICL and CoT) and further improve these existing approaches, and how to efficiently find the effective prompts for LLMs on specific tasks. Furthermore, from a practical perspective, it has become a fundamental challenge to reduce the inference cost of LLMs, especially in large-scale deployment. Another popular research direction is retrieval-augmented generation, where retrieved contexts from supporting sources are included into prompts for task solving. It has been shown that retrieval augmentation can extend the knowledge boundary and improve the question answering capacity [461], but may suffer from the effectiveness of long context utilization by LLMs [299].
**Model Utilization**. ์ž์—ฐ์–ด ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ _prompting_์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ, In-Context Learning (ICL)์€ LLM์ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋ฉฐ, ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ Fine-tuningํ•œ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋ณต์žกํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ๊ธ‰ prompting ๊ธฐ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์œผ๋ฉฐ, ํ”„๋กฌํ”„ํŠธ์— ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•˜๋Š” Chain-of-Thought (CoT) ์ „๋žต์ด ๋Œ€ํ‘œ์ ์ธ ์˜ˆ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณ„ํš์€ ๋„๊ตฌ ์‚ฌ์šฉ ๋Šฅ๋ ฅ์„ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ˜ธ์ถœํ•˜๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  prompting๊ณผ ๊ด€๋ จ๋œ ๋ช‡ ๊ฐ€์ง€ ๊ธฐ๋ณธ์ ์ธ ๋ฌธ์ œ๋Š” ์—ฌ์ „ํžˆ ๋ฏธ๊ฐœ์ฒ™ ์ƒํƒœ์ž…๋‹ˆ๋‹ค. ์ข‹์€ ํ”„๋กฌํ”„ํŠธ๋Š” ์–ด๋–ป๊ฒŒ ์ •๋‹ต์„ ์ด๋Œ์–ด๋‚ผ ์ˆ˜ ์žˆ์ง€๋งŒ ๋‚˜์œ ํ”„๋กฌํ”„ํŠธ๋Š” ๊ทธ๋ ‡์ง€ ๋ชปํ•œ๊ฐ€, ๊ณ ๊ธ‰ prompting ๋ฐฉ๋ฒ•(_์˜ˆ: ICL ๋ฐ CoT)์˜ ์ž‘๋™ ์›๋ฆฌ๋ฅผ ๋ฐํžˆ๊ณ  ๊ธฐ์กด ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•, ํŠน์ • ์ž‘์—…์—์„œ LLM์— ํšจ๊ณผ์ ์ธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฐพ๋Š” ๋ฐฉ๋ฒ• ๋“ฑ์ด ์—ฌ๊ธฐ์— ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹ค์šฉ์ ์ธ ๊ด€์ ์—์„œ ๋ณผ ๋•Œ, ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ ์‹œ LLM์˜ ์ถ”๋ก  ๋น„์šฉ์„ ์ค„์ด๋Š” ๊ฒƒ์€ ๊ธฐ๋ณธ์ ์ธ ๊ณผ์ œ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ ์ง€์› ์†Œ์Šค์—์„œ ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ํฌํ•จ์‹œ์ผœ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” Retrieval-augmented generation์ž…๋‹ˆ๋‹ค. Retrieval augmentation์€ ์ง€์‹์˜ ๊ฒฝ๊ณ„๋ฅผ ํ™•์žฅํ•˜๊ณ  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ, LLM์˜ ์žฅ๋ฌธ ๋ฌธ๋งฅ ํ™œ์šฉ ํšจ๊ณผ์— ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[461].
**๋ชจ๋ธ ํ™œ์šฉ**. ์ž์—ฐ์–ด ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ _ํ”„๋กฌํ”„ํŠธ_๋Š” ๋‹ค์–‘ํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์‚ฌ์šฉํ•˜๋Š” ๋ฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ž‘์—… ์„ค๋ช…๊ณผ ๋ฐ๋ชจ ์˜ˆ์ œ๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ๊ฒฐํ•ฉํ•จ์œผ๋กœ์จ, ์ปจํ…์ŠคํŠธ ๋‚ด ํ•™์Šต(ICL)์€ LLM์ด ์ƒˆ๋กœ์šด ์ž‘์—…์„ ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜๋ฉฐ, ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋ณต์žกํ•œ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ณ ๊ธ‰ ํ”„๋กฌํ”„ํŠธ ๊ธฐ๋ฒ•์ด ์ œ์•ˆ๋˜์—ˆ์œผ๋ฉฐ, ํ”„๋กฌํ”„ํŠธ์— ์ค‘๊ฐ„ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ํฌํ•จํ•˜๋Š” ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ(CoT) ์ „๋žต์ด ๋Œ€ํ‘œ์ ์ธ ์˜ˆ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณ„ํš์€ ๋„๊ตฌ ์‚ฌ์šฉ ๋Šฅ๋ ฅ์„ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ํ˜ธ์ถœํ•˜๋Š” ๋ณต์žกํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์œ ๋งํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ํ”„๋กฌํ”„ํŠธ์™€ ๊ด€๋ จ๋œ ๋ช‡ ๊ฐ€์ง€ ๊ธฐ๋ณธ์ ์ธ ๋ฌธ์ œ๋Š” ์—ฌ์ „ํžˆ ๋ฏธ๊ฐœ์ฒ™ ์ƒํƒœ์ž…๋‹ˆ๋‹ค. ์ข‹์€ ํ”„๋กฌํ”„ํŠธ๋Š” ์–ด๋–ป๊ฒŒ ์ •๋‹ต์„ ์ด๋Œ์–ด๋‚ผ ์ˆ˜ ์žˆ์ง€๋งŒ ๋‚˜์œ ํ”„๋กฌํ”„ํŠธ๋Š” ๊ทธ๋ ‡์ง€ ๋ชปํ•œ๊ฐ€, ๊ณ ๊ธ‰ ํ”„๋กฌํ”„ํŠธ ๋ฐฉ๋ฒ•(_์˜ˆ: ICL ๋ฐ CoT)์˜ ์ž‘๋™ ์›๋ฆฌ๋ฅผ ๋ฐํžˆ๊ณ  ๊ธฐ์กด ์ ‘๊ทผ ๋ฐฉ์‹์„ ๋”์šฑ ๊ฐœ์„ ํ•˜๋Š” ๋ฐฉ๋ฒ•, ํŠน์ • ์ž‘์—…์—์„œ LLM์— ํšจ๊ณผ์ ์ธ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํšจ์œจ์ ์œผ๋กœ ์ฐพ๋Š” ๋ฐฉ๋ฒ• ๋“ฑ์ด ์—ฌ๊ธฐ์— ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹ค์šฉ์ ์ธ ๊ด€์ ์—์„œ ๋ณผ ๋•Œ, ํŠนํžˆ ๋Œ€๊ทœ๋ชจ ๋ฐฐํฌ ์‹œ LLM์˜ ์ถ”๋ก  ๋น„์šฉ์„ ์ค„์ด๋Š” ๊ฒƒ์€ ๊ธฐ๋ณธ์ ์ธ ๊ณผ์ œ๊ฐ€ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ธ๊ธฐ ์žˆ๋Š” ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์€ ์ง€์› ์†Œ์Šค์—์„œ ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ํฌํ•จ์‹œ์ผœ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ์ž…๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰ ์ฆ๊ฐ•์€ ์ง€์‹์˜ ๊ฒฝ๊ณ„๋ฅผ ํ™•์žฅํ•˜๊ณ  ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ, LLM์˜ ์žฅ๋ฌธ ๋ฌธ๋งฅ ํ™œ์šฉ ํšจ๊ณผ์— ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ด ์ž…์ฆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[461].
Existing work considers Domain Adaptive Continual Pre-training (DACPT), a setting where a series of unlabelled domains become available to the LM sequentially and practitioners wish to train on each domain in a self-supervised fashion while retaining performance across each of them. While the objective is similar to our own, we consider general-purpose pre-training datasets that mix many domains as opposed to domain-specific datasets. Ke et al. (2022) assume data from previous domains is not available when training on new domains and develop a new technique for this setting which involves an importance mask of parameters for all previous tasks to prevent forgetting when pre-training with a masked language modeling (MLM) objective. Gururangan et al. (2020) investigated domain and task adaptive pre-training of RoBERTa (also MLM) and contributed a sample selection strategy for efficient continual pre-training. Similarly, Xie et al. (2023) also propose a data selection strategy that reduces the computational cost of continual pre-training (shown for autoregressive LLMs). Qin et al. (2023) investigate re-cycling fine-tuned adapter layers of previous base LMs as the initialization of new adapters for adapting continually updated versions of the base LM to specific tasks. Recently, Wu et al. (2024) proposed LLaMA Pro, a method for the continual pre-training of LLMs that enables learning new tasks without forgetting previous knowledge. However, unlike our work which considers adapting all existing weights, LLaMA Pro requires growing the size of the model for each new update and only adjusting the new weights.
๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ผ๋ จ์˜ ๋ผ๋ฒจ์ด ์—†๋Š” ๋„๋ฉ”์ธ์ด LM์— ์ˆœ์ฐจ์ ์œผ๋กœ ์ œ๊ณต๋˜๊ณ , ์‹ค๋ฌด์ž๊ฐ€ ๊ฐ ๋„๋ฉ”์ธ์—์„œ self-supervised ๋ฐฉ์‹์œผ๋กœ ํ•™์Šตํ•˜๋Š” ๋™์‹œ์— ๊ฐ ๋„๋ฉ”์ธ์—์„œ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๊ธฐ๋ฅผ ์›ํ•˜๋Š” ์„ค์ •์ธ Domain Adaptive Continual Pre-training (DACPT)์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋ชฉํ‘œ๋Š” ์ €ํฌ์™€ ๋น„์Šทํ•˜์ง€๋งŒ, ๋„๋ฉ”์ธ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์„ ํ˜ผํ•ฉํ•œ ๋ฒ”์šฉ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. Ke ๋“ฑ(2022)์€ ์ƒˆ๋กœ์šด ๋„๋ฉ”์ธ์—์„œ ํ•™์Šตํ•  ๋•Œ ์ด์ „ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๊ณ  masked language modeling (MLM) ๋ชฉํ‘œ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ ์žŠ์–ด๋ฒ„๋ฆฌ์ง€ ์•Š๋„๋ก ์ด์ „ ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ค‘์š”๋„ ๋งˆ์Šคํฌ๋ฅผ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ๊ธฐ๋ฒ•์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. Gururangan ๋“ฑ(2020)์€ RoBERTa(์—ญ์‹œ MLM)์˜ ๋„๋ฉ”์ธ ๋ฐ ์ž‘์—… adaptive pre-training์„ ์กฐ์‚ฌํ•˜๊ณ  ํšจ์œจ์ ์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์ƒ˜ํ”Œ ์„ ํƒ ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Xie ๋“ฑ(2023)๋„ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ค„์ด๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(autoregressive LLM์— ๋Œ€ํ•ด ์„ค๋ช…). Qin ๋“ฑ(2023)์€ ์ด์ „์— fine-tuning๋œ ์–ด๋Œ‘ํ„ฐ ๋ ˆ์ด์–ด๋ฅผ ๊ธฐ๋ณธ LM์˜ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋œ ๋ฒ„์ „์„ ํŠน์ • ์ž‘์—…์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์–ด๋Œ‘ํ„ฐ์˜ ์ดˆ๊ธฐํ™”๋กœ ์žฌํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Wu ๋“ฑ(2024)์€ ์ด์ „ ์ง€์‹์„ ์žŠ์ง€ ์•Š๊ณ  ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” LLM์˜ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์ธ LLaMA Pro๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด ๊ฐ€์ค‘์น˜๋ฅผ ๋ชจ๋‘ ์กฐ์ •ํ•˜๋Š” ์ €ํฌ์˜ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, LLaMA Pro๋Š” ์ƒˆ๋กœ์šด ์—…๋ฐ์ดํŠธ๊ฐ€ ์žˆ์„ ๋•Œ๋งˆ๋‹ค ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆฌ๊ณ  ์ƒˆ๋กœ์šด ๊ฐ€์ค‘์น˜๋งŒ ์กฐ์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์ผ๋ จ์˜ ๋ผ๋ฒจ์ด ์—†๋Š” ๋„๋ฉ”์ธ์ด LM์— ์ˆœ์ฐจ์ ์œผ๋กœ ์ œ๊ณต๋˜๊ณ , ์‹ค๋ฌด์ž๊ฐ€ ๊ฐ ๋„๋ฉ”์ธ์—์„œ ์ž๊ฐ€ ์ง€๋„ ๋ฐฉ์‹์œผ๋กœ ํ•™์Šตํ•˜๋Š” ๋™์‹œ์— ๊ฐ ๋„๋ฉ”์ธ์—์„œ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๊ธฐ๋ฅผ ์›ํ•˜๋Š” ์„ค์ •์ธ ๋„๋ฉ”์ธ ์ ์‘ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต(DACPT)์„ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ๋ชฉํ‘œ๋Š” ์ €ํฌ์™€ ๋น„์Šทํ•˜์ง€๋งŒ, ๋„๋ฉ”์ธ๋ณ„ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์•„๋‹Œ ์—ฌ๋Ÿฌ ๋„๋ฉ”์ธ์„ ํ˜ผํ•ฉํ•œ ๋ฒ”์šฉ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. Ke ๋“ฑ(2022)์€ ์ƒˆ๋กœ์šด ๋„๋ฉ”์ธ์—์„œ ํ•™์Šตํ•  ๋•Œ ์ด์ „ ๋„๋ฉ”์ธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๊ณ  ๋งˆ์Šคํ‚น๋œ ์–ธ์–ด ๋ชจ๋ธ๋ง(MLM) ๋ชฉํ‘œ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•  ๋•Œ ์žŠ์–ด๋ฒ„๋ฆฌ์ง€ ์•Š๋„๋ก ์ด์ „ ๋ชจ๋“  ์ž‘์—…์— ๋Œ€ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ์ค‘์š”๋„ ๋งˆ์Šคํฌ๋ฅผ ํฌํ•จํ•˜๋Š” ์ƒˆ๋กœ์šด ๊ธฐ๋ฒ•์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. Gururangan ๋“ฑ(2020)์€ RoBERTa(์—ญ์‹œ MLM)์˜ ๋„๋ฉ”์ธ ๋ฐ ์ž‘์—… ์ ์‘ ์‚ฌ์ „ ํ•™์Šต์„ ์กฐ์‚ฌํ•˜๊ณ  ํšจ์œจ์ ์ธ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•œ ์ƒ˜ํ”Œ ์„ ํƒ ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Xie ๋“ฑ(2023)๋„ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต์˜ ๊ณ„์‚ฐ ๋น„์šฉ์„ ์ค„์ด๋Š” ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค(์ž๋™ ํšŒ๊ท€ LLM์— ๋Œ€ํ•ด ์„ค๋ช…). Qin ๋“ฑ(2023)์€ ์ด์ „์— ๋ฏธ์„ธ ์กฐ์ •๋œ ์–ด๋Œ‘ํ„ฐ ๋ ˆ์ด์–ด๋ฅผ ๊ธฐ๋ณธ LM์˜ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋œ ๋ฒ„์ „์„ ํŠน์ • ์ž‘์—…์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ์–ด๋Œ‘ํ„ฐ์˜ ์ดˆ๊ธฐํ™”๋กœ ์žฌํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Wu ๋“ฑ(2024)์€ ์ด์ „ ์ง€์‹์„ ์žŠ์ง€ ์•Š๊ณ  ์ƒˆ๋กœ์šด ์ž‘์—…์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” LLM์˜ ์ง€์†์  ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์ธ LLaMA Pro๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด ๊ฐ€์ค‘์น˜๋ฅผ ๋ชจ๋‘ ์กฐ์ •ํ•˜๋Š” ์ €ํฌ์˜ ์—ฐ๊ตฌ์™€ ๋‹ฌ๋ฆฌ, LLaMA Pro๋Š” ์ƒˆ๋กœ์šด ์—…๋ฐ์ดํŠธ๊ฐ€ ์žˆ์„ ๋•Œ๋งˆ๋‹ค ๋ชจ๋ธ์˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆฌ๊ณ  ์ƒˆ๋กœ์šด ๊ฐ€์ค‘์น˜๋งŒ ์กฐ์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
Large Language Models (LLMs) have catalyzed significant advancements in Natural Language Processing (NLP), yet they encounter challenges such as hallucination and the need for domain-specific knowledge. To mitigate these, recent methodologies have integrated information retrieved from external resources with LLMs, substantially enhancing their performance across NLP tasks. This survey paper addresses the absence of a comprehensive overview on Retrieval-Augmented Language Models (RALMs), both Retrieval-Augmented Generation (RAG) and Retrieval-Augmented Understanding (RAU), providing an in-depth examination of their paradigm, evolution, taxonomy, and applications. The paper discusses the essential components of RALMs, including Retrievers, Language Models, and Augmentations, and how their interactions lead to diverse model structures and applications. RALMs demonstrate utility in a spectrum of tasks, from translation and dialogue systems to knowledge-intensive applications. The survey includes several evaluation methods of RALMs, emphasizing the importance of robustness, accuracy, and relevance in their assessment. It also acknowledges the limitations of RALMs, particularly in retrieval quality and computational efficiency, offering directions for future research. In conclusion, this survey aims to offer a structured insight into RALMs, their potential, and the avenues for their future development in NLP. The paper is supplemented with a Github Repository containing the surveyed works and resources for further study: [https://github.com/2471023025/RALM_Survey](https://github.com/2471023025/RALM_Survey).
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLMs)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋Œ์–ด๋ƒˆ์ง€๋งŒ, hallucination์ด๋‚˜ ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์˜ ํ•„์š”์„ฑ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ์— ์ง๋ฉดํ•ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ์˜ ๋ฐฉ๋ฒ•๋ก ์—์„œ๋Š” ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค์—์„œ ๊ฒ€์ƒ‰ํ•œ ์ •๋ณด๋ฅผ LLM๊ณผ ํ†ตํ•ฉํ•˜์—ฌ NLP ์ž‘์—… ์ „๋ฐ˜์—์„œ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ ๋ณด๊ณ ์„œ๋Š” Retrieval-Augmented Language Models (RALMs), ์ฆ‰ Retrieval-Augmented Generation (RAG)๊ณผ Retrieval-Augmented Understanding (RAU)์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ฐœ์š”๊ฐ€ ๋ถ€์กฑํ•˜๋‹ค๋Š” ์ ์„ ํ•ด๊ฒฐํ•˜๊ณ , ๊ทธ ํŒจ๋Ÿฌ๋‹ค์ž„, ์ง„ํ™”, ๋ถ„๋ฅ˜ ๋ฐ ์‘์šฉ์— ๋Œ€ํ•ด ์‹ฌ์ธต์ ์œผ๋กœ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” Retrievers, Language Models, Augmentations ๋“ฑ RALM์˜ ํ•„์ˆ˜ ๊ตฌ์„ฑ ์š”์†Œ์™€ ์ด๋“ค์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๊ตฌ์กฐ์™€ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ์ด์–ด์ง€๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. RALM์€ ๋ฒˆ์—ญ ๋ฐ ๋Œ€ํ™” ์‹œ์Šคํ…œ๋ถ€ํ„ฐ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์œ ์šฉ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—๋Š” RALM์˜ ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ํ‰๊ฐ€ ์‹œ robustness, ์ •ํ™•์„ฑ ๋ฐ ๊ด€๋ จ์„ฑ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ retrieval ํ’ˆ์งˆ๊ณผ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ์—์„œ RALM์˜ ํ•œ๊ณ„๋ฅผ ์ธ์ •ํ•˜๊ณ  ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๋ก ์ ์œผ๋กœ, ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” RALM๊ณผ ๊ทธ ์ž ์žฌ๋ ฅ, ๊ทธ๋ฆฌ๊ณ  NLP ๋ถ„์•ผ์—์„œ์˜ ํ–ฅํ›„ ๋ฐœ์ „ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ๋Š” ์กฐ์‚ฌ๋œ ์ €์ž‘๋ฌผ๊ณผ ์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๋ฆฌ์†Œ์Šค๊ฐ€ ํฌํ•จ๋œ Github ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋กœ ๋ณด์™„๋ฉ๋‹ˆ๋‹ค: [https://github.com/2471023025/RALM_Survey](https://github.com/2471023025/RALM_Survey).
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋Œ์–ด๋ƒˆ์ง€๋งŒ, ํ™˜๊ฐ์ด๋‚˜ ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์˜ ํ•„์š”์„ฑ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ์— ์ง๋ฉดํ•ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ๊ทผ์˜ ๋ฐฉ๋ฒ•๋ก ์—์„œ๋Š” ์™ธ๋ถ€ ๋ฆฌ์†Œ์Šค์—์„œ ๊ฒ€์ƒ‰ํ•œ ์ •๋ณด๋ฅผ LLM๊ณผ ํ†ตํ•ฉํ•˜์—ฌ NLP ์ž‘์—… ์ „๋ฐ˜์—์„œ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ ๋ณด๊ณ ์„œ๋Š” ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์–ธ์–ด ๋ชจ๋ธ(RALM), ์ฆ‰ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑ(RAG)๊ณผ ๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ดํ•ด(RAU)์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ๊ฐœ์š”๊ฐ€ ๋ถ€์กฑํ•˜๋‹ค๋Š” ์ ์„ ํ•ด๊ฒฐํ•˜๊ณ , ๊ทธ ํŒจ๋Ÿฌ๋‹ค์ž„, ์ง„ํ™”, ๋ถ„๋ฅ˜ ๋ฐ ์‘์šฉ์— ๋Œ€ํ•ด ์‹ฌ์ธต์ ์œผ๋กœ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๊ฒ€์ƒ‰๊ธฐ, ์–ธ์–ด ๋ชจ๋ธ, ์ฆ๊ฐ• ๋“ฑ RALM์˜ ํ•„์ˆ˜ ๊ตฌ์„ฑ ์š”์†Œ์™€ ์ด๋“ค์ด ์ƒํ˜ธ ์ž‘์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๊ตฌ์กฐ์™€ ์‘์šฉ ํ”„๋กœ๊ทธ๋žจ์œผ๋กœ ์ด์–ด์ง€๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. RALM์€ ๋ฒˆ์—ญ ๋ฐ ๋Œ€ํ™” ์‹œ์Šคํ…œ๋ถ€ํ„ฐ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์œ ์šฉ์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์„ค๋ฌธ์กฐ์‚ฌ์—๋Š” RALM์˜ ์—ฌ๋Ÿฌ ํ‰๊ฐ€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉฐ, ํ‰๊ฐ€ ์‹œ ๊ฒฌ๊ณ ์„ฑ, ์ •ํ™•์„ฑ ๋ฐ ๊ด€๋ จ์„ฑ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฒ€์ƒ‰ ํ’ˆ์งˆ๊ณผ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ์—์„œ RALM์˜ ํ•œ๊ณ„๋ฅผ ์ธ์ •ํ•˜๊ณ  ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๋ก ์ ์œผ๋กœ, ์ด ์„ค๋ฌธ์กฐ์‚ฌ๋Š” RALM๊ณผ ๊ทธ ์ž ์žฌ๋ ฅ, ๊ทธ๋ฆฌ๊ณ  NLP ๋ถ„์•ผ์—์„œ์˜ ํ–ฅํ›„ ๋ฐœ์ „ ๋ฐฉํ–ฅ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ๋Š” ์กฐ์‚ฌ๋œ ์ €์ž‘๋ฌผ๊ณผ ์ถ”๊ฐ€ ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•œ ๋ฆฌ์†Œ์Šค๊ฐ€ ํฌํ•จ๋œ Github ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋กœ ๋ณด์™„๋ฉ๋‹ˆ๋‹ค: [https://github.com/2471023025/RALM_Survey](https://github.com/2471023025/RALM_Survey).
As discussed before, language model is not a new technical concept specially for LLMs, but has evolved with the advance of artificial intelligence over the decades. Early language models mainly aim to model and generate text data, while latest language models (_e.g.,_ GPT-4) focus on complex task solving. From _language modeling_ to _task solving_, it is an important leap in scientific thinking, which is the key to understand the development of language models in the research history. From the perspective of task solving, the four generations of language models have exhibited different levels of model capacities. In Figure 2, we describe the evolution process of language models in terms of the task solving capacity. At first, statistical language models mainly assisted in some specific tasks (_e.g.,_ retrieval or speech tasks), in which the predicted or estimated probabilities can enhance the performance of task-specific approaches. Subsequently, neural language models focused on learning task-agnostic representations (_e.g.,_ features), aiming to reduce the efforts for human feature engineering. Furthermore, pre-trained language models learned context-aware representations that can be optimized according to downstream tasks. For the latest generation of language model, LLMs are enhanced by exploring the scaling effect on model capacity, which can be considered as general-purpose task solvers. To summarize, in the evolution process, the task scope that can be solved by language models have been greatly extended, and the task performance attained by language models have been significantly enhanced.
์•ž์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์–ธ์–ด ๋ชจ๋ธ์€ ํŠนํžˆ LLM์—๋งŒ ํ•ด๋‹น๋˜๋Š” ์ƒˆ๋กœ์šด ๊ธฐ์ˆ  ๊ฐœ๋…์ด ์•„๋‹ˆ๋ผ ์ˆ˜์‹ญ ๋…„์— ๊ฑธ์ณ ์ธ๊ณต ์ง€๋Šฅ์˜ ๋ฐœ์ „๊ณผ ํ•จ๊ป˜ ๋ฐœ์ „ํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์–ธ์–ด ๋ชจ๋ธ์€ ์ฃผ๋กœ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์ง€๋งŒ, ์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ(_์˜ˆ: GPT-4)์€ ๋ณต์žกํ•œ ์ž‘์—… ํ•ด๊ฒฐ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. _Language modeling_์—์„œ _task solving_๋กœ, ๊ณผํ•™์  ์‚ฌ๊ณ ์˜ ์ค‘์š”ํ•œ ๋„์•ฝ์€ ์—ฐ๊ตฌ ์—ญ์‚ฌ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์„ ์ดํ•ดํ•˜๋Š” ์—ด์‡ ์ž…๋‹ˆ๋‹ค. ์ž‘์—… ํ•ด๊ฒฐ์˜ ๊ด€์ ์—์„œ ๋ณผ ๋•Œ, 4์„ธ๋Œ€ ์–ธ์–ด ๋ชจ๋ธ์€ ์„œ๋กœ ๋‹ค๋ฅธ ์ˆ˜์ค€์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ์ง„ํ™” ๊ณผ์ •์„ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ ์ธก๋ฉด์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ํ†ต๊ณ„์  ์–ธ์–ด ๋ชจ๋ธ์ด ์ฃผ๋กœ ํŠน์ • ์ž‘์—…(_์˜ˆ: retrieval ๋˜๋Š” ์Œ์„ฑ ์ž‘์—…_)์„ ์ง€์›ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ๋Š”๋ฐ, ์˜ˆ์ธก ๋˜๋Š” ์ถ”์ • ํ™•๋ฅ ์„ ํ†ตํ•ด ์ž‘์—…๋ณ„ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์€ ์ธ๊ฐ„์˜ ํŠน์ง• ์—”์ง€๋‹ˆ์–ด๋ง์— ๋Œ€ํ•œ ๋…ธ๋ ฅ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ํ‘œํ˜„(_์˜ˆ: ํŠน์ง•_)์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ปจํ…์ŠคํŠธ ์ธ์‹ ํ‘œํ˜„์„ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ์ธ LLM์€ ์ผ๋ฐ˜์ ์ธ ๋ชฉ์ ์˜ ์ž‘์—… ์†”๋ฒ„๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ๋Œ€ํ•œ ํ™•์žฅ ํšจ๊ณผ๋ฅผ ํƒ์ƒ‰ํ•˜์—ฌ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์š”์•ฝํ•˜์ž๋ฉด, ์ง„ํ™” ๊ณผ์ •์—์„œ ์–ธ์–ด ๋ชจ๋ธ์ด ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์—… ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜์—ˆ๊ณ , ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ฌ์„ฑํ•œ ์ž‘์—… ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
์•ž์„œ ์„ค๋ช…ํ•œ ๋ฐ”์™€ ๊ฐ™์ด ์–ธ์–ด ๋ชจ๋ธ์€ ํŠนํžˆ LLM์—๋งŒ ํ•ด๋‹น๋˜๋Š” ์ƒˆ๋กœ์šด ๊ธฐ์ˆ  ๊ฐœ๋…์ด ์•„๋‹ˆ๋ผ ์ˆ˜์‹ญ ๋…„์— ๊ฑธ์ณ ์ธ๊ณต ์ง€๋Šฅ์˜ ๋ฐœ์ „๊ณผ ํ•จ๊ป˜ ๋ฐœ์ „ํ•ด ์™”์Šต๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์–ธ์–ด ๋ชจ๋ธ์€ ์ฃผ๋กœ ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ๋ธ๋งํ•˜๊ณ  ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์ง€๋งŒ, ์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ(_์˜ˆ: GPT-4)์€ ๋ณต์žกํ•œ ์ž‘์—… ํ•ด๊ฒฐ์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์–ธ์–ด ๋ชจ๋ธ๋ง_์—์„œ _์ž‘์—… ํ•ด๊ฒฐ_๋กœ, ๊ณผํ•™์  ์‚ฌ๊ณ ์˜ ์ค‘์š”ํ•œ ๋„์•ฝ์€ ์—ฐ๊ตฌ ์—ญ์‚ฌ์—์„œ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋ฐœ์ „์„ ์ดํ•ดํ•˜๋Š” ์—ด์‡ ์ž…๋‹ˆ๋‹ค. ์ž‘์—… ํ•ด๊ฒฐ์˜ ๊ด€์ ์—์„œ ๋ณผ ๋•Œ, 4์„ธ๋Œ€ ์–ธ์–ด ๋ชจ๋ธ์€ ์„œ๋กœ ๋‹ค๋ฅธ ์ˆ˜์ค€์˜ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ์ง„ํ™” ๊ณผ์ •์„ ์ž‘์—… ํ•ด๊ฒฐ ๋Šฅ๋ ฅ ์ธก๋ฉด์—์„œ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ํ†ต๊ณ„์  ์–ธ์–ด ๋ชจ๋ธ์ด ์ฃผ๋กœ ํŠน์ • ์ž‘์—…(์˜ˆ: ๊ฒ€์ƒ‰ ๋˜๋Š” ์Œ์„ฑ ์ž‘์—…)์„ ์ง€์›ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ๋Š”๋ฐ, ์˜ˆ์ธก ๋˜๋Š” ์ถ”์ • ํ™•๋ฅ ์„ ํ†ตํ•ด ์ž‘์—…๋ณ„ ์ ‘๊ทผ ๋ฐฉ์‹์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ์‹ ๊ฒฝ ์–ธ์–ด ๋ชจ๋ธ์€ ์ธ๊ฐ„์˜ ํŠน์ง• ์—”์ง€๋‹ˆ์–ด๋ง์— ๋Œ€ํ•œ ๋…ธ๋ ฅ์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์ž‘์—…์— ๊ตฌ์• ๋ฐ›์ง€ ์•Š๋Š” ํ‘œํ˜„(_์˜ˆ: ํŠน์ง•)์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‚ฌ์ „ ํ•™์Šต๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์— ๋”ฐ๋ผ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ๋Š” ์ปจํ…์ŠคํŠธ ์ธ์‹ ํ‘œํ˜„์„ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ์ตœ์‹  ์–ธ์–ด ๋ชจ๋ธ์ธ LLM์€ ์ผ๋ฐ˜์ ์ธ ๋ชฉ์ ์˜ ์ž‘์—… ์†”๋ฒ„๋กœ ๊ฐ„์ฃผํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ ์šฉ๋Ÿ‰์— ๋Œ€ํ•œ ํ™•์žฅ ํšจ๊ณผ๋ฅผ ํƒ์ƒ‰ํ•˜์—ฌ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์š”์•ฝํ•˜์ž๋ฉด, ์ง„ํ™” ๊ณผ์ •์—์„œ ์–ธ์–ด ๋ชจ๋ธ์ด ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž‘์—… ๋ฒ”์œ„๊ฐ€ ํฌ๊ฒŒ ํ™•์žฅ๋˜์—ˆ๊ณ , ์–ธ์–ด ๋ชจ๋ธ์ด ๋‹ฌ์„ฑํ•œ ์ž‘์—… ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
We perform systematic perturbations on each question/answer pair (as visualized in Figure 1. In three datasets with numerical answers (Drug Dosages, Sports Records, Latest News), we produce ten modifications that act as multipliers on the original value: \(0.1,0.2,0.4,0.8,1.2,1.5,2.0,3.0,5.0,10.0\). In the Wikipedia Years dataset, we perform ten absolute modifications in increments of 20 years for a range of \([-100,100]\). For the Wikipedia Names and Locations, the discrete categories required more hand-crafted levels of variation. For each, we performed three categorical perturbations via prompting: slight, significant,and comical. We provide the full prompts used in our study in the Appendix. For example, for a name like _Bob Green_, a slight modification implies a small tweak to another real name (_Rob Greene_), whereas a significant modification produces a similar but fictitious name (_Bilgorn Grevalle_), and a comical modification is an absurd variant (_Blob Lawnface_). For a city name like _Miami_, a slight modification changes the name of the most similar city (_Fort Lauderdale_), a significant modification produces a fictitious city name (_Marisole_), and a comical modification produces an absurd variant (_Miameme_). Because of differences in how each modified fact might appear in the retrieved text, we utilize GPT-4 to generate the perturbed excerpts for drug dosages and news. Each modified fact is replaced in the original retrieved text. Then, both the question and context are posed to GPT-4, from which the answers, along with the log probabilities of the output tokens, are collected.
๊ฐ ์งˆ๋ฌธ/๋‹ต๋ณ€ ์Œ์— ๋Œ€ํ•ด ์ฒด๊ณ„์ ์ธ perturbation์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1์—์„œ ์‹œ๊ฐํ™”๋œ ๊ฒƒ์ฒ˜๋Ÿผ). ์ˆซ์ž ๋‹ต๋ณ€์ด ์žˆ๋Š” ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ(์•ฝ๋ฌผ ๋ณต์šฉ๋Ÿ‰, ์Šคํฌ์ธ  ๊ธฐ๋ก, ์ตœ์‹  ๋‰ด์Šค)์—์„œ ์›๋ž˜ ๊ฐ’์— ๊ณฑ์…ˆ์œผ๋กœ ์ž‘์šฉํ•˜๋Š” 10๊ฐ€์ง€ ์ˆ˜์ •์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(0.1,0.2,0.4,0.8,1.2,1.5,2.0,3.0,5.0,10.0\). Wikipedia ์—ฐ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” \([-100,100]\) ๋ฒ”์œ„์— ๋Œ€ํ•ด 20๋…„ ๋‹จ์œ„๋กœ 10๊ฐœ์˜ ์ ˆ๋Œ€ ์ˆ˜์ •์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Wikipedia ์ด๋ฆ„ ๋ฐ ์œ„์น˜์˜ ๊ฒฝ์šฐ ์ด์‚ฐ ๋ฒ”์ฃผ๋Š” ์ข€ ๋” ์ˆ˜์ž‘์—…์œผ๋กœ ๋ณ€ํ˜• ์ˆ˜์ค€์„ ๋†’์—ฌ์•ผ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ๊ฐ์— ๋Œ€ํ•ด ์•ฝ๊ฐ„, ์ƒ๋‹น, ์ฝ”๋ฏน์˜ ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ perturbation์„ prompting์„ ํ†ตํ•ด ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ์ „์ฒด prompt๋Š” ๋ถ€๋ก์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, _๋ฐฅ ๊ทธ๋ฆฐ_๊ณผ ๊ฐ™์€ ์ด๋ฆ„์˜ ๊ฒฝ์šฐ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋ฉด ๋‹ค๋ฅธ ์‹ค์ œ ์ด๋ฆ„(_๋กญ ๊ทธ๋ฆฐ_)์„ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ์ด๊ณ , ์ƒ๋‹นํ•œ ์ˆ˜์ •์€ ๋น„์Šทํ•˜์ง€๋งŒ ๊ฐ€์ƒ์˜ ์ด๋ฆ„(_๋นŒ๊ณค ๊ทธ๋ ˆ๋ฐœ๋ ˆ_)์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ฝ”๋ฏนํ•œ ์ˆ˜์ •์€ ํ„ฐ๋ฌด๋‹ˆ์—†๋Š” ๋ณ€ํ˜•(_๋ธ”๋กญ ์ž”๋””๋ฐญ_)์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. _๋งˆ์ด์• ๋ฏธ_์™€ ๊ฐ™์€ ๋„์‹œ ์ด๋ฆ„์˜ ๊ฒฝ์šฐ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋ฉด ๊ฐ€์žฅ ์œ ์‚ฌํ•œ ๋„์‹œ ์ด๋ฆ„(_ํฌํŠธ ๋กœ๋”๋ฐ์ผ_)์ด ๋ณ€๊ฒฝ๋˜๊ณ , ์ƒ๋‹นํ•œ ์ˆ˜์ •์€ ๊ฐ€์ƒ์˜ ๋„์‹œ ์ด๋ฆ„(_๋งˆ๋ฆฌ์†”_)์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ฝ”๋ฏนํ•œ ์ˆ˜์ •์€ ํ„ฐ๋ฌด๋‹ˆ์—†๋Š” ๋ณ€ํ˜•(_๋ฏธ์•„๋ฉ”๋ฉ”_)์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •๋œ ๊ฐ ์‚ฌ์‹ค์ด ๊ฒ€์ƒ‰๋œ ํ…์ŠคํŠธ์— ํ‘œ์‹œ๋˜๋Š” ๋ฐฉ์‹์ด ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ์•ฝ๋ฌผ ๋ณต์šฉ๋Ÿ‰ ๋ฐ ๋‰ด์Šค์— ๋Œ€ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ perturb๋œ ๋ฐœ์ทŒ๋ฌธ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ ๊ฒ€์ƒ‰๋œ ํ…์ŠคํŠธ์—์„œ ์ˆ˜์ •๋œ ๊ฐ ์‚ฌ์‹ค์„ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์งˆ๋ฌธ๊ณผ ๋ฌธ๋งฅ์„ ๋ชจ๋‘ GPT-4์— ์ œ์‹œํ•˜๊ณ , ์ถœ๋ ฅ ํ† ํฐ์˜ ๋กœ๊ทธ ํ™•๋ฅ ๊ณผ ํ•จ๊ป˜ ๋‹ต๋ณ€์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
๊ฐ ์งˆ๋ฌธ/๋‹ต๋ณ€ ์Œ์— ๋Œ€ํ•ด ์ฒด๊ณ„์ ์ธ ์„ญ๋™์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1์—์„œ ์‹œ๊ฐํ™”๋œ ๊ฒƒ์ฒ˜๋Ÿผ). ์ˆซ์ž ๋‹ต๋ณ€์ด ์žˆ๋Š” ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ(์•ฝ๋ฌผ ๋ณต์šฉ๋Ÿ‰, ์Šคํฌ์ธ  ๊ธฐ๋ก, ์ตœ์‹  ๋‰ด์Šค)์—์„œ ์›๋ž˜ ๊ฐ’์— ๊ณฑ์…ˆ์œผ๋กœ ์ž‘์šฉํ•˜๋Š” 10๊ฐ€์ง€ ์ˆ˜์ •์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค: \(0.1,0.2,0.4,0.8,1.2,1.5,2.0,3.0,5.0,10.0\). Wikipedia ์—ฐ๋„ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ๋Š” \([-100,100]\) ๋ฒ”์œ„์— ๋Œ€ํ•ด 20๋…„ ๋‹จ์œ„๋กœ 10๊ฐœ์˜ ์ ˆ๋Œ€ ์ˆ˜์ •์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. Wikipedia ์ด๋ฆ„ ๋ฐ ์œ„์น˜์˜ ๊ฒฝ์šฐ ์ด์‚ฐ ๋ฒ”์ฃผ๋Š” ์ข€ ๋” ์ˆ˜์ž‘์—…์œผ๋กœ ๋ณ€ํ˜• ์ˆ˜์ค€์„ ๋†’์—ฌ์•ผ ํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ๊ฐ์— ๋Œ€ํ•ด ์•ฝ๊ฐ„, ์ƒ๋‹น, ์ฝ”๋ฏน์˜ ์„ธ ๊ฐ€์ง€ ๋ฒ”์ฃผ ์„ญ๋™์„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์—ฐ๊ตฌ์— ์‚ฌ์šฉ๋œ ์ „์ฒด ํ”„๋กฌํ”„ํŠธ๋Š” ๋ถ€๋ก์— ๋‚˜์™€ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, _๋ฐฅ ๊ทธ๋ฆฐ_๊ณผ ๊ฐ™์€ ์ด๋ฆ„์˜ ๊ฒฝ์šฐ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋ฉด ๋‹ค๋ฅธ ์‹ค์ œ ์ด๋ฆ„(_๋กญ ๊ทธ๋ฆฐ_)์„ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ์ด๊ณ , ์ƒ๋‹นํ•œ ์ˆ˜์ •์€ ๋น„์Šทํ•˜์ง€๋งŒ ๊ฐ€์ƒ์˜ ์ด๋ฆ„(_๋นŒ๊ณค ๊ทธ๋ ˆ๋ฐœ๋ ˆ_)์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ฝ”๋ฏนํ•œ ์ˆ˜์ •์€ ํ„ฐ๋ฌด๋‹ˆ์—†๋Š” ๋ณ€ํ˜•(_๋ธ”๋กญ ์ž”๋””๋ฐญ_)์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ด์• ๋ฏธ_์™€ ๊ฐ™์€ ๋„์‹œ ์ด๋ฆ„์˜ ๊ฒฝ์šฐ ์•ฝ๊ฐ„ ์ˆ˜์ •ํ•˜๋ฉด ๊ฐ€์žฅ ์œ ์‚ฌํ•œ ๋„์‹œ ์ด๋ฆ„(_ํฌํŠธ ๋กœ๋”๋ฐ์ผ_)์ด ๋ณ€๊ฒฝ๋˜๊ณ , ์ƒ๋‹นํ•œ ์ˆ˜์ •์€ ๊ฐ€์ƒ์˜ ๋„์‹œ ์ด๋ฆ„(_๋งˆ๋ฆฌ์†”_)์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ฝ”๋ฏนํ•œ ์ˆ˜์ •์€ ํ„ฐ๋ฌด๋‹ˆ์—†๋Š” ๋ณ€ํ˜•(_๋ฏธ์•„๋ฉ”๋ฉ”_)์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ •๋œ ๊ฐ ์‚ฌ์‹ค์ด ๊ฒ€์ƒ‰๋œ ํ…์ŠคํŠธ์— ํ‘œ์‹œ๋˜๋Š” ๋ฐฉ์‹์ด ๋‹ค๋ฅด๊ธฐ ๋•Œ๋ฌธ์— ์•ฝ๋ฌผ ๋ณต์šฉ๋Ÿ‰ ๋ฐ ๋‰ด์Šค์— ๋Œ€ํ•ด GPT-4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ญ๋™๋œ ๋ฐœ์ทŒ๋ฌธ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ ๊ฒ€์ƒ‰๋œ ํ…์ŠคํŠธ์—์„œ ์ˆ˜์ •๋œ ๊ฐ ์‚ฌ์‹ค์„ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์งˆ๋ฌธ๊ณผ ๋ฌธ๋งฅ์„ ๋ชจ๋‘ GPT-4์— ์ œ์‹œํ•˜๊ณ , ์ถœ๋ ฅ ํ† ํฐ์˜ ๋กœ๊ทธ ํ™•๋ฅ ๊ณผ ํ•จ๊ป˜ ๋‹ต๋ณ€์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Effective reward model training_. Despite that InstructGPT used a small reward model (6B GPT model), increasing work [99] has shown it is often more effective to use a large reward model (_e.g._, equal or greater than the original model size), since large reward models generally perform better in judging the quality of the LLM generated outputs. In LLMaMa 2 [99], pretrained chat model checkpoints are used to initialize the reward model, they argue that such an approach can effectively reduce the information mismatch between the model to be aligned and the reward model by sharing the same pre-training knowledge. Whereas, it is common to encounter the overfitting problem when training large-scale reward models. As a simple yet effective solution, existing work [374, 375] has introduced the LM loss on the preferred response of the input prompt from the human-annotated alignment dataset as a regularizer, which alleviates the overfitting of the reward model on the binary classification task. In addition, as there are multiple criteria for alignment (_e.g._, helpfulness and honesty), it is often difficult to train a single reward model that can satisfy all the alignment criteria. Therefore, it is useful to train multiple reward models that focus on different alignment criteria [99], and compute the final reward based on the produced ones from them via special combination strategies (_e.g._, mean pooling and weighted sum). Such a way enables more flexible rules or standards on multiple criteria, _e.g._, relaxing the requirement on helpfulness while posing more strict limits on harmfulness.
\(\bullet\)_ํšจ๊ณผ์ ์ธ reward model ํ›ˆ๋ จ_. InstructGPT๋Š” ์†Œ๊ทœ๋ชจ reward model(6B GPT ๋ชจ๋ธ)์„ ์‚ฌ์šฉํ–ˆ์ง€๋งŒ, ์ž‘์—… ์ฆ๊ฐ€[99]๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋Œ€๊ทœ๋ชจ reward model์ด LLM ์ƒ์„ฑ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํŒ๋‹จํ•˜๋Š” ๋ฐ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๊ทœ๋ชจ reward model์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ๊ณผ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. LLMaMa 2 [99]์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์ฑ„ํŒ… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ reward model์„ ์ดˆ๊ธฐํ™”ํ•˜๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ง€์‹์„ ๊ณต์œ ํ•จ์œผ๋กœ์จ ์ •๋ ฌํ•  ๋ชจ๋ธ๊ณผ reward model ๊ฐ„์˜ ์ •๋ณด ๋ถˆ์ผ์น˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋Œ€๊ทœ๋ชจ reward model์„ ํ›ˆ๋ จํ•  ๋•Œ ๊ณผ์ ํ•ฉ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ์†”๋ฃจ์…˜์œผ๋กœ, ๊ธฐ์กด ์ž‘์—… [374, 375]์—์„œ๋Š” ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์€ ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์—์„œ ์„ ํ˜ธํ•˜๋Š” ์‘๋‹ต์— ๋Œ€ํ•ด LM loss๋ฅผ ์ •๊ทœํ™” ๋„๊ตฌ๋กœ ๋„์ž…ํ•˜์—ฌ ์ด์ง„ ๋ถ„๋ฅ˜ ์ž‘์—…์—์„œ reward model์˜ ๊ณผ์ ํ•ฉ์„ ์™„ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ •๋ ฌ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ค€(_์˜ˆ: ์œ ์šฉ์„ฑ ๋ฐ ์ •์ง์„ฑ)์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ์ •๋ ฌ ๊ธฐ์ค€์„ ๋งŒ์กฑํ•  ์ˆ˜ ์žˆ๋Š” ๋‹จ์ผ reward model์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ์ข…์ข… ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„œ๋กœ ๋‹ค๋ฅธ ์ •๋ ฌ ๊ธฐ์ค€์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฌ๋Ÿฌ reward model์„ ํ›ˆ๋ จํ•˜๊ณ  ํŠน๋ณ„ํ•œ ์กฐํ•ฉ ์ „๋žต(_์˜ˆ: mean pooling ๋ฐ weighted sum)์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ reward model์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ตœ์ข… ๋ณด์ƒ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜๋ฉด ์œ ์šฉ์„ฑ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์™„ํ™”ํ•˜๋ฉด์„œ ์œ ํ•ด์„ฑ์— ๋Œ€ํ•ด์„œ๋Š” ๋” ์—„๊ฒฉํ•œ ์ œํ•œ์„ ๊ฐ€ํ•˜๋Š” ๋“ฑ ์—ฌ๋Ÿฌ ๊ธฐ์ค€์— ๋Œ€ํ•œ ๋ณด๋‹ค ์œ ์—ฐํ•œ ๊ทœ์น™์ด๋‚˜ ํ‘œ์ค€์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_ํšจ๊ณผ์ ์ธ ๋ณด์ƒ ๋ชจ๋ธ ํ›ˆ๋ จ_. InstructGPT๋Š” ์†Œ๊ทœ๋ชจ ๋ณด์ƒ ๋ชจ๋ธ(6B GPT ๋ชจ๋ธ)์„ ์‚ฌ์šฉํ–ˆ์ง€๋งŒ, ์ž‘์—… ์ฆ๊ฐ€[99]๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋Œ€๊ทœ๋ชจ ๋ณด์ƒ ๋ชจ๋ธ์ด LLM ์ƒ์„ฑ ์ถœ๋ ฅ์˜ ํ’ˆ์งˆ์„ ํŒ๋‹จํ•˜๋Š” ๋ฐ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋Œ€๊ทœ๋ชจ ๋ณด์ƒ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ๊ณผ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์ด ๋ฐํ˜€์กŒ์Šต๋‹ˆ๋‹ค. LLMaMa 2 [99]์—์„œ๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ์ฑ„ํŒ… ๋ชจ๋ธ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•˜๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ง€์‹์„ ๊ณต์œ ํ•จ์œผ๋กœ์จ ์ •๋ ฌํ•  ๋ชจ๋ธ๊ณผ ๋ณด์ƒ ๋ชจ๋ธ ๊ฐ„์˜ ์ •๋ณด ๋ถˆ์ผ์น˜๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ๋Œ€๊ทœ๋ชจ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ๋•Œ ๊ณผ์ ํ•ฉ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•˜๋ฉด์„œ๋„ ํšจ๊ณผ์ ์ธ ์†”๋ฃจ์…˜์œผ๋กœ, ๊ธฐ์กด ์ž‘์—… [374, 375]์—์„œ๋Š” ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์€ ์ •๋ ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ์—์„œ ์„ ํ˜ธํ•˜๋Š” ์‘๋‹ต์— ๋Œ€ํ•ด LM ์†์‹ค์„ ์ •๊ทœํ™” ๋„๊ตฌ๋กœ ๋„์ž…ํ•˜์—ฌ ์ด์ง„ ๋ถ„๋ฅ˜ ์ž‘์—…์—์„œ ๋ณด์ƒ ๋ชจ๋ธ์˜ ๊ณผ์ ํ•ฉ์„ ์™„ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ •๋ ฌ์—๋Š” ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ๊ธฐ์ค€(_์˜ˆ: ์œ ์šฉ์„ฑ ๋ฐ ์ •์ง์„ฑ)์ด ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ๋ชจ๋“  ์ •๋ ฌ ๊ธฐ์ค€์„ ๋งŒ์กฑํ•  ์ˆ˜ ์žˆ๋Š” ๋‹จ์ผ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ์ด ์ข…์ข… ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์„œ๋กœ ๋‹ค๋ฅธ ์ •๋ ฌ ๊ธฐ์ค€์— ์ดˆ์ ์„ ๋งž์ถ˜ ์—ฌ๋Ÿฌ ๋ณด์ƒ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํŠน๋ณ„ํ•œ ์กฐํ•ฉ ์ „๋žต(_์˜ˆ: ํ‰๊ท  ํ’€๋ง ๋ฐ ๊ฐ€์ค‘ ํ•ฉ๊ณ„)์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋ณด์ƒ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ตœ์ข… ๋ณด์ƒ์„ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•˜๋ฉด ์œ ์šฉ์„ฑ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์™„ํ™”ํ•˜๋ฉด์„œ ์œ ํ•ด์„ฑ์— ๋Œ€ํ•ด์„œ๋Š” ๋” ์—„๊ฒฉํ•œ ์ œํ•œ์„ ๊ฐ€ํ•˜๋Š” ๋“ฑ ์—ฌ๋Ÿฌ ๊ธฐ์ค€์— ๋Œ€ํ•œ ๋ณด๋‹ค ์œ ์—ฐํ•œ ๊ทœ์น™์ด๋‚˜ ํ‘œ์ค€์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
The key characteristic of our benchmark is that for each logical reasoning problem, we **synthetically generate variants** with **different premise orders.** Specifically, we denote the order that conforms to the ground truth proof with forward chaining as the _forward_ order, where the rule applied in each derivation step is sequentially presented in the problem description. Intuitively, presenting premises in the forward order simplifies the problem for humans, as this allows us to write the proof on-the-fly while reading the premises. Conversely, a premise ordering that is more random increases the task difficulty, since carrying out the derivation requires us to repetitively look for premises for each reasoning step. Motivated by this intuition, we categorize different premise orders based on their Kendall tau distance \(\tau\)(Cicirello, 2019; Sen, 1968) to the forward order, normalized into the range \([-1,1]\). Specifically, \(\tau=1\) is the _forward_ order, and we denote the order with \(\tau=-1\) as the _backward_ order, which is the reverse of the forward order and aligns with the proof via backward chaining. \(\tau\approx 0\) suggests that there is no strong correlation between the premise order in the problem description and the proof. To thoroughly investigate the LLM preference on different premise orders, we evaluate the model performance on \(\tau=0.5\), \(0\) and \(-0.5\), in addition to the forward (\(\tau=1\)) and backward (\(\tau=-1\)) orders. We present examples with \(\tau=1\) and \(0\) in Figure 1, and defer examples with other \(\tau\) values to Figure 11 in Appendix B.
์šฐ๋ฆฌ ๋ฒค์น˜๋งˆํฌ์˜ ์ฃผ์š” ํŠน์ง•์€ ๊ฐ ๋…ผ๋ฆฌ ์ถ”๋ก  ๋ฌธ์ œ์— ๋Œ€ํ•ด **์ „์ œ ์ˆœ์„œ๊ฐ€ ๋‹ค๋ฅธ ๋ณ€ํ˜•**์„ **ํ•ฉ์„ฑ์ ์œผ๋กœ ์ƒ์„ฑ**ํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, forward chaining์œผ๋กœ ์ ‘๊ทผํ•˜๋Š” ์ง€์ƒ ์‹ค์ฆ์— ๋ถ€ํ•ฉํ•˜๋Š” ์ˆœ์„œ๋ฅผ _forward_ ์ˆœ์„œ๋กœ ํ‘œ์‹œํ•˜๋ฉฐ, ๊ฐ ๋„์ถœ ๋‹จ๊ณ„์—์„œ ์ ์šฉ๋œ ๊ทœ์น™์ด ๋ฌธ์ œ ์„ค๋ช…์— ์ˆœ์ฐจ์ ์œผ๋กœ ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ, ์ „์ œ๋ฅผ forward ์ˆœ์„œ๋กœ ์ œ์‹œํ•˜๋ฉด ์ „์ œ๋ฅผ ์ฝ์œผ๋ฉด์„œ ์ฆ๋ช…์„ ์ฆ‰์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ๋žŒ์—๊ฒŒ๋Š” ๋ฌธ์ œ๊ฐ€ ๋” ๊ฐ„๋‹จํ•ด์ง‘๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ๋” ๋ฌด์ž‘์œ„์ ์ธ ์ „์ œ ์ˆœ์„œ๋ฅผ ์ œ์‹œํ•˜๋ฉด ๋„์ถœ์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์ถ”๋ก  ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ „์ œ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ฐพ์•„์•ผ ํ•˜๋ฏ€๋กœ ์ž‘์—… ๋‚œ์ด๋„๊ฐ€ ๋†’์•„์ง‘๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง๊ด€์— ํž˜์ž…์–ด, ์šฐ๋ฆฌ๋Š” Kendall tau ๊ฑฐ๋ฆฌ(\(\tau\))(Cicirello, 2019; Sen, 1968)๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋‹ค์–‘ํ•œ ์ „์ œ ์ˆœ์„œ๋ฅผ forward ์ˆœ์„œ๋กœ ์ •๊ทœํ™”ํ•˜์—ฌ \([-1,1]\) ๋ฒ”์œ„๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(\tau=1\)์€ _forward_ ์ˆœ์„œ์ด๋ฉฐ, \(\tau=-1\)์˜ ์ˆœ์„œ๋ฅผ forward ์ˆœ์„œ์˜ ๋ฐ˜๋Œ€์ด๋ฉฐ backward chaining์„ ํ†ตํ•ด ์ฆ๋ช…๊ณผ ์ผ์น˜ํ•˜๋Š” _backward_ ์ˆœ์„œ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. \(\tau\approx 0\)์€ ๋ฌธ์ œ ์„ค๋ช…์˜ ์ „์ œ ์ˆœ์„œ์™€ ์ฆ๋ช… ์‚ฌ์ด์— ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†์Œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ LLM ์„ ํ˜ธ๋„๋ฅผ ์ฒ ์ €ํžˆ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด forward(\(\tau=1\)) ๋ฐ backward(\(\tau=-1\)) ์ˆœ์„œ ์™ธ์—๋„ \(\tau=0.5\), \(0\) ๋ฐ \(-0.5\)์— ๋Œ€ํ•œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์— \(\tau=1\) ๋ฐ \(0\)์˜ ์˜ˆ๋ฅผ ์ œ์‹œํ•˜๊ณ , ๋‹ค๋ฅธ \(\tau\) ๊ฐ’์˜ ์˜ˆ๋Š” ๋ถ€๋ก B์˜ ๊ทธ๋ฆผ 11๋กœ ๋ฏธ๋ฃจ์–ด ๋‘ก๋‹ˆ๋‹ค.
์šฐ๋ฆฌ ๋ฒค์น˜๋งˆํฌ์˜ ์ฃผ์š” ํŠน์ง•์€ ๊ฐ ๋…ผ๋ฆฌ ์ถ”๋ก  ๋ฌธ์ œ์— ๋Œ€ํ•ด **์ „์ œ ์ˆœ์„œ๊ฐ€ ๋‹ค๋ฅธ ๋ณ€ํ˜•**์„ **ํ•ฉ์„ฑ์ ์œผ๋กœ ์ƒ์„ฑ**ํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ํฌ์›Œ๋“œ ์ฒด์ธ์œผ๋กœ ์ ‘๊ทผํ•˜๋Š” ์ง€์ƒ ์‹ค์ฆ์— ๋ถ€ํ•ฉํ•˜๋Š” ์ˆœ์„œ๋ฅผ ํฌ์›Œ๋“œ ์ˆœ์„œ๋กœ ํ‘œ์‹œํ•˜๋ฉฐ, ๊ฐ ๋„์ถœ ๋‹จ๊ณ„์—์„œ ์ ์šฉ๋œ ๊ทœ์น™์ด ๋ฌธ์ œ ์„ค๋ช…์— ์ˆœ์ฐจ์ ์œผ๋กœ ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ, ์ „์ œ๋ฅผ ์ˆœ๋ฐฉํ–ฅ ์ˆœ์„œ๋กœ ์ œ์‹œํ•˜๋ฉด ์ „์ œ๋ฅผ ์ฝ์œผ๋ฉด์„œ ์ฆ๋ช…์„ ์ฆ‰์‹œ ์ž‘์„ฑํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์‚ฌ๋žŒ์—๊ฒŒ๋Š” ๋ฌธ์ œ๊ฐ€ ๋” ๊ฐ„๋‹จํ•ด์ง‘๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ, ๋” ๋ฌด์ž‘์œ„์ ์ธ ์ „์ œ ์ˆœ์„œ๋ฅผ ์ œ์‹œํ•˜๋ฉด ๋„์ถœ์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ ์ถ”๋ก  ๋‹จ๊ณ„์— ๋Œ€ํ•œ ์ „์ œ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ฐพ์•„์•ผ ํ•˜๋ฏ€๋กœ ์ž‘์—… ๋‚œ์ด๋„๊ฐ€ ๋†’์•„์ง‘๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง๊ด€์— ํž˜์ž…์–ด, ์šฐ๋ฆฌ๋Š” ์ผ„๋‹ฌ ํƒ€์šฐ ๊ฑฐ๋ฆฌ(\tau)(Cicirello, 2019; Sen, 1968)๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋‹ค์–‘ํ•œ ์ „์ œ ์ˆœ์„œ๋ฅผ ํฌ์›Œ๋“œ ์ˆœ์„œ๋กœ ์ •๊ทœํ™”ํ•˜์—ฌ \([-1,1]\) ๋ฒ”์œ„๋กœ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, \(\tau=1\)์€ _ํฌ์›Œ๋“œ_ ์ˆœ์„œ์ด๋ฉฐ, \(\tau=-1\)์˜ ์ˆœ์„œ๋ฅผ ํฌ์›Œ๋“œ ์ˆœ์„œ์˜ ๋ฐ˜๋Œ€์ด๋ฉฐ ๋ฐฑ์›Œ๋“œ ์ฒด์ธ์„ ํ†ตํ•ด ์ฆ๋ช…๊ณผ ์ผ์น˜ํ•˜๋Š” _๋ฐฑ์›Œ๋“œ_ ์ˆœ์„œ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. \(\tau\approx 0\)์€ ๋ฌธ์ œ ์„ค๋ช…์˜ ์ „์ œ ์ˆœ์„œ์™€ ์ฆ๋ช… ์‚ฌ์ด์— ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์—†์Œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ „์ œ ์ˆœ์„œ์— ๋Œ€ํ•œ LLM ์„ ํ˜ธ๋„๋ฅผ ์ฒ ์ €ํžˆ ์กฐ์‚ฌํ•˜๊ธฐ ์œ„ํ•ด ํฌ์›Œ๋“œ(\(\tau=1\)) ๋ฐ ๋ฐฑ์›Œ๋“œ(\(\tau=-1\)) ์ˆœ์„œ ์™ธ์—๋„ \(\tau=0.5\), \(0\) ๋ฐ \(-0.5\)์— ๋Œ€ํ•œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 1์— \(\tau=1\) ๋ฐ \(0\)์˜ ์˜ˆ๋ฅผ ์ œ์‹œํ•˜๊ณ , ๋‹ค๋ฅธ \(\tau\) ๊ฐ’์˜ ์˜ˆ๋Š” ๋ถ€๋ก B์˜ ๊ทธ๋ฆผ 11๋กœ ๋ฏธ๋ฃจ์–ด ๋‘ก๋‹ˆ๋‹ค.
**Formatting Daily Chat Data.** Despite that a large number of training instances have been formatted with instructions, they mainly come from public NLP datasets, either lacking instruction diversity or mismatching with real human needs [66]. To overcome this issue, InstructGPT [66] proposes to take the queries that real users have submitted to the OpenAI API as the task descriptions. Additionally, to enrich the task diversity, human labelers are also asked to compose the instructions for real-life tasks, including open-ended generation, open question answering, brainstorming, and chatting. Then, they let another group of labelers directly answer these instructions as the output. Finally, they pair one instruction (_i.e.,_ the collected user query) and the expected output (_i.e.,_ the human-written answer) as a training instance. Note that InstructGPT also employs these real-world tasks formatted in natural language for alignment tuning (discussed in Section 5.2). Further, GPT-4 [46] has designed potentially high-risk instructions and guided the model to reject these instructions through supervised fine-tuning for safety concerns. Considering the absence of high-quality public chat data, several studies have also collected users' chat requests as input data, and then utilized ChatGPT or GPT-4 to generate responses as output data. A notable example of such a dataset is the conversational data from ShareGPT [148]. Additionally, Dolly [172] and OpenAssistant [173] have further released their conversation data, which has been carefully labeled by human annotators to attain a high level of quality.
**์ผ์ผ ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ ํฌ๋งทํŒ…**. ๋งŽ์€ ์ˆ˜์˜ ํ•™์Šต ์ธ์Šคํ„ด์Šค๊ฐ€ instructions์— ๋”ฐ๋ผ ํฌ๋งท๋˜์—ˆ์ง€๋งŒ, ์ฃผ๋กœ ๊ณต๊ฐœ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์œผ๋กœ instruction์˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์š”๊ตฌ ์‚ฌํ•ญ๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค[66]. ์ด ๋ฌธ์ œ๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด InstructGPT [66]๋Š” ์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ OpenAI API์— ์ œ์ถœํ•œ ์ฟผ๋ฆฌ๋ฅผ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์—๊ฒŒ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ, ๊ฐœ๋ฐฉํ˜• ์งˆ๋ฌธ ๋‹ต๋ณ€, ๋ธŒ๋ ˆ์ธ์Šคํ† ๋ฐ, ์ฑ„ํŒ… ๋“ฑ ์‹ค์ œ ์ž‘์—…์— ๋Œ€ํ•œ instructions์„ ์ž‘์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‹ค๋ฅธ ๋ผ๋ฒจ๋Ÿฌ ๊ทธ๋ฃน์—๊ฒŒ ์ด๋Ÿฌํ•œ instructions์— ์ง์ ‘ ์ถœ๋ ฅ์œผ๋กœ ๋‹ตํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ•˜๋‚˜์˜ instruction(_์ฆ‰, ์ˆ˜์ง‘๋œ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ)๊ณผ ์˜ˆ์ƒ ์ถœ๋ ฅ(_์ฆ‰, ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋‹ต๋ณ€)์„ ์Œ์œผ๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ํ•™์Šต ์ธ์Šคํ„ด์Šค๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. InstructGPT๋Š” alignment tuning(5.2์ ˆ์—์„œ ์„ค๋ช…)์„ ์œ„ํ•ด ์ž์—ฐ์–ด๋กœ ํฌ๋งท๋œ ์ด๋Ÿฌํ•œ ์‹ค์ œ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-4 [46]๋Š” ์ž ์žฌ์ ์œผ๋กœ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ๋Š” instructions์„ ์„ค๊ณ„ํ•˜๊ณ  ์•ˆ์ „ ๋ฌธ์ œ๋ฅผ ์œ„ํ•ด instructions์„ supervised Fine-tuning์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ instructions์„ ๊ฑฐ๋ถ€ํ•˜๋„๋ก ๋ชจ๋ธ์„ ์•ˆ๋‚ดํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ๊ณต๊ฐœ ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ์ž์˜ ์ฑ„ํŒ… ์š”์ฒญ์„ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋กœ ์ˆ˜์ง‘ํ•œ ํ›„ ChatGPT ๋˜๋Š” GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‘๋‹ต์„ ์ถœ๋ ฅ ๋ฐ์ดํ„ฐ๋กœ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” ShareGPT [148]์˜ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Dolly [172]์™€ OpenAssistant [173]๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๊ฐ€ ์‹ ์ค‘ํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
**์ผ์ผ ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ ํฌ๋งทํŒ…**. ๋งŽ์€ ์ˆ˜์˜ ํ•™์Šต ์ธ์Šคํ„ด์Šค๊ฐ€ ์ง€์นจ์— ๋”ฐ๋ผ ํฌ๋งท๋˜์—ˆ์ง€๋งŒ, ์ฃผ๋กœ ๊ณต๊ฐœ NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๊ฐ€์ ธ์˜จ ๊ฒƒ์œผ๋กœ ์ง€์นจ์˜ ๋‹ค์–‘์„ฑ์ด ๋ถ€์กฑํ•˜๊ฑฐ๋‚˜ ์‹ค์ œ ์‚ฌ์šฉ์ž์˜ ์š”๊ตฌ ์‚ฌํ•ญ๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค[66]. ์ด ๋ฌธ์ œ๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด InstructGPT [66]๋Š” ์‹ค์ œ ์‚ฌ์šฉ์ž๊ฐ€ OpenAI API์— ์ œ์ถœํ•œ ์ฟผ๋ฆฌ๋ฅผ ์ž‘์—… ์„ค๋ช…์œผ๋กœ ์‚ฌ์šฉํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž‘์—…์˜ ๋‹ค์–‘์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ๋ผ๋ฒจ๋Ÿฌ์—๊ฒŒ ๊ฐœ๋ฐฉํ˜• ์ƒ์„ฑ, ๊ฐœ๋ฐฉํ˜• ์งˆ๋ฌธ ๋‹ต๋ณ€, ๋ธŒ๋ ˆ์ธ์Šคํ† ๋ฐ, ์ฑ„ํŒ… ๋“ฑ ์‹ค์ œ ์ž‘์—…์— ๋Œ€ํ•œ ์ง€์นจ์„ ์ž‘์„ฑํ•˜๋„๋ก ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋‹ค๋ฅธ ๋ผ๋ฒจ๋Ÿฌ ๊ทธ๋ฃน์—๊ฒŒ ์ด๋Ÿฌํ•œ ์ง€์นจ์— ์ง์ ‘ ์ถœ๋ ฅ์œผ๋กœ ๋‹ตํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ํ•˜๋‚˜์˜ ์ง€์นจ(_์ฆ‰, ์ˆ˜์ง‘๋œ ์‚ฌ์šฉ์ž ์ฟผ๋ฆฌ)๊ณผ ์˜ˆ์ƒ ์ถœ๋ ฅ(_์ฆ‰, ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ๋‹ต๋ณ€)์„ ์Œ์œผ๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ ํ•™์Šต ์ธ์Šคํ„ด์Šค๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. InstructGPT๋Š” ์ •๋ ฌ ํŠœ๋‹(5.2์ ˆ์—์„œ ์„ค๋ช…)์„ ์œ„ํ•ด ์ž์—ฐ์–ด๋กœ ํฌ๋งท๋œ ์ด๋Ÿฌํ•œ ์‹ค์ œ ์ž‘์—…์„ ์‚ฌ์šฉํ•˜๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ GPT-4 [46]๋Š” ์ž ์žฌ์ ์œผ๋กœ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ๋Š” ์ง€์นจ์„ ์„ค๊ณ„ํ•˜๊ณ  ์•ˆ์ „ ๋ฌธ์ œ๋ฅผ ์œ„ํ•ด ์ง€์นจ์„ ๊ฐ๋… ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด ์ด๋Ÿฌํ•œ ์ง€์นจ์„ ๊ฑฐ๋ถ€ํ•˜๋„๋ก ๋ชจ๋ธ์„ ์•ˆ๋‚ดํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ์˜ ๊ณต๊ฐœ ์ฑ„ํŒ… ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•˜๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•˜์—ฌ, ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ์ž์˜ ์ฑ„ํŒ… ์š”์ฒญ์„ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋กœ ์ˆ˜์ง‘ํ•œ ํ›„ ChatGPT ๋˜๋Š” GPT-4๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‘๋‹ต์„ ์ถœ๋ ฅ ๋ฐ์ดํ„ฐ๋กœ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” ShareGPT [148]์˜ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ Dolly [172]์™€ OpenAssistant [173]๋Š” ๋†’์€ ์ˆ˜์ค€์˜ ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ธ๊ฐ„ ์ฃผ์„๊ฐ€๊ฐ€ ์‹ ์ค‘ํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•œ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€๋กœ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
In these works, multiple LLM agents are used to improve the performance of LLMs. For instance, LLM-Debate (Du et al., 2023) employs multiple LLM agents in a debate form. The reasoning performance is improved by creating a framework that allows more than one agent to "debate" the final answer of arithmetic tasks. They show performance improvements compared to using one single agent. Similarly, CoT-SC (Wang et al., 2023) generates multiple thought chains and picks the most self-consistent one as the final answer. The reasoning performance is improved by involving more thought chains compared to chain-of-thought (CoT) (Wei et al., 2022) which employs a single thought chain. Incidentally, from the data analysis of these works, we can notice the effects of putting multiple agents together, to some extent, can lead to a performance improvement in certain problems. For example, in Table 10 of Section 3.3 of LLM-Debate (Du et al., 2023), the authors have reported a preliminary curve: the accuracy of a math problem increases with the number of debating agents (although the number was simply increased from 1 to 7). Also, in Wang et al. (2023), involving more chains-of-thought pipelines (termed as a "sample-and-marginalize" decoding procedure), can lead to a performance gain. We realize that the LLM performance may likely be improved by a brute-force scaling up the number of agents instantiated. However, since the scaling property of "raw" agents is not the focus of these works, the scenarios/tasks and experiments considered are limited. So far, there lacks a dedicated in-depth study on this phenomenon.
์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ๋Š” LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ LLM ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLM-Debate(Du et al., 2023)๋Š” ํ† ๋ก  ํ˜•์‹์œผ๋กœ ์—ฌ๋Ÿฌ LLM ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ํ•œ ๋ช… ์ด์ƒ์˜ ์—์ด์ „ํŠธ๊ฐ€ ์‚ฐ์ˆ  ๊ณผ์ œ์˜ ์ตœ์ข… ๋‹ต์„ "ํ† ๋ก "ํ•  ์ˆ˜ ์žˆ๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด ์ถ”๋ก  ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹จ์ผ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, CoT-SC(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒ๊ฐ ์‚ฌ์Šฌ์„ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ€์žฅ ์ž๊ธฐ ์ผ๊ด€์„ฑ์ด ์žˆ๋Š” ๊ฒƒ์„ ์ตœ์ข… ๋‹ต์œผ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ๋‹จ์ผ ์‚ฌ๊ณ  ์‚ฌ์Šฌ์„ ์‚ฌ์šฉํ•˜๋Š” Chain-of-Thought (CoT)(Wei et al., 2022)์— ๋น„ํ•ด ๋” ๋งŽ์€ ์‚ฌ๊ณ  ์‚ฌ์Šฌ์„ ํฌํ•จํ•จ์œผ๋กœ์จ ์ถ”๋ก  ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์˜ ๋ฐ์ดํ„ฐ ๋ถ„์„์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ํ•จ๊ป˜ ๋ฐฐ์น˜ํ•˜๋ฉด ์–ด๋–ค ๋ฌธ์ œ์—์„œ๋Š” ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLM-Debate(Du et al., 2023)์˜ ์„น์…˜ 3.3์˜ ํ‘œ 10์—์„œ ์ €์ž๋“ค์€ ํ† ๋ก  ์—์ด์ „ํŠธ์˜ ์ˆ˜์— ๋”ฐ๋ผ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ์ •ํ™•๋„๊ฐ€ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์˜ˆ๋น„์ ์ธ ๊ณก์„ ์„ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค(์ˆ˜๋Š” ๋‹จ์ˆœํžˆ 1์—์„œ 7๋กœ ์ฆ๊ฐ€ํ–ˆ์ง€๋งŒ). ๋˜ํ•œ Wang et al. (2023)์—์„œ๋Š” ๋” ๋งŽ์€ Chain-of-Thought ํŒŒ์ดํ”„๋ผ์ธ(์ด๋ฅผ "์ƒ˜ํ”Œ ๋ฐ ํ•œ๊ณ„ํ™”" ๋””์ฝ”๋”ฉ ์ ˆ์ฐจ๋ผ๊ณ  ํ•จ)์„ ํฌํ•จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ธ์Šคํ„ด์Šคํ™”๋œ ์—์ด์ „ํŠธ์˜ ์ˆ˜๋ฅผ ๋ฌด์ฐจ๋ณ„ ๋Œ€์ž…ํ•˜์—ฌ LLM ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ž‘์—…์˜ ์ดˆ์ ์€ "์›์‹œ" ์—์ด์ „ํŠธ์˜ ํ™•์žฅ ์†์„ฑ์ด ์•„๋‹ˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ๋ ค๋œ ์‹œ๋‚˜๋ฆฌ์˜ค/์ž‘์—…๊ณผ ์‹คํ—˜์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์— ๋Œ€ํ•œ ์ „๋‹ด์ ์ธ ์‹ฌ์ธต ์—ฐ๊ตฌ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์ž‘์—…์—์„œ๋Š” LLM์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์—ฌ๋Ÿฌ LLM ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLM-Debate(Du et al., 2023)๋Š” ํ† ๋ก  ํ˜•์‹์œผ๋กœ ์—ฌ๋Ÿฌ LLM ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ํ•œ ๋ช… ์ด์ƒ์˜ ์—์ด์ „ํŠธ๊ฐ€ ์‚ฐ์ˆ  ๊ณผ์ œ์˜ ์ตœ์ข… ๋‹ต์„ "ํ† ๋ก "ํ•  ์ˆ˜ ์žˆ๋Š” ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ๋งŒ๋“ค์–ด ์ถ”๋ก  ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋“ค์€ ๋‹จ์ผ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์— ๋น„ํ•ด ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, CoT-SC(Wang et al., 2023)๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒ๊ฐ ์‚ฌ์Šฌ์„ ์ƒ์„ฑํ•˜๊ณ  ๊ฐ€์žฅ ์ž๊ธฐ ์ผ๊ด€์„ฑ์ด ์žˆ๋Š” ๊ฒƒ์„ ์ตœ์ข… ๋‹ต์œผ๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ๋‹จ์ผ ์‚ฌ๊ณ  ์‚ฌ์Šฌ์„ ์‚ฌ์šฉํ•˜๋Š” ์‚ฌ๊ณ  ์‚ฌ์Šฌ(CoT)(Wei et al., 2022)์— ๋น„ํ•ด ๋” ๋งŽ์€ ์‚ฌ๊ณ  ์‚ฌ์Šฌ์„ ํฌํ•จํ•จ์œผ๋กœ์จ ์ถ”๋ก  ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์˜ ๋ฐ์ดํ„ฐ ๋ถ„์„์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ํ•จ๊ป˜ ๋ฐฐ์น˜ํ•˜๋ฉด ์–ด๋–ค ๋ฌธ์ œ์—์„œ๋Š” ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, LLM-Debate(Du et al., 2023)์˜ ์„น์…˜ 3.3์˜ ํ‘œ 10์—์„œ ์ €์ž๋“ค์€ ํ† ๋ก  ์—์ด์ „ํŠธ์˜ ์ˆ˜์— ๋”ฐ๋ผ ์ˆ˜ํ•™ ๋ฌธ์ œ์˜ ์ •ํ™•๋„๊ฐ€ ์ฆ๊ฐ€ํ•œ๋‹ค๋Š” ์˜ˆ๋น„์ ์ธ ๊ณก์„ ์„ ๋ณด๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค(์ˆ˜๋Š” ๋‹จ์ˆœํžˆ 1์—์„œ 7๋กœ ์ฆ๊ฐ€ํ–ˆ์ง€๋งŒ). ๋˜ํ•œ Wang et al. (2023)์—์„œ๋Š” ๋” ๋งŽ์€ ์‚ฌ๊ณ ์˜ ์‚ฌ์Šฌ ํŒŒ์ดํ”„๋ผ์ธ(์ด๋ฅผ "์ƒ˜ํ”Œ ๋ฐ ํ•œ๊ณ„ํ™”" ๋””์ฝ”๋”ฉ ์ ˆ์ฐจ๋ผ๊ณ  ํ•จ)์„ ํฌํ•จํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ธ์Šคํ„ด์Šคํ™”๋œ ์—์ด์ „ํŠธ์˜ ์ˆ˜๋ฅผ ๋ฌด์ฐจ๋ณ„ ๋Œ€์ž…ํ•˜์—ฌ LLM ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ž‘์—…์˜ ์ดˆ์ ์€ "์›์‹œ" ์—์ด์ „ํŠธ์˜ ํ™•์žฅ ์†์„ฑ์ด ์•„๋‹ˆ๊ธฐ ๋•Œ๋ฌธ์— ๊ณ ๋ ค๋œ ์‹œ๋‚˜๋ฆฌ์˜ค/์ž‘์—…๊ณผ ์‹คํ—˜์ด ์ œํ•œ์ ์ž…๋‹ˆ๋‹ค. ์ง€๊ธˆ๊นŒ์ง€ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์— ๋Œ€ํ•œ ์ „๋‹ด์ ์ธ ์‹ฌ์ธต ์—ฐ๊ตฌ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค.
**Pros and Cons of SFT**. SFT has been shown to be an effective approach to boosting the performance of LLMs on various benchmarks [67, 69, 137, 138], which can largely enhance the task generalization ability and flexibly endow specific functions (_e.g.,_ establishing the chatbot's identity). More discussions about the usefulness of SFT can be found in Section 5.1.3. It has been widely recognized that SFT mainly _unlocks_ the abilities but not _inject_ new abilities into LLMs. Thus, it might become problematic when one tries to stimulate the non-endogenous abilities of LLMs via SFT. As a concrete scenario, it would potentially advocate the hallucination behaviors when demonstration data is beyond the knowledge or ability scope of LLMs, _e.g.,_ training a LLM to answer questions about its unknown facts. An interesting viewpoint from John Schulman's talk on RLHF [395] is that distilling superior models to train less capable models (_e.g.,_ prompting GPT-4 to generate the response as fine-tuning data) might increase the possibilities of generating the hallucinated texts, thus likely affecting the factual accuracy of LLMs. Furthermore, as a behavior cloning method, SFT aims to imitate the behaviors (without explorations) of the experts who construct the demonstration data. However, there often exist variations among different annotators on the writing styles, quality, and preferences of demonstration data, which tends to affect the learning performance of SFT. Thus, high-quality instruction data (but not the quantity) is the primary factor for effective training of LLMs during the SFT stage [99].
**SFT์˜ ์žฅ๋‹จ์ **. Supervised Fine-tuning (SFT)๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ Large Language Models (LLMs)์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ[67, 69, 137, 138], ์ด๋ฅผ ํ†ตํ•ด ์ž‘์—… ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ํŠน์ • ๊ธฐ๋Šฅ(์˜ˆ: ์ฑ—๋ด‡์˜ ์‹ ์› ์„ค์ •)์„ ์œ ์—ฐํ•˜๊ฒŒ ๋ถ€์—ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SFT์˜ ์œ ์šฉ์„ฑ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋…ผ์˜๋Š” ์„น์…˜ 5.1.3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SFT๋Š” ์ฃผ๋กœ ๋Šฅ๋ ฅ์„ '์ž ๊ธˆ ํ•ด์ œ'ํ•˜๋Š” ๊ฒƒ์ด์ง€ LLM์— ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ '์ฃผ์ž…'ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์ด ๋„๋ฆฌ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ SLM์ด ์•Œ๋ ค์ง€์ง€ ์•Š์€ ์‚ฌ์‹ค์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์ด, SFT๋ฅผ ํ†ตํ•ด LLM์˜ ๋น„๋‚ด์ƒ์  ๋Šฅ๋ ฅ์„ ์ž๊ทนํ•˜๋ ค๊ณ  ํ•  ๋•Œ ๋ฌธ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋กœ, ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ๊ฐ€ LLM์˜ ์ง€์‹์ด๋‚˜ ๋Šฅ๋ ฅ ๋ฒ”์œ„๋ฅผ ๋„˜์–ด์„œ๋Š” ๊ฒฝ์šฐ, ์ฆ‰, Hallucination ํ–‰๋™์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กด ์Š๋งŒ์˜ Reinforcement Learning from Human Feedback (RLHF)์— ๋Œ€ํ•œ ๊ฐ•์—ฐ์—์„œ ํฅ๋ฏธ๋กœ์šด ๊ด€์ ์€ ๋œ ์œ ๋Šฅํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ์„ ์ฆ๋ฅ˜ํ•˜๋Š” ๊ฒƒ(์˜ˆ: GPT-4๊ฐ€ Fine-tuning ๋ฐ์ดํ„ฐ๋กœ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ)์ด ํ™˜๊ฐ์ ์ธ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์—ฌ LLM์˜ ์‚ฌ์‹ค ์ •ํ™•์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–‰๋™ ๋ณต์ œ ๋ฐฉ๋ฒ•์œผ๋กœ์„œ SFT๋Š” ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ํ–‰๋™(ํƒ์ƒ‰ ์—†์ด)์„ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ์˜ ์ž‘์„ฑ ์Šคํƒ€์ผ, ํ’ˆ์งˆ ๋ฐ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด์„œ๋Š” ์ฃผ๊ด€์ž๋งˆ๋‹ค ์ฐจ์ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์ด๋Š” SFT์˜ ํ•™์Šต ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ ํ’ˆ์งˆ์˜ Instruction ๋ฐ์ดํ„ฐ(์ˆ˜๋Ÿ‰์ด ์•„๋‹Œ)๋Š” SFT ๋‹จ๊ณ„์—์„œ LLM์˜ ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ์„ ์œ„ํ•œ ์ฃผ์š” ์š”์†Œ์ž…๋‹ˆ๋‹ค[99].
**SFT์˜ ์žฅ๋‹จ์ **. SFT๋Š” ๋‹ค์–‘ํ•œ ๋ฒค์น˜๋งˆํฌ์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ[67, 69, 137, 138], ์ด๋ฅผ ํ†ตํ•ด ์ž‘์—… ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ํŠน์ • ๊ธฐ๋Šฅ(์˜ˆ: ์ฑ—๋ด‡์˜ ์‹ ์› ์„ค์ •)์„ ์œ ์—ฐํ•˜๊ฒŒ ๋ถ€์—ฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SFT์˜ ์œ ์šฉ์„ฑ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋…ผ์˜๋Š” ์„น์…˜ 5.1.3์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SFT๋Š” ์ฃผ๋กœ ๋Šฅ๋ ฅ์„ '์ž ๊ธˆ ํ•ด์ œ'ํ•˜๋Š” ๊ฒƒ์ด์ง€ LLM์— ์ƒˆ๋กœ์šด ๋Šฅ๋ ฅ์„ '์ฃผ์ž…'ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ๋Š” ์ ์ด ๋„๋ฆฌ ์•Œ๋ ค์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ SLM์ด ์•Œ๋ ค์ง€์ง€ ์•Š์€ ์‚ฌ์‹ค์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋„๋ก LLM์„ ํ›ˆ๋ จํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์ด, SFT๋ฅผ ํ†ตํ•ด LLM์˜ ๋น„๋‚ด์ƒ์  ๋Šฅ๋ ฅ์„ ์ž๊ทนํ•˜๋ ค๊ณ  ํ•  ๋•Œ ๋ฌธ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์ธ ์‹œ๋‚˜๋ฆฌ์˜ค๋กœ, ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ๊ฐ€ LLM์˜ ์ง€์‹์ด๋‚˜ ๋Šฅ๋ ฅ ๋ฒ”์œ„๋ฅผ ๋„˜์–ด์„œ๋Š” ๊ฒฝ์šฐ, ์ฆ‰, ํ™€๋ฃจ์‹œ๋„ค์ด์…˜ ํ–‰๋™์„ ์ด‰์ง„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กด ์Š๋งŒ์˜ RLHF์— ๋Œ€ํ•œ ๊ฐ•์—ฐ์—์„œ ํฅ๋ฏธ๋กœ์šด ๊ด€์ ์€ ๋œ ์œ ๋Šฅํ•œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ธฐ ์œ„ํ•ด ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ์„ ์ฆ๋ฅ˜ํ•˜๋Š” ๊ฒƒ(์˜ˆ: GPT-4๊ฐ€ ๋ฏธ์„ธ ์กฐ์ • ๋ฐ์ดํ„ฐ๋กœ ์‘๋‹ต์„ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ)์ด ํ™˜๊ฐ์ ์ธ ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•  ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์—ฌ LLM์˜ ์‚ฌ์‹ค ์ •ํ™•์„ฑ์— ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ํ–‰๋™ ๋ณต์ œ ๋ฐฉ๋ฒ•์œผ๋กœ์„œ SFT๋Š” ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ์ „๋ฌธ๊ฐ€์˜ ํ–‰๋™(ํƒ์ƒ‰ ์—†์ด)์„ ๋ชจ๋ฐฉํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ฐ๋ชจ ๋ฐ์ดํ„ฐ์˜ ์ž‘์„ฑ ์Šคํƒ€์ผ, ํ’ˆ์งˆ ๋ฐ ์„ ํ˜ธ๋„์— ๋Œ€ํ•ด์„œ๋Š” ์ฃผ๊ด€์ž๋งˆ๋‹ค ์ฐจ์ด๊ฐ€ ์žˆ๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์ด๋Š” SFT์˜ ํ•™์Šต ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ ํ’ˆ์งˆ์˜ ์ง€์นจ ๋ฐ์ดํ„ฐ(์ˆ˜๋Ÿ‰์ด ์•„๋‹Œ)๋Š” SFT ๋‹จ๊ณ„์—์„œ LLM์˜ ํšจ๊ณผ์ ์ธ ํ›ˆ๋ จ์„ ์œ„ํ•œ ์ฃผ์š” ์š”์†Œ์ž…๋‹ˆ๋‹ค[99].
In the post-deep learning era, the Transformer architecture has demonstrated its powerful performance across pre-trained big models and various downstream tasks. However, the enormous computational demands of this architecture have deterred many researchers. To further reduce the complexity of attention models, numerous efforts have been made to design more efficient methods. Among them, the State Space Model (SSM), as a possible replacement for the self-attention based Transformer model, has drawn more and more attention in recent years. In this paper, we give the first comprehensive review of these works and also provide experimental comparisons and analysis to better demonstrate the features and advantages of SSM. Specifically, we first give a detailed description of principles to help the readers quickly capture the key ideas of SSM. After that, we dive into the reviews of existing SSMs and their various applications, including natural language processing, computer vision, graph, multi-modal and multi-media, point cloud/event stream, time series data, and other domains. In addition, we give statistical comparisons and analysis of these models and hope it helps the readers to understand the effectiveness of different structures on various tasks. Then, we propose possible research points in this direction to better promote the development of the theoretical model and application of SSM. More related works will be continuously updated on the following GitHub [https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List](https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List).
๋”ฅ๋Ÿฌ๋‹ ์ดํ›„ Transformer ์•„ํ‚คํ…์ฒ˜๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋Œ€ํ˜• ๋ชจ๋ธ๊ณผ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ์ „๋ฐ˜์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์•„ํ‚คํ…์ฒ˜์˜ ๋ง‰๋Œ€ํ•œ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์œผ๋กœ ์ธํ•ด ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋ง‰ํžˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ attention ๋ชจ๋ธ์˜ ๋ณต์žก์„ฑ์„ ๋”์šฑ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋” ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ˆ˜๋งŽ์€ ๋…ธ๋ ฅ์ด ๊ณ„์†๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ์ค‘์—์„œ๋„ self-attention ๊ธฐ๋ฐ˜ Transformer ๋ชจ๋ธ์˜ ๋Œ€์ฒด ๊ฐ€๋Šฅ์„ฑ์œผ๋กœ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ๋Š” State Space Model (SSM)์€ ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ์ ์  ๋” ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ์ข…ํ•ฉ์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์‹คํ—˜์  ๋น„๊ต์™€ ๋ถ„์„์„ ํ†ตํ•ด SSM์˜ ํŠน์ง•๊ณผ ์žฅ์ ์„ ๋” ์ž˜ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ๋„๋ก ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๋จผ์ € ๋…์ž๊ฐ€ SSM์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋ฅผ ๋น ๋ฅด๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ์›๋ฆฌ๋ฅผ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „, ๊ทธ๋ž˜ํ”„, ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด, ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ/์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ๋ฐ ๊ธฐํƒ€ ์˜์—ญ์„ ํฌํ•จํ•œ ๊ธฐ์กด SSM๊ณผ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ์— ๋Œ€ํ•œ ๊ฒ€ํ† ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ†ต๊ณ„์  ๋น„๊ต ๋ฐ ๋ถ„์„์„ ์ œ๊ณตํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๊ตฌ์กฐ์˜ ํšจ๊ณผ๋ฅผ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋ก ์  ๋ชจ๋ธ์˜ ๋ฐœ์ „๊ณผ SSM์˜ ์ ์šฉ์„ ๋” ์ž˜ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ ํฌ์ธํŠธ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ ์ž‘์—…์€ ๋‹ค์Œ GitHub [https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List](https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List)์—์„œ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
๋”ฅ๋Ÿฌ๋‹ ์ดํ›„ ํŠธ๋žœ์Šคํฌ๋จธ ์•„ํ‚คํ…์ฒ˜๋Š” ์‚ฌ์ „ ํ•™์Šต๋œ ๋Œ€ํ˜• ๋ชจ๋ธ๊ณผ ๋‹ค์–‘ํ•œ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—… ์ „๋ฐ˜์—์„œ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ์•„ํ‚คํ…์ฒ˜์˜ ๋ง‰๋Œ€ํ•œ ๊ณ„์‚ฐ ์š”๊ตฌ ์‚ฌํ•ญ์œผ๋กœ ์ธํ•ด ๋งŽ์€ ์—ฐ๊ตฌ์ž๋“ค์ด ๋ง‰ํžˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์‹ฌ ๋ชจ๋ธ์˜ ๋ณต์žก์„ฑ์„ ๋”์šฑ ์ค„์ด๊ธฐ ์œ„ํ•ด ๋” ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์„ ์„ค๊ณ„ํ•˜๊ธฐ ์œ„ํ•œ ์ˆ˜๋งŽ์€ ๋…ธ๋ ฅ์ด ๊ณ„์†๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ์ค‘์—์„œ๋„ ์ž๊ธฐ ์ฃผ์˜ ๊ธฐ๋ฐ˜ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์˜ ๋Œ€์ฒด ๊ฐ€๋Šฅ์„ฑ์œผ๋กœ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ๋Š” ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(SSM)์€ ์ตœ๊ทผ ๋ช‡ ๋…„ ๋™์•ˆ ์ ์  ๋” ๋งŽ์€ ๊ด€์‹ฌ์„ ๋ฐ›๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ์ตœ์ดˆ์˜ ์ข…ํ•ฉ์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ œ๊ณตํ•˜๊ณ , ์‹คํ—˜์  ๋น„๊ต์™€ ๋ถ„์„์„ ํ†ตํ•ด SSM์˜ ํŠน์ง•๊ณผ ์žฅ์ ์„ ๋” ์ž˜ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ๋„๋ก ๋…ธ๋ ฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, ๋จผ์ € ๋…์ž๊ฐ€ SSM์˜ ํ•ต์‹ฌ ์•„์ด๋””์–ด๋ฅผ ๋น ๋ฅด๊ฒŒ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ์›๋ฆฌ๋ฅผ ์ž์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ, ์ปดํ“จํ„ฐ ๋น„์ „, ๊ทธ๋ž˜ํ”„, ๋‹ค์ค‘ ๋ชจ๋‹ฌ ๋ฐ ๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด, ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ/์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ๋ฐ ๊ธฐํƒ€ ์˜์—ญ์„ ํฌํ•จํ•œ ๊ธฐ์กด SSM๊ณผ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ์— ๋Œ€ํ•œ ๊ฒ€ํ† ๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ํ†ต๊ณ„์  ๋น„๊ต ๋ฐ ๋ถ„์„์„ ์ œ๊ณตํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ ๋‹ค์–‘ํ•œ ๊ตฌ์กฐ์˜ ํšจ๊ณผ๋ฅผ ์ดํ•ดํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ๋ฐฉํ–ฅ์œผ๋กœ ์ด๋ก ์  ๋ชจ๋ธ์˜ ๋ฐœ์ „๊ณผ SSM์˜ ์ ์šฉ์„ ๋” ์ž˜ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐ€๋Šฅํ•œ ์—ฐ๊ตฌ ํฌ์ธํŠธ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๊ด€๋ จ ์ž‘์—…์€ ๋‹ค์Œ GitHub [https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List](https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List)์—์„œ ์ง€์†์ ์œผ๋กœ ์—…๋ฐ์ดํŠธ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.
**Stage 1 (new input embeddings):** Initially, our focus is narrow yet critical: to learn the input embeddings of the newly added tokens while freezing all other model parameters. This stage is foundational, allowing the model to adjust its recognition and processing of these tokens from the beginning. The pre-initialized embeddings serve as a starting point, guiding the model to better utilize these new tokens in its existing framework. Our principal hypothesis here is that if the input and output token sequences in causal language modeling can be differentiated, by utilizing both the old and new tokenizers at the same time, the model can more efficiently and effectively learn new vocabulary embeddings, as it could leverage its established knowledge in the embedding spaces from old tokens. However, employing distinct tokenizers for input and output sequences at once poses implementation challenges, such as the difficulty of applying teacher forcing due to mismatched input/output sequences. Here, the subword-based embedding initialization (Sec 2.2) provides a proxy for using the old tokenizer for output sequences, such that the model is tasked to generate the subword token (old) given the whole word token (new). In other words, the model could learn to align their representations for generating the new token with that for generating its first subword token, by optimizing only the input embeddings without any modification of input/output token sequences as described in Figure 2. However, at this stage, the model is not yet able to distinguish between tokens sharing the same hidden state.
**1๋‹จ๊ณ„(์ƒˆ input embeddings):** ์ฒ˜์Œ์—๋Š” ๋‹ค๋ฅธ ๋ชจ๋“  ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋™๊ฒฐํ•œ ์ƒํƒœ์—์„œ ์ƒˆ๋กœ ์ถ”๊ฐ€๋œ ํ† ํฐ์˜ input embeddings์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถฅ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ชจ๋ธ์ด ์ฒ˜์Œ๋ถ€ํ„ฐ ์ด๋Ÿฌํ•œ ํ† ํฐ์— ๋Œ€ํ•œ ์ธ์‹๊ณผ ์ฒ˜๋ฆฌ๋ฅผ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ธฐ์ดˆ์ ์ธ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ์ดˆ๊ธฐํ™”๋œ embeddings์€ ์‹œ์ž‘์  ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋ชจ๋ธ์ด ๊ธฐ์กด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ํ† ํฐ์„ ๋” ์ž˜ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ์˜ ์ฃผ์š” ๊ฐ€์„ค์€ ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ input ๋ฐ output token sequences๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ๊ธฐ์กด ํ† ํฐ๊ณผ ์ƒˆ๋กœ์šด tokenizer๋ฅผ ๋™์‹œ์— ํ™œ์šฉํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์ด embedding ๊ณต๊ฐ„์—์„œ ๊ธฐ์กด ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ƒˆ๋กœ์šด ์–ดํœ˜ embeddings์„ ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ input ๋ฐ output sequences์— ์„œ๋กœ ๋‹ค๋ฅธ tokenizer๋ฅผ ํ•œ ๋ฒˆ์— ์‚ฌ์šฉํ•˜๋ฉด input/output sequences ๋ถˆ์ผ์น˜๋กœ ์ธํ•œ teacher forcing ์ ์šฉ์˜ ์–ด๋ ค์›€ ๋“ฑ ๊ตฌํ˜„์ƒ์˜ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ subword ๊ธฐ๋ฐ˜ embedding ์ดˆ๊ธฐํ™”(Sec 2.2)๋Š” output sequences์— ๋Œ€ํ•ด ์ด์ „ tokenizer๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ํ”„๋ก์‹œ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๋ชจ๋ธ์ด ์ „์ฒด ๋‹จ์–ด token(์ƒˆ๋กœ์šด)์ด ์ฃผ์–ด์ง€๋ฉด subword token(์ด์ „)์„ ์ƒ์„ฑํ•˜๋„๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ์€ ๊ทธ๋ฆผ 2์— ์„ค๋ช…๋œ ๋Œ€๋กœ input/output token sequences๋ฅผ ์ˆ˜์ •ํ•˜์ง€ ์•Š๊ณ  input embeddings๋งŒ ์ตœ์ ํ™”ํ•˜์—ฌ ์ƒˆ๋กœ์šด token์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ‘œํ˜„๊ณผ ์ฒซ ๋ฒˆ์งธ subword token์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ๊ณต์œ ํ•˜๋Š” token์„ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
**1๋‹จ๊ณ„(์ƒˆ ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ):** ์ฒ˜์Œ์—๋Š” ๋‹ค๋ฅธ ๋ชจ๋“  ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋™๊ฒฐํ•œ ์ƒํƒœ์—์„œ ์ƒˆ๋กœ ์ถ”๊ฐ€๋œ ํ† ํฐ์˜ ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ์„ ํ•™์Šตํ•˜๋Š” ๋ฐ ์ดˆ์ ์„ ๋งž์ถฅ๋‹ˆ๋‹ค. ์ด ๋‹จ๊ณ„๋Š” ๋ชจ๋ธ์ด ์ฒ˜์Œ๋ถ€ํ„ฐ ์ด๋Ÿฌํ•œ ํ† ํฐ์— ๋Œ€ํ•œ ์ธ์‹๊ณผ ์ฒ˜๋ฆฌ๋ฅผ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ธฐ์ดˆ์ ์ธ ๋‹จ๊ณ„์ž…๋‹ˆ๋‹ค. ์‚ฌ์ „ ์ดˆ๊ธฐํ™”๋œ ์ž„๋ฒ ๋”ฉ์€ ์‹œ์ž‘์  ์—ญํ• ์„ ํ•˜๋ฉฐ, ๋ชจ๋ธ์ด ๊ธฐ์กด ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ ์ด๋Ÿฌํ•œ ์ƒˆ๋กœ์šด ํ† ํฐ์„ ๋” ์ž˜ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์šฐ๋ฆฌ์˜ ์ฃผ์š” ๊ฐ€์„ค์€ ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง์—์„œ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ํ† ํฐ ์‹œํ€€์Šค๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ๋‹ค๋ฉด, ๊ธฐ์กด ํ† ํฐ๊ณผ ์ƒˆ๋กœ์šด ํ† ํฐํ™”๊ธฐ๋ฅผ ๋™์‹œ์— ํ™œ์šฉํ•จ์œผ๋กœ์จ ๋ชจ๋ธ์ด ์ž„๋ฒ ๋”ฉ ๊ณต๊ฐ„์—์„œ ๊ธฐ์กด ์ง€์‹์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์ƒˆ๋กœ์šด ์–ดํœ˜ ์ž„๋ฒ ๋”ฉ์„ ๋ณด๋‹ค ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ž…๋ ฅ ๋ฐ ์ถœ๋ ฅ ์‹œํ€€์Šค์— ์„œ๋กœ ๋‹ค๋ฅธ ํ† ํฐํ™”๊ธฐ๋ฅผ ํ•œ ๋ฒˆ์— ์‚ฌ์šฉํ•˜๋ฉด ์ž…๋ ฅ/์ถœ๋ ฅ ์‹œํ€€์Šค ๋ถˆ์ผ์น˜๋กœ ์ธํ•œ ๊ต์‚ฌ ๊ฐ•์ œ ์ ์šฉ์˜ ์–ด๋ ค์›€ ๋“ฑ ๊ตฌํ˜„์ƒ์˜ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ํ•˜์œ„ ๋‹จ์–ด ๊ธฐ๋ฐ˜ ์ž„๋ฒ ๋”ฉ ์ดˆ๊ธฐํ™”(Sec 2.2)๋Š” ์ถœ๋ ฅ ์‹œํ€€์Šค์— ๋Œ€ํ•ด ์ด์ „ ํ† ํฐํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ํ”„๋ก์‹œ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๋ชจ๋ธ์ด ์ „์ฒด ๋‹จ์–ด ํ† ํฐ(์ƒˆ๋กœ์šด)์ด ์ฃผ์–ด์ง€๋ฉด ํ•˜์œ„ ๋‹จ์–ด ํ† ํฐ(์ด์ „)์„ ์ƒ์„ฑํ•˜๋„๋ก ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋ชจ๋ธ์€ ๊ทธ๋ฆผ 2์— ์„ค๋ช…๋œ ๋Œ€๋กœ ์ž…๋ ฅ/์ถœ๋ ฅ ํ† ํฐ ์‹œํ€€์Šค๋ฅผ ์ˆ˜์ •ํ•˜์ง€ ์•Š๊ณ  ์ž…๋ ฅ ์ž„๋ฒ ๋”ฉ๋งŒ ์ตœ์ ํ™”ํ•˜์—ฌ ์ƒˆ๋กœ์šด ํ† ํฐ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ‘œํ˜„๊ณผ ์ฒซ ๋ฒˆ์งธ ํ•˜์œ„ ๋‹จ์–ด ํ† ํฐ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์ด ๋™์ผํ•œ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ๊ณต์œ ํ•˜๋Š” ํ† ํฐ์„ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.
In recent times, substantial advancements have been witnessed in large language models (LLMs), exemplified by ChatGPT, showcasing remarkable proficiency across a range of complex tasks. However, many mainstream LLMs (e.g. LLaMA) are pretrained on English-dominant corpus, which limits their performance in other non-English languages. In this paper, we focus on how to effectively transfer the capabilities of language generation and following instructions to a non-English language. To answer this question, we conduct an extensive empirical investigation based on LLaMA, accumulating over 1440 GPU hours. We analyze the impact of key factors such as vocabulary extension, further pretraining, and instruction tuning on transfer. To accurately assess the model's level of knowledge, we employ four widely used standardized testing benchmarks: C-Eval, MMLU, AGI-Eval, and GAOKAO-Bench. Furthermore, a comprehensive evaluation of the model's response quality is conducted, considering aspects such as accuracy, fluency, informativeness, logical coherence, and harmlessness, based on LLM-Eval, a benchmarks consisting instruction tasks from 17 diverse categories. Our evaluation results demonstrate that comparable performance to state-of-the-art transfer models can be achieved with less than \(1\%\) of the pretraining data, both in terms of knowledge alignment and response quality. Furthermore, the experimental outcomes across the thirteen low-resource languages also exhibit similar trends. We anticipate that the conclusions revealed by the experiments will aid the community in developing non-English LLMs.
์ตœ๊ทผ์—๋Š” ChatGPT๋ฅผ ๋Œ€ํ‘œ๋กœ ํ•˜๋Š” Large Language Models (LLMs)์ด ํฌ๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋งŽ์€ ์ฃผ๋ฅ˜ LLMs (์˜ˆ: LLaMA)์€ ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ์Šค๋กœ ์‚ฌ์ „ ํ•™์Šต๋˜์–ด ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์—์„œ ์„ฑ๋Šฅ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์–ธ์–ด ์ƒ์„ฑ ๋ฐ Instruction ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํšจ๊ณผ์ ์œผ๋กœ ์ „์†กํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1440์‹œ๊ฐ„ ์ด์ƒ์˜ GPU ์‹œ๊ฐ„์„ ์ถ•์ ํ•˜์—ฌ ๊ด‘๋ฒ”์œ„ํ•œ ์‹ค์ฆ ์กฐ์‚ฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ดํœ˜ ํ™•์žฅ, ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต, Instruction Tuning๊ณผ ๊ฐ™์€ ์ฃผ์š” ์š”์†Œ๊ฐ€ ์ „์†ก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์ง€์‹ ์ˆ˜์ค€์„ ์ •ํ™•ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” 4๊ฐ€์ง€ ํ‘œ์ค€ํ™” ํ…Œ์ŠคํŠธ ๋ฒค์น˜๋งˆํฌ์ธ C-Eval, MMLU, AGI-Eval, GAOKAO-Bench๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 17๊ฐœ์˜ ๋‹ค์–‘ํ•œ ์นดํ…Œ๊ณ ๋ฆฌ์˜ Instruction ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ์ธ LLM-Eval์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•์„ฑ, ์œ ์ฐฝ์„ฑ, ์œ ์ต์„ฑ, ๋…ผ๋ฆฌ์  ์ผ๊ด€์„ฑ, ๋ฌดํ•ด์„ฑ ๋“ฑ์˜ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ ๋ชจ๋ธ์˜ ์‘๋‹ต ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ์ง€์‹ ์ •๋ ฌ๊ณผ ์‘๋‹ต ํ’ˆ์งˆ ์ธก๋ฉด์—์„œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ 1% ๋ฏธ๋งŒ์œผ๋กœ๋„ ์ตœ์‹  ์ „์†ก ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 13๊ฐœ์˜ ์ €์ž์› ์–ธ์–ด์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋„ ์œ ์‚ฌํ•œ ์ถ”์„ธ๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์‹คํ—˜์„ ํ†ตํ•ด ๋ฐํ˜€์ง„ ๊ฒฐ๋ก ์ด ๋น„์˜์–ด LLMs์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
์ตœ๊ทผ์—๋Š” ChatGPT๋ฅผ ๋Œ€ํ‘œ๋กœ ํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ํฌ๊ฒŒ ๋ฐœ์ „ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ณต์žกํ•œ ์ž‘์—…์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ณด์—ฌ์ฃผ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋งŽ์€ ์ฃผ๋ฅ˜ LLM(์˜ˆ: LLaMA)์€ ์˜์–ด๊ฐ€ ์ง€๋ฐฐ์ ์ธ ์ฝ”ํผ์Šค๋กœ ์‚ฌ์ „ ํ•™์Šต๋˜์–ด ๋‹ค๋ฅธ ๋น„์˜์–ด ์–ธ์–ด์—์„œ ์„ฑ๋Šฅ์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ์–ธ์–ด ์ƒ์„ฑ ๋ฐ ๋ช…๋ น์–ด ์ˆ˜ํ–‰ ๊ธฐ๋Šฅ์„ ๋น„์˜์–ด ์–ธ์–ด๋กœ ํšจ๊ณผ์ ์œผ๋กœ ์ „์†กํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด LLaMA๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 1440์‹œ๊ฐ„ ์ด์ƒ์˜ GPU ์‹œ๊ฐ„์„ ์ถ•์ ํ•˜์—ฌ ๊ด‘๋ฒ”์œ„ํ•œ ์‹ค์ฆ ์กฐ์‚ฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ดํœ˜ ํ™•์žฅ, ์ถ”๊ฐ€ ์‚ฌ์ „ ํ•™์Šต, ๋ช…๋ น์–ด ํŠœ๋‹๊ณผ ๊ฐ™์€ ์ฃผ์š” ์š”์†Œ๊ฐ€ ์ „์†ก์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์˜ ์ง€์‹ ์ˆ˜์ค€์„ ์ •ํ™•ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” 4๊ฐ€์ง€ ํ‘œ์ค€ํ™” ํ…Œ์ŠคํŠธ ๋ฒค์น˜๋งˆํฌ์ธ C-Eval, MMLU, AGI-Eval, GAOKAO-Bench๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 17๊ฐœ์˜ ๋‹ค์–‘ํ•œ ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๋ช…๋ น์–ด ๊ณผ์ œ๋กœ ๊ตฌ์„ฑ๋œ ๋ฒค์น˜๋งˆํฌ์ธ LLM-Eval์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •ํ™•์„ฑ, ์œ ์ฐฝ์„ฑ, ์œ ์ต์„ฑ, ๋…ผ๋ฆฌ์  ์ผ๊ด€์„ฑ, ๋ฌดํ•ด์„ฑ ๋“ฑ์˜ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ ๋ชจ๋ธ์˜ ์‘๋‹ต ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ข…ํ•ฉ์ ์ธ ํ‰๊ฐ€๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ‰๊ฐ€ ๊ฒฐ๊ณผ, ์ง€์‹ ์ •๋ ฌ๊ณผ ์‘๋‹ต ํ’ˆ์งˆ ์ธก๋ฉด์—์„œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ 1% ๋ฏธ๋งŒ์œผ๋กœ๋„ ์ตœ์‹  ์ „์†ก ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ 13๊ฐœ์˜ ์ €์ž์› ์–ธ์–ด์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋„ ์œ ์‚ฌํ•œ ์ถ”์„ธ๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์‹คํ—˜์„ ํ†ตํ•ด ๋ฐํ˜€์ง„ ๊ฒฐ๋ก ์ด ๋น„์˜์–ด LLM์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๊ธฐ๋Œ€ํ•ฉ๋‹ˆ๋‹ค.
Immediate hazardsImmediate hazards are sources of harm that are already being presented by existing frontier and production-ready models. This includes enabling scams and fraud [86], terrorist activity [87, 88], disinformation campaigns [89, 90, 91], creation of child sexual abuse material[92], encouraging suicide and self-harm [93], cyber attacks and malware[94, 95], amongst many others [96]. Another concern is factual errors and "hallucinations". This is a substantial risk when models are faced with questions about events that happened after their training cutoff date if they do not have access to external sources of up-to-date information [6, 97, 98]. Generative AI has been shown to increase the scale and severity of these hazards by reducing organizational and material barriers. For instance, the media has reported that criminals have used text-to-speech models to run realistic banking scans where they mass-call people and pretend to be one of their relations in need of immediate financial assistance [99]. The risk of bias, unfairness, and discrimination in AI models is a longstanding concern, supported by a large body of research [100, 101, 44, 102]. Recent work also shows that out-the-box models can also be easily adjusted with a small fine-tuning budget to readily generate toxic, hateful, offensive, and deeply biased content [103, 104, 68]. And substantial work has focused on developing human- and machine-understandable attack methods to cause models to regurgitate private information [105], 'forget' their safety filters [106] or reveal vulnerabilities in their design [107].
์ฆ‰๊ฐ์ ์ธ ์œ„ํ—˜์ฆ‰๊ฐ์ ์ธ ์œ„ํ—˜์€ ๊ธฐ์กด์˜ ํ”„๋ก ํ‹ฐ์–ด ๋ฐ ์ƒ์‚ฐ ์ค€๋น„ ๋ชจ๋ธ์—์„œ ์ด๋ฏธ ๋ฐœ์ƒํ•˜๊ณ  ์žˆ๋Š” ์œ„ํ•ด ์š”์†Œ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์‚ฌ๊ธฐ ๋ฐ ์‚ฌ๊ธฐ ๊ฐ€๋Šฅ[86], ํ…Œ๋Ÿฌ ํ™œ๋™[87, 88], ํ—ˆ์œ„ ์ •๋ณด ์บ ํŽ˜์ธ[89, 90, 91], ์•„๋™ ์„ฑ ํ•™๋Œ€ ์ž๋ฃŒ ์ œ์ž‘[92], ์ž์‚ด ๋ฐ ์žํ•ด ์กฐ์žฅ[93], ์‚ฌ์ด๋ฒ„ ๊ณต๊ฒฉ ๋ฐ ๋ฉ€์›จ์–ด[94, 95] ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์šฐ๋ ค๋Š” ์‚ฌ์‹ค ์˜ค๋ฅ˜์™€ "hallucinations"์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ตœ์‹  ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์ด ๊ต์œก ์ข…๋ฃŒ์ผ ์ดํ›„์— ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ์ƒ๋‹นํ•œ ์œ„ํ—˜์ด ๋ฉ๋‹ˆ๋‹ค[6, 97, 98]. ์ƒ์„ฑํ˜• AI๋Š” ์กฐ์ง์  ๋ฐ ๋ฌผ์งˆ์  ์žฅ๋ฒฝ์„ ๋‚ฎ์ถค์œผ๋กœ์จ ์ด๋Ÿฌํ•œ ์œ„ํ—˜์˜ ๊ทœ๋ชจ์™€ ์‹ฌ๊ฐ์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์–ธ๋ก ์—์„œ๋Š” ๋ฒ”์ฃ„์ž๋“ค์ด ํ…์ŠคํŠธ ์Œ์„ฑ ๋ณ€ํ™˜ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋Œ€๋Ÿ‰์œผ๋กœ ์ „ํ™”๋ฅผ ๊ฑธ์–ด ์ฆ‰๊ฐ์ ์ธ ๊ธˆ์ „์  ์ง€์›์ด ํ•„์š”ํ•œ ์นœ์ฒ™ ์ค‘ ํ•œ ๋ช…์ธ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฐ€์žฅํ•˜๋Š” ์‚ฌ์‹ค์ ์ธ ๋ฑ…ํ‚น ์Šค์บ”์„ ์‹คํ–‰ํ–ˆ๋‹ค๊ณ  ๋ณด๋„ํ–ˆ์Šต๋‹ˆ๋‹ค[99]. AI ๋ชจ๋ธ์˜ ํŽธํ–ฅ์„ฑ, ๋ถˆ๊ณต์ •์„ฑ, ์ฐจ๋ณ„์„ฑ์— ๋Œ€ํ•œ ์œ„ํ—˜์€ ์˜ค๋žœ ๊ธฐ๊ฐ„ ๋™์•ˆ ์šฐ๋ ค๋˜์–ด ์™”์œผ๋ฉฐ, ์ด๋Š” ๋งŽ์€ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ์˜ํ•ด ๋’ท๋ฐ›์นจ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[100, 101, 44, 102]. ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด, ์ƒ์ž ๋ฐ– ๋ชจ๋ธ๋„ ์†Œ๊ทœ๋ชจ Fine-tuning ์˜ˆ์‚ฐ์œผ๋กœ ์‰ฝ๊ฒŒ ์กฐ์ •ํ•˜์—ฌ ๋…์„ฑ์ด ์žˆ๊ณ , ํ˜์˜ค์ ์ด๋ฉฐ, ๋ถˆ์พŒ๊ฐ์„ ์ฃผ๊ณ , ๊นŠ์€ ํŽธ๊ฒฌ์„ ๊ฐ€์ง„ ์ฝ˜ํ…์ธ ๋ฅผ ์‰ฝ๊ฒŒ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[103, 104, 68]. ๊ทธ๋ฆฌ๊ณ  ์ƒ๋‹นํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋ชจ๋ธ์ด ์‚ฌ์ ์ธ ์ •๋ณด๋ฅผ ํ† ํ•ด๋‚ด๊ฒŒ ํ•˜๊ฑฐ๋‚˜[105], ์•ˆ์ „ ํ•„ํ„ฐ๋ฅผ '์žŠ์–ด๋ฒ„๋ฆฌ๊ฒŒ' ํ•˜๊ฑฐ๋‚˜[106], ์„ค๊ณ„์ƒ์˜ ์ทจ์•ฝ์ ์„ ๋“œ๋Ÿฌ๋‚ด๊ฒŒ ํ•˜๋Š” ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„๊ฐ€ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฒฉ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ง‘์ค‘๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค[107].
์ฆ‰๊ฐ์ ์ธ ์œ„ํ—˜์ฆ‰๊ฐ์ ์ธ ์œ„ํ—˜์€ ๊ธฐ์กด์˜ ํ”„๋ก ํ‹ฐ์–ด ๋ฐ ์ƒ์‚ฐ ์ค€๋น„ ๋ชจ๋ธ์—์„œ ์ด๋ฏธ ๋ฐœ์ƒํ•˜๊ณ  ์žˆ๋Š” ์œ„ํ•ด ์š”์†Œ์ž…๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์‚ฌ๊ธฐ ๋ฐ ์‚ฌ๊ธฐ ๊ฐ€๋Šฅ[86], ํ…Œ๋Ÿฌ ํ™œ๋™[87, 88], ํ—ˆ์œ„ ์ •๋ณด ์บ ํŽ˜์ธ[89, 90, 91], ์•„๋™ ์„ฑ ํ•™๋Œ€ ์ž๋ฃŒ ์ œ์ž‘[92], ์ž์‚ด ๋ฐ ์žํ•ด ์กฐ์žฅ[93], ์‚ฌ์ด๋ฒ„ ๊ณต๊ฒฉ ๋ฐ ๋ฉ€์›จ์–ด[94, 95] ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์šฐ๋ ค๋Š” ์‚ฌ์‹ค ์˜ค๋ฅ˜์™€ 'ํ™˜๊ฐ'์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ตœ์‹  ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์™ธ๋ถ€ ์†Œ์Šค์— ์•ก์„ธ์Šคํ•  ์ˆ˜ ์—†๋Š” ๊ฒฝ์šฐ, ๋ชจ๋ธ์ด ๊ต์œก ์ข…๋ฃŒ์ผ ์ดํ›„์— ๋ฐœ์ƒํ•œ ์‚ฌ๊ฑด์— ๋Œ€ํ•œ ์งˆ๋ฌธ์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ์ƒ๋‹นํ•œ ์œ„ํ—˜์ด ๋ฉ๋‹ˆ๋‹ค[6, 97, 98]. ์ƒ์„ฑํ˜• AI๋Š” ์กฐ์ง์  ๋ฐ ๋ฌผ์งˆ์  ์žฅ๋ฒฝ์„ ๋‚ฎ์ถค์œผ๋กœ์จ ์ด๋Ÿฌํ•œ ์œ„ํ—˜์˜ ๊ทœ๋ชจ์™€ ์‹ฌ๊ฐ์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์–ธ๋ก ์—์„œ๋Š” ๋ฒ”์ฃ„์ž๋“ค์ด ํ…์ŠคํŠธ ์Œ์„ฑ ๋ณ€ํ™˜ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ๋“ค์—๊ฒŒ ๋Œ€๋Ÿ‰์œผ๋กœ ์ „ํ™”๋ฅผ ๊ฑธ์–ด ์ฆ‰๊ฐ์ ์ธ ๊ธˆ์ „์  ์ง€์›์ด ํ•„์š”ํ•œ ์นœ์ฒ™ ์ค‘ ํ•œ ๋ช…์ธ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฐ€์žฅํ•˜๋Š” ์‚ฌ์‹ค์ ์ธ ๋ฑ…ํ‚น ์Šค์บ”์„ ์‹คํ–‰ํ–ˆ๋‹ค๊ณ  ๋ณด๋„ํ–ˆ์Šต๋‹ˆ๋‹ค[99]. AI ๋ชจ๋ธ์˜ ํŽธํ–ฅ์„ฑ, ๋ถˆ๊ณต์ •์„ฑ, ์ฐจ๋ณ„์„ฑ์— ๋Œ€ํ•œ ์œ„ํ—˜์€ ์˜ค๋žœ ๊ธฐ๊ฐ„ ๋™์•ˆ ์šฐ๋ ค๋˜์–ด ์™”์œผ๋ฉฐ, ์ด๋Š” ๋งŽ์€ ์—ฐ๊ตฌ ๊ฒฐ๊ณผ์— ์˜ํ•ด ๋’ท๋ฐ›์นจ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[100, 101, 44, 102]. ์ตœ๊ทผ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด, ์ƒ์ž ๋ฐ– ๋ชจ๋ธ๋„ ์†Œ๊ทœ๋ชจ ๋ฏธ์„ธ ์กฐ์ • ์˜ˆ์‚ฐ์œผ๋กœ ์‰ฝ๊ฒŒ ์กฐ์ •ํ•˜์—ฌ ๋…์„ฑ์ด ์žˆ๊ณ , ํ˜์˜ค์ ์ด๋ฉฐ, ๋ถˆ์พŒ๊ฐ์„ ์ฃผ๊ณ , ๊นŠ์€ ํŽธ๊ฒฌ์„ ๊ฐ€์ง„ ์ฝ˜ํ…์ธ ๋ฅผ ์‰ฝ๊ฒŒ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[103, 104, 68]. ๊ทธ๋ฆฌ๊ณ  ์ƒ๋‹นํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋ชจ๋ธ์ด ์‚ฌ์ ์ธ ์ •๋ณด๋ฅผ ํ† ํ•ด๋‚ด๊ฒŒ ํ•˜๊ฑฐ๋‚˜[105], ์•ˆ์ „ ํ•„ํ„ฐ๋ฅผ '์žŠ์–ด๋ฒ„๋ฆฌ๊ฒŒ' ํ•˜๊ฑฐ๋‚˜[106], ์„ค๊ณ„์ƒ์˜ ์ทจ์•ฝ์ ์„ ๋“œ๋Ÿฌ๋‚ด๊ฒŒ ํ•˜๋Š” ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„๊ฐ€ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ฒฉ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์ง‘์ค‘๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค[107].
To further decrease the computing cost, while capturing long-range dependency and maintaining high performance, many new sparse attention based models or new neural network paradigms are proposed [24, 25, 26, 27, 28]. Among them, State Space Model (e.g., Mamba [12], S4 [29], S4nd [30]), as shown in Fig. 1, becomes the center of attention. As shown in the left part of Fig. 2, the amount of SSM-related papers released shows the trend of explosive growth. The State Space Model (SSM) is a framework initially proposed to model a dynamic system using state variables in the field of control theory, computational neuroscience, etc 1. When adapting this concept for deep learning, we usually refer to linear invariant (or stationary) systems. The original SSM is a continuous-dynamic system that can be discretized for _recurrent_ and _convolutional_ views for the computer to handle. SSMs can be adopted for various data processing and feature learning, including image/video data, text data, structured graph data, event streams/point cloud data, multi-modal/multi-media data, audio and speech, time series data, tabular data, etc. It can also be utilized to build efficient generative models, such as SSMs-based diffusion generative models [31, 32, 33]. In order to help readers better understand the SSM and keep track of the latest research progress and various applications, this paper conducts a systematic review of the field and verifies the performance of the SSM model in downstream tasks experimentally. It is hoped that this review can better lead and promote the development of the field of SSM.
์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๊ณ  ๋†’์€ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๋”์šฑ ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ๋งŽ์€ ์ƒˆ๋กœ์šด ํฌ์†Œ attention ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๋˜๋Š” ์ƒˆ๋กœ์šด neural network ํŒจ๋Ÿฌ๋‹ค์ž„์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[24, 25, 26, 27, 28]. ๊ทธ์ค‘์—์„œ๋„ ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด State Space Model(์˜ˆ: Mamba[12], S4[29], S4nd[30])์ด ๊ด€์‹ฌ์˜ ์ค‘์‹ฌ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์˜ ์™ผ์ชฝ์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ SSM ๊ด€๋ จ ๋…ผ๋ฌธ ๋ฐœํ‘œ๋Ÿ‰์€ ํญ๋ฐœ์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” ์ถ”์„ธ๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. State Space Model(SSM)์€ ์ œ์–ด ์ด๋ก , computational neuroscience ๋“ฑ์˜ ๋ถ„์•ผ์—์„œ state ๋ณ€์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์  ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ ํ”„๋ ˆ์ž„์›Œํฌ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐœ๋…์„ deep learning์— ์ ์šฉํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ linear invariant(๋˜๋Š” stationary) ์‹œ์Šคํ…œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜์˜ SSM์€ ์—ฐ์†์ ์ธ ๋™์  ์‹œ์Šคํ…œ์œผ๋กœ, ์ปดํ“จํ„ฐ๊ฐ€ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก _recurrent_ ๋ฐ _convolutional_ ๋ทฐ๋กœ ์ด์‚ฐํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SSM์€ ์ด๋ฏธ์ง€/๋น„๋””์˜ค ๋ฐ์ดํ„ฐ, ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ, ๊ตฌ์กฐํ™”๋œ ๊ทธ๋ž˜ํ”„ ๋ฐ์ดํ„ฐ, ์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ/ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ๋ฐ์ดํ„ฐ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ/๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๋ฐ์ดํ„ฐ, ์˜ค๋””์˜ค ๋ฐ ์Œ์„ฑ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ, ํ‘œ ํ˜•์‹ ๋ฐ์ดํ„ฐ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ๋ฐ ํŠน์ง• ํ•™์Šต์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ SSM ๊ธฐ๋ฐ˜ diffusion generative ๋ชจ๋ธ[31, 32, 33]๊ณผ ๊ฐ™์€ ํšจ์œจ์ ์ธ ์ƒ์„ฑ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ์—๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋…์ž๋“ค์ด SSM์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ์ตœ์‹  ์—ฐ๊ตฌ ์ง„ํ–‰ ์ƒํ™ฉ๊ณผ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ด ๋ถ„์•ผ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ SSM ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์‹คํ—˜์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฆฌ๋ทฐ๊ฐ€ SSM ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ ๋” ์ž˜ ์ด๋Œ๊ณ  ์ด‰์ง„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
์žฅ๊ฑฐ๋ฆฌ ์ข…์†์„ฑ์„ ํฌ์ฐฉํ•˜๊ณ  ๋†’์€ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ๊ณ„์‚ฐ ๋น„์šฉ์„ ๋”์šฑ ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ๋งŽ์€ ์ƒˆ๋กœ์šด ํฌ์†Œ ์ฃผ์˜ ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ๋˜๋Š” ์ƒˆ๋กœ์šด ์‹ ๊ฒฝ๋ง ํŒจ๋Ÿฌ๋‹ค์ž„์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค[24, 25, 26, 27, 28]. ๊ทธ์ค‘์—์„œ๋„ ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(์˜ˆ: ๋ง˜๋ฐ”[12], S4[29], S4nd[30])์ด ๊ด€์‹ฌ์˜ ์ค‘์‹ฌ์ด ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 2์˜ ์™ผ์ชฝ์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ SSM ๊ด€๋ จ ๋…ผ๋ฌธ ๋ฐœํ‘œ๋Ÿ‰์€ ํญ๋ฐœ์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” ์ถ”์„ธ๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค. ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(SSM)์€ ์ œ์–ด ์ด๋ก , ๊ณ„์‚ฐ ์‹ ๊ฒฝ๊ณผํ•™ ๋“ฑ์˜ ๋ถ„์•ผ์—์„œ ์ƒํƒœ ๋ณ€์ˆ˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋™์  ์‹œ์Šคํ…œ์„ ๋ชจ๋ธ๋งํ•˜๊ธฐ ์œ„ํ•ด ์ฒ˜์Œ ์ œ์•ˆ๋œ ํ”„๋ ˆ์ž„์›Œํฌ์ž…๋‹ˆ๋‹ค. ์ด ๊ฐœ๋…์„ ๋”ฅ๋Ÿฌ๋‹์— ์ ์šฉํ•  ๋•Œ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์„ ํ˜• ๋ถˆ๋ณ€(๋˜๋Š” ์ •์ง€) ์‹œ์Šคํ…œ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜์˜ SSM์€ ์—ฐ์†์ ์ธ ๋™์  ์‹œ์Šคํ…œ์œผ๋กœ, ์ปดํ“จํ„ฐ๊ฐ€ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก _๋ฐ˜๋ณต_ ๋ฐ _์ปจ๋ณผ๋ฃจ์…˜_ ๋ทฐ๋กœ ์ด์‚ฐํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. SSM์€ ์ด๋ฏธ์ง€/๋น„๋””์˜ค ๋ฐ์ดํ„ฐ, ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ, ๊ตฌ์กฐํ™”๋œ ๊ทธ๋ž˜ํ”„ ๋ฐ์ดํ„ฐ, ์ด๋ฒคํŠธ ์ŠคํŠธ๋ฆผ/ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ๋ฐ์ดํ„ฐ, ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ/๋ฉ€ํ‹ฐ๋ฏธ๋””์–ด ๋ฐ์ดํ„ฐ, ์˜ค๋””์˜ค ๋ฐ ์Œ์„ฑ, ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ, ํ‘œ ํ˜•์‹ ๋ฐ์ดํ„ฐ ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ๋ฐ ํŠน์ง• ํ•™์Šต์— ์ ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ SSM ๊ธฐ๋ฐ˜ ํ™•์‚ฐ ์ƒ์„ฑ ๋ชจ๋ธ[31, 32, 33]๊ณผ ๊ฐ™์€ ํšจ์œจ์ ์ธ ์ƒ์„ฑ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๋Š” ๋ฐ์—๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐฑ์„œ์—์„œ๋Š” ๋…์ž๋“ค์ด SSM์„ ๋” ์ž˜ ์ดํ•ดํ•˜๊ณ  ์ตœ์‹  ์—ฐ๊ตฌ ์ง„ํ–‰ ์ƒํ™ฉ๊ณผ ๋‹ค์–‘ํ•œ ์‘์šฉ ๋ถ„์•ผ๋ฅผ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ด ๋ถ„์•ผ์— ๋Œ€ํ•œ ์ฒด๊ณ„์ ์ธ ๊ฒ€ํ† ๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์—์„œ SSM ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ์‹คํ—˜์ ์œผ๋กœ ๊ฒ€์ฆํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฆฌ๋ทฐ๊ฐ€ SSM ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ ๋” ์ž˜ ์ด๋Œ๊ณ  ์ด‰์ง„ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜๊ธฐ๋ฅผ ๋ฐ”๋ž๋‹ˆ๋‹ค.
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call _Direct Preference Optimization_ (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.
๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ์–ธ์–ด ๋ชจ๋ธ(LM)์€ ๊ด‘๋ฒ”์œ„ํ•œ ์„ธ๊ณ„ ์ง€์‹๊ณผ ์ผ๋ถ€ ์ถ”๋ก  ๊ธฐ์ˆ ์„ ํ•™์Šตํ•˜์ง€๋งŒ, ์™„์ „ํžˆ ๋น„์ง€๋„์ ์ธ ํ›ˆ๋ จ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ํ–‰๋™์„ ์ •๋ฐ€ํ•˜๊ฒŒ ์ œ์–ดํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐ์ข…์„ฑ์„ ์–ป๊ธฐ ์œ„ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์ƒ์„ฑ์˜ ์ƒ๋Œ€์  ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ๋ ˆ์ด๋ธ”์„ ์ˆ˜์ง‘ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋น„์ง€๋„ LM์„ Fine-tuningํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์ข…์ข… Reinforcement Learning from Human Feedback (RLHF)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RLHF๋Š” ๋ณต์žกํ•˜๊ณ  ์ข…์ข… ๋ถˆ์•ˆ์ •ํ•œ ์ ˆ์ฐจ๋กœ, ๋จผ์ € ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ์„ ๋งž์ถ˜ ๋‹ค์Œ, Reinforcement Learning์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ LM์„ Fine-tuningํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์—์„œ ๋„ˆ๋ฌด ๋ฉ€๋ฆฌ ๋ฒ—์–ด๋‚˜์ง€ ์•Š์œผ๋ฉด์„œ ์ด ์˜ˆ์ƒ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” RLHF์—์„œ ๋ณด์ƒ ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋ฅผ ์†Œ๊ฐœํ•˜์—ฌ ํ•ด๋‹น ์ตœ์ ์˜ ์ •์ฑ…์„ ํ์‡„ ํ˜•์‹์œผ๋กœ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ๋ถ„๋ฅ˜ Loss๋งŒ์œผ๋กœ ํ‘œ์ค€ RLHF ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. Direct Preference Optimization (DPO)๋ผ๊ณ  ํ•˜๋Š” ๊ฒฐ๊ณผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์•ˆ์ •์ ์ด๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ ๊ฒŒ ๋“ค๊ธฐ ๋•Œ๋ฌธ์— Fine-tuning ์ค‘์— LM์—์„œ ์ƒ˜ํ”Œ๋งํ•˜๊ฑฐ๋‚˜ ์ค‘์š”ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์กฐ์ •ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, DPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Fine-tuningํ•˜๋ฉด ๊ธฐ์กด ๋ฐฉ๋ฒ•๋งŒํผ ๋˜๋Š” ๊ทธ ์ด์ƒ์œผ๋กœ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ LM์„ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, DPO๋กœ Fine-tuningํ•˜๋ฉด ์ƒ์„ฑ์˜ ๊ฐ์„ฑ์„ ์ œ์–ดํ•˜๋Š” ๋Šฅ๋ ฅ์ด PPO ๊ธฐ๋ฐ˜ RLHF๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ๊ตฌํ˜„ ๋ฐ ํ›ˆ๋ จ์ด ํ›จ์”ฌ ๊ฐ„๋‹จํ•˜๋ฉด์„œ ์š”์•ฝ ๋ฐ ๋‹จ๋ฐœ ๋Œ€ํ™”์—์„œ ์‘๋‹ต ํ’ˆ์งˆ์„ ๋”ฐ๋ผ์žก๊ฑฐ๋‚˜ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ ์–ธ์–ด ๋ชจ๋ธ(LM)์€ ๊ด‘๋ฒ”์œ„ํ•œ ์„ธ๊ณ„ ์ง€์‹๊ณผ ์ผ๋ถ€ ์ถ”๋ก  ๊ธฐ์ˆ ์„ ํ•™์Šตํ•˜์ง€๋งŒ, ์™„์ „ํžˆ ๋น„์ง€๋„์ ์ธ ํ›ˆ๋ จ ํŠน์„ฑ์œผ๋กœ ์ธํ•ด ํ–‰๋™์„ ์ •๋ฐ€ํ•˜๊ฒŒ ์ œ์–ดํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐ์ข…์„ฑ์„ ์–ป๊ธฐ ์œ„ํ•œ ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ ์ƒ์„ฑ์˜ ์ƒ๋Œ€์  ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ธ๊ฐ„์˜ ๋ ˆ์ด๋ธ”์„ ์ˆ˜์ง‘ํ•˜๊ณ  ์ด๋Ÿฌํ•œ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ ๋น„์ง€๋„ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์ข…์ข… ์ธ๊ฐ„ ํ”ผ๋“œ๋ฐฑ(RLHF)์„ ํ†ตํ•œ ๊ฐ•ํ™” ํ•™์Šต(RLHF)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ RLHF๋Š” ๋ณต์žกํ•˜๊ณ  ์ข…์ข… ๋ถˆ์•ˆ์ •ํ•œ ์ ˆ์ฐจ๋กœ, ๋จผ์ € ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ๋ณด์ƒ ๋ชจ๋ธ์„ ๋งž์ถ˜ ๋‹ค์Œ, ๊ฐ•ํ™” ํ•™์Šต์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€๊ทœ๋ชจ ๋น„์ง€๋„ LM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ์›๋ž˜ ๋ชจ๋ธ์—์„œ ๋„ˆ๋ฌด ๋ฉ€๋ฆฌ ๋ฒ—์–ด๋‚˜์ง€ ์•Š์œผ๋ฉด์„œ ์ด ์˜ˆ์ƒ ๋ณด์ƒ์„ ์ตœ๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์—์„œ๋Š” RLHF์—์„œ ๋ณด์ƒ ๋ชจ๋ธ์˜ ์ƒˆ๋กœ์šด ๋งค๊ฐœ๋ณ€์ˆ˜ํ™”๋ฅผ ์†Œ๊ฐœํ•˜์—ฌ ํ•ด๋‹น ์ตœ์ ์˜ ์ •์ฑ…์„ ํ์‡„ ํ˜•์‹์œผ๋กœ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜์—ฌ ๊ฐ„๋‹จํ•œ ๋ถ„๋ฅ˜ ์†์‹ค๋งŒ์œผ๋กœ ํ‘œ์ค€ RLHF ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ง์ ‘ ์„ ํ˜ธ๋„ ์ตœ์ ํ™”(_Direct Preference Optimization_ (DPO))๋ผ๊ณ  ํ•˜๋Š” ๊ฒฐ๊ณผ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์•ˆ์ •์ ์ด๊ณ  ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ๊ณ„์‚ฐ ๋น„์šฉ์ด ์ ๊ฒŒ ๋“ค๊ธฐ ๋•Œ๋ฌธ์— ๋ฏธ์„ธ ์กฐ์ • ์ค‘์— LM์—์„œ ์ƒ˜ํ”Œ๋งํ•˜๊ฑฐ๋‚˜ ์ค‘์š”ํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์กฐ์ •ํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ, DPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉด ๊ธฐ์กด ๋ฐฉ๋ฒ•๋งŒํผ ๋˜๋Š” ๊ทธ ์ด์ƒ์œผ๋กœ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„์— ๋งž๊ฒŒ LM์„ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, DPO๋กœ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋ฉด ์ƒ์„ฑ์˜ ๊ฐ์„ฑ์„ ์ œ์–ดํ•˜๋Š” ๋Šฅ๋ ฅ์ด PPO ๊ธฐ๋ฐ˜ RLHF๋ฅผ ๋Šฅ๊ฐ€ํ•˜๋ฉฐ, ๊ตฌํ˜„ ๋ฐ ํ›ˆ๋ จ์ด ํ›จ์”ฌ ๊ฐ„๋‹จํ•˜๋ฉด์„œ ์š”์•ฝ ๋ฐ ๋‹จ๋ฐœ ๋Œ€ํ™”์—์„œ ์‘๋‹ต ํ’ˆ์งˆ์„ ๋”ฐ๋ผ์žก๊ฑฐ๋‚˜ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Underperforming specialized generation._ Although LLMs have learned general language patterns to generate coherent text, their proficiency in generation might be constrained when dealing with a specialized domain or task. For instance, a language model that has been trained on general web articles may face challenges when generating a medical report which involves many medical jargon and methods. Intuitively, domain knowledge should be critical for model specialization. However, it is not easy to inject such specialized knowledge into LLMs. As discussed in recent analyses [47, 648], when LLMs are trained to exhibit some specific ability that allows them to excel in some areas, they might struggle in others. Such an issue is related to _catastrophic forgetting_[649, 650] in training neural networks, which refers to the conflict phenomenon of integrating new and old knowledge. Similar cases also occur in human alignment of LLMs, where _"alignment tax"_[66] (_e.g.,_ a potential loss in the in-context learning ability) has to be paid for aligning to human values and needs. Moreover, due to the limitations of sequence modeling architecture, LLMs still face challenges in the understanding and generation of structured data. Consequently, they often fall behind task-specific models on complex structured data tasks, such as knowledge-base question answering and semantic parsing [651, 458]. Therefore, it is important to develop effective model specialization methods that can flexibly adapt LLMs to various task scenarios, meanwhile retaining the original abilities as possible.
\(\bullet\)_์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ํŠน์ˆ˜ ์ƒ์„ฑ._ LLM์€ ์ผ๊ด€์„ฑ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ํŒจํ„ด์„ ํ•™์Šตํ–ˆ์ง€๋งŒ, ํŠน์ˆ˜ ๋„๋ฉ”์ธ์ด๋‚˜ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ฐ˜ ์›น ๊ธฐ์‚ฌ๋ฅผ ํ•™์Šตํ•œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋งŽ์€ ์˜ํ•™ ์ „๋ฌธ ์šฉ์–ด์™€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋œ ์˜๋ฃŒ ๋ณด๊ณ ์„œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ ๋„๋ฉ”์ธ ์ง€์‹์€ ๋ชจ๋ธ ์ „๋ฌธํ™”์— ๋งค์šฐ ์ค‘์š”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์„ LLM์— ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ถ„์„[47, 648]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์ด ์ผ๋ถ€ ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŠน์ • ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ›ˆ๋ จํ•˜๋ฉด ๋‹ค๋ฅธ ์˜์—ญ์—์„œ๋Š” ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋Š” ์‹ ๊ฒฝ๋ง ํ›ˆ๋ จ์—์„œ _catastrophic forgetting_[649, 650]๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ƒˆ๋กœ์šด ์ง€์‹๊ณผ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฐˆ๋“ฑ ํ˜„์ƒ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ํ•„์š”์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด _"alignment tax"_[66](_์˜ˆ:_ In-Context Learning ๋Šฅ๋ ฅ์˜ ์ž ์žฌ์  ์†์‹ค)์„ ์ง€๋ถˆํ•ด์•ผ ํ•˜๋Š” LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ์—์„œ๋„ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„๋กœ ์ธํ•ด LLM์€ ์—ฌ์ „ํžˆ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ์ดํ•ด์™€ ์ƒ์„ฑ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ฐ ์‹œ๋งจํ‹ฑ ๊ตฌ๋ฌธ ๋ถ„์„[651, 458]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ์ž‘์—…์—์„œ ์ข…์ข… ์ž‘์—…๋ณ„ ๋ชจ๋ธ์— ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์ž‘์—… ์‹œ๋‚˜๋ฆฌ์˜ค์— LLM์„ ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉํ•˜๋Š” ๋™์‹œ์— ์›๋ž˜์˜ ๋Šฅ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ์ ์ธ ๋ชจ๋ธ ์ „๋ฌธํ™” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_์„ฑ๋Šฅ์ด ์ €์กฐํ•œ ํŠน์ˆ˜ ์ƒ์„ฑ._ LLM์€ ์ผ๊ด€์„ฑ ์žˆ๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ฐ˜์ ์ธ ์–ธ์–ด ํŒจํ„ด์„ ํ•™์Šตํ–ˆ์ง€๋งŒ, ํŠน์ˆ˜ ๋„๋ฉ”์ธ์ด๋‚˜ ์ž‘์—…์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” ์ƒ์„ฑ ๋Šฅ๋ ฅ์ด ์ œํ•œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ผ๋ฐ˜ ์›น ๊ธฐ์‚ฌ๋ฅผ ํ•™์Šตํ•œ ์–ธ์–ด ๋ชจ๋ธ์€ ๋งŽ์€ ์˜ํ•™ ์ „๋ฌธ ์šฉ์–ด์™€ ๋ฐฉ๋ฒ•์ด ํฌํ•จ๋œ ์˜๋ฃŒ ๋ณด๊ณ ์„œ๋ฅผ ์ƒ์„ฑํ•  ๋•Œ ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ง๊ด€์ ์œผ๋กœ ๋„๋ฉ”์ธ ์ง€์‹์€ ๋ชจ๋ธ ์ „๋ฌธํ™”์— ๋งค์šฐ ์ค‘์š”ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ „๋ฌธ ์ง€์‹์„ LLM์— ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์‰ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ ๋ถ„์„[47, 648]์—์„œ ๋…ผ์˜ํ•œ ๋ฐ”์™€ ๊ฐ™์ด, LLM์ด ์ผ๋ถ€ ์˜์—ญ์—์„œ ๋›ฐ์–ด๋‚œ ๋Šฅ๋ ฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ๋„๋ก ํŠน์ • ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋„๋ก ํ›ˆ๋ จํ•˜๋ฉด ๋‹ค๋ฅธ ์˜์—ญ์—์„œ๋Š” ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋Š” ์‹ ๊ฒฝ๋ง ํ›ˆ๋ จ์—์„œ _์žฌ์•™์  ๋ง๊ฐ_[649, 650]๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š”๋ฐ, ์ด๋Š” ์ƒˆ๋กœ์šด ์ง€์‹๊ณผ ๊ธฐ์กด ์ง€์‹์„ ํ†ตํ•ฉํ•˜๋Š” ๊ฐˆ๋“ฑ ํ˜„์ƒ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‚ฌ๋ก€๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ํ•„์š”์— ๋งž์ถ”๊ธฐ ์œ„ํ•ด _"์ •๋ ฌ ์„ธ๊ธˆ"_[66](_์˜ˆ:_ ์ธ์ปจํ…์ŠคํŠธ ํ•™์Šต ๋Šฅ๋ ฅ์˜ ์ž ์žฌ์  ์†์‹ค)์„ ์ง€๋ถˆํ•ด์•ผ ํ•˜๋Š” LLM์˜ ์ธ๊ฐ„ ์ •๋ ฌ์—์„œ๋„ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‹œํ€€์Šค ๋ชจ๋ธ๋ง ์•„ํ‚คํ…์ฒ˜์˜ ํ•œ๊ณ„๋กœ ์ธํ•ด LLM์€ ์—ฌ์ „ํžˆ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ์˜ ์ดํ•ด์™€ ์ƒ์„ฑ์— ์–ด๋ ค์›€์„ ๊ฒช๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ง€์‹ ๊ธฐ๋ฐ˜ ์งˆ์˜์‘๋‹ต ๋ฐ ์‹œ๋งจํ‹ฑ ๊ตฌ๋ฌธ ๋ถ„์„[651, 458]๊ณผ ๊ฐ™์€ ๋ณต์žกํ•œ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ์ž‘์—…์—์„œ ์ข…์ข… ์ž‘์—…๋ณ„ ๋ชจ๋ธ์— ๋’ค์ฒ˜์ง‘๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ์ž‘์—… ์‹œ๋‚˜๋ฆฌ์˜ค์— LLM์„ ์œ ์—ฐํ•˜๊ฒŒ ์ ์šฉํ•˜๋Š” ๋™์‹œ์— ์›๋ž˜์˜ ๋Šฅ๋ ฅ์„ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ํšจ๊ณผ์ ์ธ ๋ชจ๋ธ ์ „๋ฌธํ™” ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Some works focus on mitigating factual mistakes. He et al. (2023) introduces external knowledge to evaluate reasoning chains and votes to filter out chains that contain factual mistakes but without correcting them Wang et al. (2023) adopts a similar way, with the difference that it additionally introduces a reflection mechanism to correct low-scoring reasoning. Zhao et al. (2023) filters out low-confidence reasoning by consistency and guides models to re-reasoning based on relevant external knowledge. While the aforementioned methods work well on knowledge-intensive tasks, they fall short in addressing the challenge of contextual inconsistencies. Zhang et al. (2023) explores the hallucination snowballing phenomena during the reasoning process. Others aim to address the inconsistency issues. Radhakrishnan et al. (2023) observes that models are more faithful when dealing with simple questions. Thus, it improves faithfulness through question decomposition. Faithful CoT (Lyu et al., 2023) initially generates symbolic reasoning chains and later deterministically executes symbolic functions, mitigating reasoning inconsistencies. Lanham et al. (2023) explores the factors that influence faithfulness, which provides an empirical perspective. It finds faithfulness varies on different tasks and decreases as the model size increases. CoNLI (Lei et al., 2023) proposes a post-editing strategy to diminish the hallucinations. SynTra (Jones et al., 2023) performs prefix-tuning on a synthetic dataset designed to elicit hallucination easily, and then transfers this capability to real tasks.
์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ์‹ค ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. He ๋“ฑ(2023)์€ ์ถ”๋ก  ์ฒด์ธ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‚ฌ์‹ค ์˜ค๋ฅ˜๊ฐ€ ํฌํ•จ๋œ ์ฒด์ธ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜์ง€๋งŒ ์ด๋ฅผ ์ˆ˜์ •ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. Wang ๋“ฑ(2023)์€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์ง€๋งŒ, ์ ์ˆ˜๊ฐ€ ๋‚ฎ์€ ์ถ”๋ก ์„ ์ˆ˜์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ˜์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ๋„์ž…ํ•œ๋‹ค๋Š” ์ฐจ์ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Zhao ๋“ฑ(2023)์€ ์ผ๊ด€์„ฑ์„ ํ†ตํ•ด ์‹ ๋ขฐ๋„๊ฐ€ ๋‚ฎ์€ ์ถ”๋ก ์„ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๊ด€๋ จ ์™ธ๋ถ€ ์ง€์‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์ด ๋‹ค์‹œ ์ถ”๋ก ํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฐฉ๋ฒ•์€ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…์—์„œ ์ž˜ ์ž‘๋™ํ•˜์ง€๋งŒ, ์ƒํ™ฉ์— ๋”ฐ๋ฅธ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. Zhang ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์—์„œ hallucination snowballing ํ˜„์ƒ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ๋“ค์€ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. Radhakrishnan ๋“ฑ(2023)์€ ๋ชจ๋ธ์ด ๋‹จ์ˆœํ•œ ์งˆ๋ฌธ์„ ๋‹ค๋ฃฐ ๋•Œ ๋” ์ถฉ์‹คํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์งˆ๋ฌธ ๋ถ„ํ•ด๋ฅผ ํ†ตํ•ด ์ถฉ์‹ค๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. Faithful CoT(Lyu et al., 2023)๋Š” ์ฒ˜์Œ์— ๊ธฐํ˜ธ ์ถ”๋ก  ์ฒด์ธ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๋‚˜์ค‘์— ๊ธฐํ˜ธ ํ•จ์ˆ˜๋ฅผ ๊ฒฐ์ •๋ก ์ ์œผ๋กœ ์‹คํ–‰ํ•˜์—ฌ ์ถ”๋ก  ๋ถˆ์ผ์น˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. Lanham ๋“ฑ(2023)์€ ์ถฉ์‹ค๋„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์š”์ธ์„ ํƒ๊ตฌํ•˜์—ฌ ๊ฒฝํ—˜์  ๊ด€์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ถฉ์‹ค๋„๊ฐ€ ์ž‘์—…๋งˆ๋‹ค ๋‹ค๋ฅด๋ฉฐ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์ถฉ์‹ค๋„๊ฐ€ ๊ฐ์†Œํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. CoNLI(Lei et al., 2023)๋Š” hallucinations์„ ์ค„์ด๊ธฐ ์œ„ํ•œ ์‚ฌํ›„ ํŽธ์ง‘ ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. SynTra(Jones et al., 2023)๋Š” hallucination์„ ์‰ฝ๊ฒŒ ์œ ๋ฐœํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด prefix-tuning์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ ์ด ๊ธฐ๋Šฅ์„ ์‹ค์ œ ์ž‘์—…์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค.
์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ์‚ฌ์‹ค ์˜ค๋ฅ˜๋ฅผ ์™„ํ™”ํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. He ๋“ฑ(2023)์€ ์ถ”๋ก  ์ฒด์ธ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์‚ฌ์‹ค ์˜ค๋ฅ˜๊ฐ€ ํฌํ•จ๋œ ์ฒด์ธ์„ ๊ฑธ๋Ÿฌ๋‚ด๊ธฐ ์œ„ํ•ด ์™ธ๋ถ€ ์ง€์‹์„ ๋„์ž…ํ•˜์ง€๋งŒ ์ด๋ฅผ ์ˆ˜์ •ํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. Wang ๋“ฑ(2023)์€ ์œ ์‚ฌํ•œ ๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜์ง€๋งŒ, ์ ์ˆ˜๊ฐ€ ๋‚ฎ์€ ์ถ”๋ก ์„ ์ˆ˜์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ˜์˜ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์ถ”๊ฐ€๋กœ ๋„์ž…ํ•œ๋‹ค๋Š” ์ฐจ์ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. Zhao ๋“ฑ(2023)์€ ์ผ๊ด€์„ฑ์„ ํ†ตํ•ด ์‹ ๋ขฐ๋„๊ฐ€ ๋‚ฎ์€ ์ถ”๋ก ์„ ํ•„ํ„ฐ๋งํ•˜๊ณ  ๊ด€๋ จ ์™ธ๋ถ€ ์ง€์‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์ด ๋‹ค์‹œ ์ถ”๋ก ํ•˜๋„๋ก ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค. ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๋ฐฉ๋ฒ•์€ ์ง€์‹ ์ง‘์•ฝ์ ์ธ ์ž‘์—…์—์„œ ์ž˜ ์ž‘๋™ํ•˜์ง€๋งŒ, ์ƒํ™ฉ์— ๋”ฐ๋ฅธ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. Zhang ๋“ฑ(2023)์€ ์ถ”๋ก  ๊ณผ์ •์—์„œ ํ™€๋ฃจ์‹œ๋„ค์ด์…˜ ์Šค๋…ธ์šฐ๋ณผ๋ง ํ˜„์ƒ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์—ฐ๊ตฌ๋“ค์€ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. Radhakrishnan ๋“ฑ(2023)์€ ๋ชจ๋ธ์ด ๋‹จ์ˆœํ•œ ์งˆ๋ฌธ์„ ๋‹ค๋ฃฐ ๋•Œ ๋” ์ถฉ์‹คํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์งˆ๋ฌธ ๋ถ„ํ•ด๋ฅผ ํ†ตํ•ด ์ถฉ์‹ค๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. Faithful CoT(Lyu et al., 2023)๋Š” ์ฒ˜์Œ์— ๊ธฐํ˜ธ ์ถ”๋ก  ์ฒด์ธ์„ ์ƒ์„ฑํ•œ ๋‹ค์Œ ๋‚˜์ค‘์— ๊ธฐํ˜ธ ํ•จ์ˆ˜๋ฅผ ๊ฒฐ์ •๋ก ์ ์œผ๋กœ ์‹คํ–‰ํ•˜์—ฌ ์ถ”๋ก  ๋ถˆ์ผ์น˜๋ฅผ ์™„ํ™”ํ•ฉ๋‹ˆ๋‹ค. Lanham ๋“ฑ(2023)์€ ์ถฉ์‹ค๋„์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์š”์ธ์„ ํƒ๊ตฌํ•˜์—ฌ ๊ฒฝํ—˜์  ๊ด€์ ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ๋Š” ์ถฉ์‹ค๋„๊ฐ€ ์ž‘์—…๋งˆ๋‹ค ๋‹ค๋ฅด๋ฉฐ ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ปค์งˆ์ˆ˜๋ก ์ถฉ์‹ค๋„๊ฐ€ ๊ฐ์†Œํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. CoNLI(Lei et al., 2023)๋Š” ํ™˜๊ฐ์„ ์ค„์ด๊ธฐ ์œ„ํ•œ ์‚ฌํ›„ ํŽธ์ง‘ ์ „๋žต์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. SynTra(Jones et al., 2023)๋Š” ํ™˜๊ฐ์„ ์‰ฝ๊ฒŒ ์œ ๋ฐœํ•˜๋„๋ก ์„ค๊ณ„๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด ์ ‘๋‘์‚ฌ ํŠœ๋‹์„ ์ˆ˜ํ–‰ํ•œ ๋‹ค์Œ ์ด ๊ธฐ๋Šฅ์„ ์‹ค์ œ ์ž‘์—…์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค.
In this paper, we introduce MiniCPM, a series of SLMs, which primarily builds on two models, endowed with 2.4B and 1.2B non-embedding parameters respectively, and they rank preeminently in their respective 2B and 1B scale categories. MiniCPM also exhibits comparable capabilities to those of 7B\(\sim\)13B language models, such as Llama2-7B (Touvron et al., 2023), Mistral-7B (Jiang et al., 2023), Gemma-7B (Banks and Warkentin, 2024), and Llama-13B (Touvron et al., 2023), etc. Notwithstanding their small model sizes, our training methodology is meticulously designed to facilitate seamless scaling of both model scale and data horizons. This is exemplified through our model wind tunnel experiments that encompasses comprehensive hyper-parameter optimization (Section 3), and the deployment of a WSD (Warmup-Stable-Decay) learning rate scheduler (Section 4). The latter is tailored for continuous training with an un-predefined pre-training token number, and makes the reusing of model intermediate checkpoints highly feasible. A detailed analysis of the training dynamics of MiniCPM is presented, suggesting that the WSD scheduler demonstrates the intriguing loss landscape of model pre-training. With WSD scheduler, we are now also capable of studying the data-model scaling law with linear effort on model axis and a negligible effort on data axis, while the traditional ones need quadratic effort considering the scaling along both model and data axes. The result of scaling law indicates a much higher data size / model size ratio compared with Chinchilla Optimal (Hoffmann et al., 2022).
์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ฃผ๋กœ 2.4B์™€ 1.2B์˜ ๋น„์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–์ถ˜ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์ผ๋ จ์˜ SLM์ธ MiniCPM์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ๊ฐ๊ฐ์˜ 2B ๋ฐ 1B ๊ทœ๋ชจ ๋ฒ”์ฃผ์—์„œ ์šฐ์ˆ˜ํ•œ ์ˆœ์œ„๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ MiniCPM์€ Llama2-7B(Touvron et al., 2023), Mistral-7B(Jiang et al., 2023), Gemma-7B(Banks and Warkentin, 2024), Llama-13B(Touvron et al., 2023) ๋“ฑ 7B~13B Large Language Models (LLMs)๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ž‘๋”๋ผ๋„, ์ €ํฌ์˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์€ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ๋ฐ์ดํ„ฐ ์ง€ํ‰์„ ์„ ์›ํ™œํ•˜๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํฌ๊ด„์ ์ธ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ตœ์ ํ™”(์„น์…˜ 3)๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜๊ณผ WSD(Warmup-Stable-Decay) Learning Rate ์Šค์ผ€์ค„๋Ÿฌ(์„น์…˜ 4)๋ฅผ ํ†ตํ•ด ๊ตฌํ˜„๋ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฏธ๋ฆฌ ์ •์˜๋˜์ง€ ์•Š์€ ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ ์ˆ˜๋กœ ์—ฐ์† ํ•™์Šต์„ ์œ„ํ•ด ๋งž์ถคํ™”๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์žฌ์‚ฌ์šฉ์ด ๋งค์šฐ ์šฉ์ดํ•ฉ๋‹ˆ๋‹ค. MiniCPM์˜ ํ›ˆ๋ จ ๋™์—ญํ•™์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์ด ์ œ์‹œ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” WSD ์Šค์ผ€์ค„๋Ÿฌ๊ฐ€ ๋ชจ๋ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํฅ๋ฏธ๋กœ์šด Loss ํ™˜๊ฒฝ์„ ๋ณด์—ฌ์ค€๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. WSD ์Šค์ผ€์ค„๋Ÿฌ๋ฅผ ํ†ตํ•ด ์ด์ œ ๋ชจ๋ธ ์ถ•์˜ ์„ ํ˜• ๋…ธ๋ ฅ๊ณผ ๋ฐ์ดํ„ฐ ์ถ•์˜ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ๋…ธ๋ ฅ์œผ๋กœ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ Scaling Law๋ฅผ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์ถ• ๋ชจ๋‘๋ฅผ ๋”ฐ๋ผ ์Šค์ผ€์ผ๋งํ•ด์•ผ ํ•˜๋ฏ€๋กœ ์ด์ฐจ์ ์ธ ๋…ธ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. Scaling Law์˜ ๊ฒฐ๊ณผ๋Š” Chinchilla Optimal(Hoffmann et al., 2022)๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ/๋ชจ๋ธ ํฌ๊ธฐ ๋น„์œจ์ด ํ›จ์”ฌ ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ์ฃผ๋กœ 2.4B์™€ 1.2B์˜ ๋น„์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–์ถ˜ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ์ผ๋ จ์˜ SLM์ธ MiniCPM์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ๊ฐ๊ฐ์˜ 2B ๋ฐ 1B ๊ทœ๋ชจ ๋ฒ”์ฃผ์—์„œ ์šฐ์ˆ˜ํ•œ ์ˆœ์œ„๋ฅผ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ MiniCPM์€ Llama2-7B(Touvron et al., 2023), Mistral-7B(Jiang et al., 2023), Gemma-7B(Banks and Warkentin, 2024), Llama-13B(Touvron et al., 2023) ๋“ฑ 7B\(\sim\)13B ์–ธ์–ด ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋ธ ํฌ๊ธฐ๊ฐ€ ์ž‘๋”๋ผ๋„, ์ €ํฌ์˜ ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก ์€ ๋ชจ๋ธ ๊ทœ๋ชจ์™€ ๋ฐ์ดํ„ฐ ์ง€ํ‰์„ ์„ ์›ํ™œํ•˜๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ธ์‹ฌํ•˜๊ฒŒ ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํฌ๊ด„์ ์ธ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ตœ์ ํ™”(์„น์…˜ 3)๋ฅผ ํฌํ•จํ•˜๋Š” ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜๊ณผ WSD(Warmup-Stable-Decay) ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„๋Ÿฌ(์„น์…˜ 4)๋ฅผ ํ†ตํ•ด ๊ตฌํ˜„๋ฉ๋‹ˆ๋‹ค. ํ›„์ž๋Š” ๋ฏธ๋ฆฌ ์ •์˜๋˜์ง€ ์•Š์€ ์‚ฌ์ „ ํ•™์Šต ํ† ํฐ ์ˆ˜๋กœ ์—ฐ์† ํ•™์Šต์„ ์œ„ํ•ด ๋งž์ถคํ™”๋˜์—ˆ์œผ๋ฉฐ, ๋ชจ๋ธ ์ค‘๊ฐ„ ์ฒดํฌํฌ์ธํŠธ์˜ ์žฌ์‚ฌ์šฉ์ด ๋งค์šฐ ์šฉ์ดํ•ฉ๋‹ˆ๋‹ค. MiniCPM์˜ ํ›ˆ๋ จ ๋™์—ญํ•™์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋ถ„์„์ด ์ œ์‹œ๋˜์–ด ์žˆ์œผ๋ฉฐ, ์ด๋Š” WSD ์Šค์ผ€์ค„๋Ÿฌ๊ฐ€ ๋ชจ๋ธ ์‚ฌ์ „ ํ›ˆ๋ จ์˜ ํฅ๋ฏธ๋กœ์šด ์†์‹ค ํ™˜๊ฒฝ์„ ๋ณด์—ฌ์ค€๋‹ค๋Š” ๊ฒƒ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค. WSD ์Šค์ผ€์ค„๋Ÿฌ๋ฅผ ํ†ตํ•ด ์ด์ œ ๋ชจ๋ธ ์ถ•์˜ ์„ ํ˜• ๋…ธ๋ ฅ๊ณผ ๋ฐ์ดํ„ฐ ์ถ•์˜ ๋ฌด์‹œํ•  ์ˆ˜ ์žˆ๋Š” ๋…ธ๋ ฅ์œผ๋กœ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ๊ธฐ์กด ๋ฐฉ๋ฒ•์€ ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์ถ• ๋ชจ๋‘๋ฅผ ๋”ฐ๋ผ ์Šค์ผ€์ผ๋งํ•ด์•ผ ํ•˜๋ฏ€๋กœ ์ด์ฐจ์ ์ธ ๋…ธ๋ ฅ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์˜ ๊ฒฐ๊ณผ๋Š” ์นœ์น ๋ผ ์ตœ์ (Hoffmann et al., 2022)๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ/๋ชจ๋ธ ํฌ๊ธฐ ๋น„์œจ์ด ํ›จ์”ฌ ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
* **Retrieval-on-demand** ([Retrieval]): Given an input and previous-step generation (if applicable), an LM determines whether the continuation requires factual grounding. No indicates retrieval is unnecessary as the sequence does not require factual grounding or may not be enhanced by knowledge retrieval, Yes indicates retrieval is necessary. We additionally have continue to use evidence, which indicates that a model can continue to use the evidence retrieved previously. For instance, a passage may contain rich factual information, and thus Self-Rag generates multiple segments based on the passage. * **Relevant** ([IsIsIs] ): Retrieved knowledge may not be always relevant to the input. This aspect indicates whether the evidence provides useful information (Relevant) or not (Irrelevant). * **Supported** ([IsIsIs] ): Attribution is the concept of whether the output is fully supported by certain evidence (Menick et al., 2022; Bohnet et al., 2022). This aspect judges how much information in the output is entailed by the evidence. We evaluate attributions in three scale, Fully supported, Partially supported, and No support / Contradictory, following Yue et al. (2023); Nakano et al. (2021). * **Useful** ([IsIsIs] ): Following the definitions from Liu et al. (2023), we define the perceived utility as whether the response is a helpful and informative answer to the query, independently from whether it is in fact factual or not. This can be also viewed as plausibility in Menick et al. (2022). For usefulness, we use a five-scale evaluation (1 is the lowest and 5 is the highest).
* **Retrieval-on-demand** ([Retrieval]): ์ž…๋ ฅ๊ณผ ์ด์ „ ๋‹จ๊ณ„ ์ƒ์„ฑ์ด ์ฃผ์–ด์ง€๋ฉด LM์€ ์—ฐ์†์— ์‚ฌ์‹ค์  ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์•„๋‹ˆ์š”๋Š” ์‹œํ€€์Šค์— ์‚ฌ์‹ค์  ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์ง€์‹ ๊ฒ€์ƒ‰์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์˜ˆ๋Š” ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „์— ๊ฒ€์ƒ‰ํ•œ ์ฆ๊ฑฐ๋ฅผ ๊ณ„์† ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ฆ๊ฑฐ๋ฅผ ๊ณ„์† ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•œ ๊ตฌ์ ˆ์— ํ’๋ถ€ํ•œ ์‚ฌ์‹ค ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ Self-RAG์€ ํ•ด๋‹น ๊ตฌ์ ˆ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. * **Relevant** ([IsIsIs]): ๊ฒ€์ƒ‰๋œ ์ง€์‹์ด ํ•ญ์ƒ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ ์ฆ๊ฑฐ๊ฐ€ ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š”์ง€(Relevant) ๋˜๋Š” ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”์ง€(Irrelevant)๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * **Supported** ([IsIsIs]): Attribution์€ ์ถœ๋ ฅ์ด ํŠน์ • ์ฆ๊ฑฐ์— ์˜ํ•ด ์™„์ „ํžˆ ์ง€์›๋˜๋Š”์ง€ ์—ฌ๋ถ€(Menick et al., 2022; Bohnet et al., 2022)๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ ์ถœ๋ ฅ์˜ ์ •๋ณด๊ฐ€ ์ฆ๊ฑฐ์— ์˜ํ•ด ์–ผ๋งˆ๋‚˜ ๋งŽ์ด ์ˆ˜๋ฐ˜๋˜๋Š”์ง€๋ฅผ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ Yue et al. (2023); Nakano et al. (2021)์— ๋”ฐ๋ผ ์„ธ ๊ฐ€์ง€ ์ฒ™๋„, ์ฆ‰ ์™„์ „ํžˆ ์ง€์›, ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›, ์ง€์› ์—†์Œ/๋ชจ์ˆœ์œผ๋กœ ๊ธฐ์—ฌ๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. * **Useful** ([IsIsIs]): Liu et al. (2023)์˜ ์ •์˜์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ์ธ์ง€๋œ ์œ ์šฉ์„ฑ์„ ๋‹ต๋ณ€์ด ์‚ฌ์‹ค์— ๊ทผ๊ฑฐํ•˜๋“  ๊ทธ๋ ‡์ง€ ์•Š๋“  ์ƒ๊ด€์—†์ด ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ์œ ์šฉํ•˜๊ณ  ์œ ์ตํ•œ ๋‹ต๋ณ€์ธ์ง€ ์—ฌ๋ถ€๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” Menick et al. (2022)์—์„œ ํƒ€๋‹น์„ฑ์œผ๋กœ ๋ณผ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์šฉ์„ฑ์˜ ๊ฒฝ์šฐ 5๋‹จ๊ณ„ ํ‰๊ฐ€(1์€ ๊ฐ€์žฅ ๋‚ฎ๊ณ  5๋Š” ๊ฐ€์žฅ ๋†’์Œ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
* **์ˆ˜์š” ์‹œ ๊ฒ€์ƒ‰** ([๊ฒ€์ƒ‰]): ์ž…๋ ฅ๊ณผ ์ด์ „ ๋‹จ๊ณ„ ์ƒ์„ฑ์ด ์ฃผ์–ด์ง€๋ฉด LM์€ ์—ฐ์†์— ์‚ฌ์‹ค์  ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ•œ์ง€ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ์•„๋‹ˆ์š”๋Š” ์‹œํ€€์Šค์— ์‚ฌ์‹ค์  ๊ทผ๊ฑฐ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ์ง€์‹ ๊ฒ€์ƒ‰์œผ๋กœ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์—†์Œ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์˜ˆ๋Š” ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋˜ํ•œ ์ด์ „์— ๊ฒ€์ƒ‰ํ•œ ์ฆ๊ฑฐ๋ฅผ ๊ณ„์† ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ฆ๊ฑฐ๋ฅผ ๊ณ„์† ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ํ•œ ๊ตฌ์ ˆ์— ํ’๋ถ€ํ•œ ์‚ฌ์‹ค ์ •๋ณด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์„ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ Self-Rag์€ ํ•ด๋‹น ๊ตฌ์ ˆ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์—ฌ๋Ÿฌ ์„ธ๊ทธ๋จผํŠธ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. * **๊ด€๋ จ์„ฑ** ([IsIsIs]): ๊ฒ€์ƒ‰๋œ ์ง€์‹์ด ํ•ญ์ƒ ์ž…๋ ฅ๊ณผ ๊ด€๋ จ์ด ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ ์ฆ๊ฑฐ๊ฐ€ ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š”์ง€(๊ด€๋ จ์„ฑ) ๋˜๋Š” ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”์ง€(๊ด€๋ จ์„ฑ ์—†์Œ)๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. * **์ง€์›๋จ** ([IsIsIs]): ๊ธฐ์—ฌ๋„๋Š” ์ถœ๋ ฅ์ด ํŠน์ • ์ฆ๊ฑฐ์— ์˜ํ•ด ์™„์ „ํžˆ ์ง€์›๋˜๋Š”์ง€ ์—ฌ๋ถ€(Menick et al., 2022; Bohnet et al., 2022)๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฐœ๋…์ž…๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ ์ถœ๋ ฅ์˜ ์ •๋ณด๊ฐ€ ์ฆ๊ฑฐ์— ์˜ํ•ด ์–ผ๋งˆ๋‚˜ ๋งŽ์ด ์ˆ˜๋ฐ˜๋˜๋Š”์ง€๋ฅผ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค. ์ด ์ธก๋ฉด์€ Yue et al. (2023); Nakano et al. (2021)์— ๋”ฐ๋ผ ์„ธ ๊ฐ€์ง€ ์ฒ™๋„, ์ฆ‰ ์™„์ „ํžˆ ์ง€์›, ๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›, ์ง€์› ์—†์Œ/๋ชจ์ˆœ์œผ๋กœ ๊ธฐ์—ฌ๋„๋ฅผ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. * **์œ ์šฉ์„ฑ** ([IsIsIs]): Liu et al. (2023)์˜ ์ •์˜์— ๋”ฐ๋ผ, ์šฐ๋ฆฌ๋Š” ์ธ์ง€๋œ ์œ ์šฉ์„ฑ์„ ๋‹ต๋ณ€์ด ์‚ฌ์‹ค์— ๊ทผ๊ฑฐํ•˜๋“  ๊ทธ๋ ‡์ง€ ์•Š๋“  ์ƒ๊ด€์—†์ด ์ฟผ๋ฆฌ์— ๋Œ€ํ•œ ์œ ์šฉํ•˜๊ณ  ์œ ์ตํ•œ ๋‹ต๋ณ€์ธ์ง€ ์—ฌ๋ถ€๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” Menick et al. (2022)์—์„œ ํƒ€๋‹น์„ฑ์œผ๋กœ ๋ณผ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ์šฉ์„ฑ์˜ ๊ฒฝ์šฐ 5๋‹จ๊ณ„ ํ‰๊ฐ€(1์€ ๊ฐ€์žฅ ๋‚ฎ๊ณ  5๋Š” ๊ฐ€์žฅ ๋†’์Œ)๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
**Scalable Pre-training Strategies.** Since the discovery of scaling law (Kaplan et al., 2020; Rae et al., 2021; Aghajanyan et al., 2023), scientifically and predictably (Achiam et al., 2023; Hu et al., 2023; Du et al., 2024) scaling up the LLMs has been pursued from diverse perspectives, especially for the pre-training stage. In terms of training stability, the Tensor Program series (Yang et al., 2022; 2023) is introduced to ensure optimal hyper-parameter consistency across varying model scales, a technique employed in training CerebrasGPT (Dey et al., 2023). Furthermore, Wortsman et al. (2023) suggest leveraging smaller models to anticipate and mitigate instabilities in larger model training. From the training data standpoint, various data-centric strategies have been advocated (Xie et al., 2024; Shi et al., 2023; Ye et al., 2024). In the realm of training methodologies, prior research has delved into diverse learning rate schedulers (LRS) (Howard and Ruder, 2018; Raffel et al., 2020; Hundt et al., 2019), with the Cosine LRS (Loshchilov and Hutter, 2016) emerging as the predominant choice in LLMs. Kaplan et al. (2020) and Hoffmann et al. (2022) have meticulously examined the hyper-parameters of Cosine LRS, thereby laying a foundational groundwork for subsequent pre-training works. Of these, DeepSeek (Bi et al., 2024) bears the closest resemblance to our proposed WSD LRS. Concerning batch size scheduling, Smith et al. (2017) advocates for incrementing batch size as an alternative to diminishing learning rate, a strategy recently adopted by Yi-9B (Young et al., 2024).
**ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์‚ฌ์ „ ํ•™์Šต ์ „๋žต.** ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(Kaplan et al., 2020; Rae et al., 2021; Aghajanyan et al., 2023)์˜ ๋ฐœ๊ฒฌ ์ดํ›„, ๊ณผํ•™์ ์ด๊ณ  ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ(Achiam et al., 2023; Hu et al., 2023; Du et al., 2024) ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ํŠนํžˆ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ ์ถ”๊ตฌ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์•ˆ์ •์„ฑ ์ธก๋ฉด์—์„œ, ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ตœ์ ์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด Tensor Program ์‹œ๋ฆฌ์ฆˆ(Yang et al., 2022; 2023)๊ฐ€ ๋„์ž…๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” CerebrasGPT(Dey et al., 2023)๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ Wortsman et al. (2023)์€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๋” ํฐ ๋ชจ๋ธ ํ›ˆ๋ จ์˜ ๋ถˆ์•ˆ์ •์„ฑ์„ ์˜ˆ์ธกํ•˜๊ณ  ์™„ํ™”ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๊ด€์ ์—์„œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ ์ „๋žต์ด ์˜นํ˜ธ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Xie et al., 2024; Shi et al., 2023; Ye et al., 2024). ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก  ์˜์—ญ์—์„œ ์„ ํ–‰ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ Learning Rate Schedulers (LRS)(Howard and Ruder, 2018; Raffel et al., 2020; Hundt et al., 2019)๋ฅผ ์—ฐ๊ตฌํ–ˆ์œผ๋ฉฐ, Cosine LRS(Loshchilov and Hutter, 2016)๊ฐ€ LLM์—์„œ ์ฃผ๋œ ์„ ํƒ์œผ๋กœ ๋ถ€์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. Kaplan et al. (2020)๊ณผ Hoffmann et al. (2022)์€ Cosine LRS์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ฉด๋ฐ€ํžˆ ์กฐ์‚ฌํ•˜์—ฌ ํ›„์† ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์„ ์œ„ํ•œ ๊ธฐ์ดˆ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘์—์„œ๋„ DeepSeek(Bi et al., 2024)๋Š” ์ €ํฌ๊ฐ€ ์ œ์•ˆํ•œ WSD LRS์™€ ๊ฐ€์žฅ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. Batch Size ์Šค์ผ€์ค„๋ง๊ณผ ๊ด€๋ จํ•˜์—ฌ Smith et al. (2017)์€ Learning Rate ๊ฐ์†Œ์˜ ๋Œ€์•ˆ์œผ๋กœ Batch Size๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์„ ์˜นํ˜ธํ•˜๋Š”๋ฐ, ์ด ์ „๋žต์€ ์ตœ๊ทผ Yi-9B(Young et al., 2024)์—์„œ ์ฑ„ํƒํ•œ ์ „๋žต์ž…๋‹ˆ๋‹ค.
**ํ™•์žฅ ๊ฐ€๋Šฅํ•œ ์‚ฌ์ „ ํ•™์Šต ์ „๋žต.** ์Šค์ผ€์ผ๋ง ๋ฒ•์น™(Kaplan et al., 2020; Rae et al., 2021; Aghajanyan et al., 2023)์˜ ๋ฐœ๊ฒฌ ์ดํ›„, ๊ณผํ•™์ ์ด๊ณ  ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ(Achiam et al., 2023; Hu et al., 2023; Du et al., 2024) ๋ฐฉ์‹์œผ๋กœ LLM์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ํŠนํžˆ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๊ด€์ ์—์„œ ์ถ”๊ตฌ๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์•ˆ์ •์„ฑ ์ธก๋ฉด์—์„œ, ๋‹ค์–‘ํ•œ ๋ชจ๋ธ ๊ทœ๋ชจ์—์„œ ์ตœ์ ์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์ผ๊ด€์„ฑ์„ ๋ณด์žฅํ•˜๊ธฐ ์œ„ํ•ด Tensor Program ์‹œ๋ฆฌ์ฆˆ(Yang et al., 2022; 2023)๊ฐ€ ๋„์ž…๋˜์—ˆ์œผ๋ฉฐ, ์ด๋Š” CerebrasGPT(Dey et al., 2023)๋ฅผ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ Wortsman et al. (2023)์€ ๋” ์ž‘์€ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ๋” ํฐ ๋ชจ๋ธ ํ›ˆ๋ จ์˜ ๋ถˆ์•ˆ์ •์„ฑ์„ ์˜ˆ์ธกํ•˜๊ณ  ์™„ํ™”ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ๊ด€์ ์—์„œ ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ ์ค‘์‹ฌ ์ „๋žต์ด ์˜นํ˜ธ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค(Xie et al., 2024; Shi et al., 2023; Ye et al., 2024). ํ›ˆ๋ จ ๋ฐฉ๋ฒ•๋ก  ์˜์—ญ์—์„œ ์„ ํ–‰ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„๋Ÿฌ(LRS)(Howard and Ruder, 2018; Raffel et al., 2020; Hundt et al., 2019)๋ฅผ ์—ฐ๊ตฌํ–ˆ์œผ๋ฉฐ, ์ฝ”์‚ฌ์ธ LRS(Loshchilov and Hutter, 2016)๊ฐ€ LLM์—์„œ ์ฃผ๋œ ์„ ํƒ์œผ๋กœ ๋ถ€์ƒํ–ˆ์Šต๋‹ˆ๋‹ค. Kaplan et al. (2020)๊ณผ Hoffmann et al. (2022)์€ ์ฝ”์‚ฌ์ธ LRS์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ฉด๋ฐ€ํžˆ ์กฐ์‚ฌํ•˜์—ฌ ํ›„์† ์‚ฌ์ „ ํ•™์Šต ์ž‘์—…์„ ์œ„ํ•œ ๊ธฐ์ดˆ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ค‘์—์„œ๋„ DeepSeek(Bi et al., 2024)๋Š” ์ €ํฌ๊ฐ€ ์ œ์•ˆํ•œ WSD LRS์™€ ๊ฐ€์žฅ ์œ ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ฐฐ์น˜ ํฌ๊ธฐ ์Šค์ผ€์ค„๋ง๊ณผ ๊ด€๋ จํ•˜์—ฌ Smith et al. (2017)์€ ํ•™์Šต ์†๋„ ๊ฐ์†Œ์˜ ๋Œ€์•ˆ์œผ๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์„ ์˜นํ˜ธํ•˜๋Š”๋ฐ, ์ด ์ „๋žต์€ ์ตœ๊ทผ Yi-9B(Young et al., 2024)์—์„œ ์ฑ„ํƒํ•œ ์ „๋žต์ž…๋‹ˆ๋‹ค.
The striking progress of AI in the last few years can be largely attributed to major efforts throughout the world towards _scaling-up_ to ever-larger models and datasets. Large Language Models (LLMs) have steadily increased in size from a mere billion parameters just five years ago (GPT-2 had 1.5 billion parameters [RWC\({}^{+}\)19]) to trillion parameters today. The impetus for this effort originates in the seemingly predictable improvement one obtains by training large models, the so-called _scaling laws_[KMH\({}^{+}\)20, HBM\({}^{+}\)22, MRB\({}^{+}\)23]. However these laws assume a "fixed" data source. This assumption is now significantly disrupted by the existence of frontier LLMs themselves, which allow us to interact with data in novel ways. In our previous works on the phi models [GZA\({}^{+}\)23, LBE\({}^{+}\)23, JBA\({}^{+}\)23] it was shown that a combination of LLM-based filtering of web data, and LLM-created synthetic data, enable performance in smaller language models that were typically seen only in much larger models. For example our previous model trained on this data recipe, **phi-2** (2.7B parameters), matched the performance of models 25 times larger trained on regular data. In this report we present a new model, **phi-3-mini** (3.8B parameters), trained for 3.3T tokens on larger and more advanced versions of the datasets used in **phi-2**. With its small size, **phi-3-mini** can easily be inferenced locally on a modern phone (see Figure 1), yet it achieves a quality that seems on-par with models such as Miktral 8x7B [JSR\({}^{+}\)24] and GPT-3.5.
์ง€๋‚œ ๋ช‡ ๋…„ ๋™์•ˆ AI๊ฐ€ ๋ˆˆ์— ๋„๊ฒŒ ๋ฐœ์ „ํ•œ ๊ฒƒ์€ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ _scaling-up_ํ•˜๊ธฐ ์œ„ํ•œ ์ฃผ์š” ๋…ธ๋ ฅ์— ํž˜์ž…์€ ๋ฐ” ํฌ๋‹ค. Large Language Models (LLMs)์€ ๋ถˆ๊ณผ 5๋…„ ์ „ 10์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ(GPT-2๋Š” 15์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ [RWC\({}^{+}\)19])์— ๋ถˆ๊ณผํ–ˆ๋˜ ๊ทœ๋ชจ๊ฐ€ ์˜ค๋Š˜๋‚ ์—๋Š” ์ˆ˜์กฐ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊พธ์ค€ํžˆ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์˜ ์›๋™๋ ฅ์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ , ์ฆ‰ ์†Œ์œ„ _scaling laws_[KMH\({}^{+}\)20, HBM\({}^{+}\)22, MRB\({}^{+}\)23]์—์„œ ๋น„๋กฏ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฒ•์น™์€ "๊ณ ์ •๋œ" ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ€์ •์€ ์ด์ œ ์ƒˆ๋กœ์šด ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ์™€ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์ฒจ๋‹จ LLM์˜ ์กด์žฌ๋กœ ์ธํ•ด ํฌ๊ฒŒ ๊นจ์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. phi ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด์ „ ์—ฐ๊ตฌ[GZA\({}^{+}\)23, LBE\({}^{+}\)23, JBA\({}^{+}\)23]์—์„œ ์›น ๋ฐ์ดํ„ฐ์˜ LLM ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ๋ง๊ณผ LLM ์ƒ์„ฑ synthetic data์˜ ์กฐํ•ฉ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ํ›จ์”ฌ ๋” ํฐ ๋ชจ๋ธ์—์„œ๋งŒ ๋ณผ ์ˆ˜ ์žˆ์—ˆ๋˜ ์†Œํ˜• ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด ๋ฐ์ดํ„ฐ ๋ ˆ์‹œํ”ผ๋กœ ํ›ˆ๋ จ๋œ ์ด์ „ ๋ชจ๋ธ์ธ **phi-2**(ํŒŒ๋ผ๋ฏธํ„ฐ 27์–ต ๊ฐœ)๋Š” ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ 25๋ฐฐ ๋” ํฐ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ณด๊ณ ์„œ์—์„œ๋Š” **phi-2**์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋” ํฌ๊ณ  ๋ฐœ์ „๋œ ๋ฒ„์ „์—์„œ 3.3T ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ์ƒˆ๋กœ์šด ๋ชจ๋ธ์ธ **phi-3-mini**(ํŒŒ๋ผ๋ฏธํ„ฐ 38์–ต ๊ฐœ)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํฌ๊ธฐ๊ฐ€ ์ž‘์€ **phi-3-mini**๋Š” ์ตœ์‹  ํœด๋Œ€ํฐ์—์„œ ์‰ฝ๊ฒŒ ๋กœ์ปฌ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1 ์ฐธ์กฐ), ๊ทธ๋Ÿฌ๋‚˜ Miktral 8x7B [JSR\({}^{+}\)24] ๋ฐ GPT-3.5์™€ ๊ฐ™์€ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
์ง€๋‚œ ๋ช‡ ๋…„ ๋™์•ˆ AI๊ฐ€ ๋ˆˆ์— ๋„๊ฒŒ ๋ฐœ์ „ํ•œ ๊ฒƒ์€ ์ „ ์„ธ๊ณ„์ ์œผ๋กœ ๋” ํฐ ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ _ํ™•์žฅ_ํ•˜๊ธฐ ์œ„ํ•œ ์ฃผ์š” ๋…ธ๋ ฅ์— ํž˜์ž…์€ ๋ฐ” ํฌ๋‹ค. ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ๋ถˆ๊ณผ 5๋…„ ์ „ 10์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ(GPT-2๋Š” 15์–ต ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ [RWC\({}^{+}\)19])์— ๋ถˆ๊ณผํ–ˆ๋˜ ๊ทœ๋ชจ๊ฐ€ ์˜ค๋Š˜๋‚ ์—๋Š” ์ˆ˜์กฐ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๊พธ์ค€ํžˆ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋…ธ๋ ฅ์˜ ์›๋™๋ ฅ์€ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•จ์œผ๋กœ์จ ์–ป์„ ์ˆ˜ ์žˆ๋Š” ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ๊ฐœ์„ , ์ฆ‰ ์†Œ์œ„ _์Šค์ผ€์ผ๋ง ๋ฒ•์น™_[KMH\({}^{+}\)20, HBM\({}^{+}\)22, MRB\({}^{+}\)23]์—์„œ ๋น„๋กฏ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฒ•์น™์€ "๊ณ ์ •๋œ" ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ€์ •์€ ์ด์ œ ์ƒˆ๋กœ์šด ๋ฐฉ์‹์œผ๋กœ ๋ฐ์ดํ„ฐ์™€ ์ƒํ˜ธ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์ฒจ๋‹จ LLM์˜ ์กด์žฌ๋กœ ์ธํ•ด ํฌ๊ฒŒ ๊นจ์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. phi ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด์ „ ์—ฐ๊ตฌ[GZA\({}^{+}\)23, LBE\({}^{+}\)23, JBA\({}^{+}\)23]์—์„œ ์›น ๋ฐ์ดํ„ฐ์˜ LLM ๊ธฐ๋ฐ˜ ํ•„ํ„ฐ๋ง๊ณผ LLM ์ƒ์„ฑ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์กฐํ•ฉ์ด ์ผ๋ฐ˜์ ์œผ๋กœ ํ›จ์”ฌ ๋” ํฐ ๋ชจ๋ธ์—์„œ๋งŒ ๋ณผ ์ˆ˜ ์žˆ์—ˆ๋˜ ์†Œํ˜• ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ด ๋ฐ์ดํ„ฐ ๋ ˆ์‹œํ”ผ๋กœ ํ›ˆ๋ จ๋œ ์ด์ „ ๋ชจ๋ธ์ธ **phi-2**(ํŒŒ๋ผ๋ฏธํ„ฐ 27์–ต ๊ฐœ)๋Š” ์ผ๋ฐ˜ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จ๋œ 25๋ฐฐ ๋” ํฐ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ๊ณผ ์ผ์น˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ณด๊ณ ์„œ์—์„œ๋Š” **phi-2**์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋” ํฌ๊ณ  ๋ฐœ์ „๋œ ๋ฒ„์ „์—์„œ 3.3T ํ† ํฐ์— ๋Œ€ํ•ด ํ›ˆ๋ จ๋œ ์ƒˆ๋กœ์šด ๋ชจ๋ธ์ธ **phi-3-mini**(ํŒŒ๋ผ๋ฏธํ„ฐ 38์–ต ๊ฐœ)๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํฌ๊ธฐ๊ฐ€ ์ž‘์€ **phi-3-mini**๋Š” ์ตœ์‹  ํœด๋Œ€ํฐ์—์„œ ์‰ฝ๊ฒŒ ๋กœ์ปฌ ์ถ”๋ก ์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ๋ฆผ 1 ์ฐธ์กฐ), ๊ทธ๋Ÿฌ๋‚˜ Miktral 8x7B [JSR\({}^{+}\)24] ๋ฐ GPT-3.5์™€ ๊ฐ™์€ ๋ชจ๋ธ๊ณผ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ํ’ˆ์งˆ์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.
Pipelines for web data.Massive web datasets are typically built upon CommonCrawl, a publicly available scrape of the internet, which has now been running for 12 years and has collected petabytes of data. Working with data scraped from all over the internet presents unique challenges: notably, a significant portion is low-quality machine-generated spam or pornographic content (Trinh and Le, 2018; Kreutzer et al., 2022). Accordingly, training on unfiltered web data is undesirable, resulting in poorly performing models (Raffel et al., 2020). Modern pipelines focus on filtering out this undesirable content (Wenzek et al., 2020). Broadly speaking, these pipelines usually combine a variety of stages: (1) _language identification_, leveraging inexpensive n-gram models (e.g., fastText Joulin et al. (2016)); (2) _filtering rules and heuristics_, such as only keeping lines with valid punctuation, discarding lines with too many symbols, or removing documents containing banned words (Grave et al., 2018; Raffel et al., 2020); (3) _ML-based quality filtering_, using lightweight models trained on known gold data to identify similar high-quality web documents (Wenzek et al., 2020; Brown et al., 2020); (4) _deduplication_, removing either exact duplicate spans or similar documents (Lee et al., 2022). While some filtering is necessary, excessive filtering can introduce undesirable biases in the model. This can overly impact minorities (Dodge et al., 2021), motivating the adoption of practices such as pseudo-crawling, wherein allowed URLs are manually curated (Laurencon et al., 2022).
์›น ๋ฐ์ดํ„ฐ๋ฅผ ์œ„ํ•œ ํŒŒ์ดํ”„๋ผ์ธ. ๋Œ€๊ทœ๋ชจ ์›น ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ 12๋…„ ๋™์•ˆ ์‹คํ–‰๋˜์–ด ํŽ˜ํƒ€๋ฐ”์ดํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ธํ„ฐ๋„ท ์Šคํฌ๋žฉ์ธ CommonCrawl์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋ฉ๋‹ˆ๋‹ค. ์ธํ„ฐ๋„ท ์ „์ฒด์—์„œ ์Šคํฌ๋žฉํ•œ ๋ฐ์ดํ„ฐ๋กœ ์ž‘์—…ํ•˜๋Š” ๊ฒƒ์€ ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ์•ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ƒ๋‹น ๋ถ€๋ถ„์ด ๊ธฐ๊ณ„์—์„œ ์ƒ์„ฑํ•œ ์ŠคํŒธ์ด๋‚˜ ์Œ๋ž€ ์ฝ˜ํ…์ธ (Trinh and Le, 2018; Kreutzer et al., 2022)๋ผ๋Š” ์ ์ด ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•„ํ„ฐ๋ง๋˜์ง€ ์•Š์€ ์›น ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๊ฒฐ๊ณผ์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๋ชจ๋ธ์ด ๋งŒ๋“ค์–ด์ง‘๋‹ˆ๋‹ค(Raffel et al., 2020). ์ตœ์‹  ํŒŒ์ดํ”„๋ผ์ธ์€ ์ด๋Ÿฌํ•œ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค(Wenzek et al., 2020). ๋„“๊ฒŒ๋Š” ์ด๋Ÿฌํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„๋ฅผ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค: (1) _language identification_, ์ €๋ ดํ•œ n-๊ทธ๋žจ ๋ชจ๋ธ(์˜ˆ: fastText Joulin et al. (2016))์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ; (2) _filtering rules and heuristics_, ์˜ˆ๋ฅผ ๋“ค์–ด ์˜ฌ๋ฐ”๋ฅธ ๊ตฌ๋‘์ ์ด ์žˆ๋Š” ์ค„๋งŒ ์œ ์ง€ํ•˜๊ฑฐ๋‚˜, ๊ธฐํ˜ธ๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์€ ์ค„์€ ์‚ญ์ œํ•˜๊ฑฐ๋‚˜, ๊ธˆ์ง€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ(Grave et al., 2018; Raffel et al., 2020); (3) _ML-based quality filtering_, ์•Œ๋ ค์ง„ ๊ณจ๋“œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๊ฒฝ๋Ÿ‰ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ์‚ฌํ•œ ๊ณ ํ’ˆ์งˆ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค(Wenzek et al., 2020; Brown et al., 2020); (4) _deduplication_, ์ •ํ™•ํ•œ ์ค‘๋ณต ๋ฒ”์œ„ ๋˜๋Š” ์œ ์‚ฌํ•œ ๋ฌธ์„œ ์ œ๊ฑฐ(Lee et al., 2022). ์ผ๋ถ€ ํ•„ํ„ฐ๋ง์€ ํ•„์š”ํ•˜์ง€๋งŒ ๊ณผ๋„ํ•œ ํ•„ํ„ฐ๋ง์€ ๋ชจ๋ธ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํŽธํ–ฅ์„ฑ์„ ์œ ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์†Œ์ˆ˜์ž์—๊ฒŒ ์ง€๋‚˜์น˜๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Dodge et al., 2021), ํ—ˆ์šฉ๋œ URL์„ ์ˆ˜๋™์œผ๋กœ ํ๋ ˆ์ด์…˜ํ•˜๋Š” pseudo-crawling๊ณผ ๊ฐ™์€ ๊ด€ํ–‰์„ ์ฑ„ํƒํ•˜๋„๋ก ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค(Laurencon et al., 2022).
์›น ๋ฐ์ดํ„ฐ๋ฅผ ์œ„ํ•œ ํŒŒ์ดํ”„๋ผ์ธ.๋Œ€๊ทœ๋ชจ ์›น ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ 12๋…„ ๋™์•ˆ ์‹คํ–‰๋˜์–ด ํŽ˜ํƒ€๋ฐ”์ดํŠธ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ธํ„ฐ๋„ท ์Šคํฌ๋žฉ์ธ CommonCrawl์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ์ถ•๋ฉ๋‹ˆ๋‹ค. ์ธํ„ฐ๋„ท ์ „์ฒด์—์„œ ์Šคํฌ๋žฉํ•œ ๋ฐ์ดํ„ฐ๋กœ ์ž‘์—…ํ•˜๋Š” ๊ฒƒ์€ ๊ณ ์œ ํ•œ ๊ณผ์ œ๋ฅผ ์•ˆ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์ƒ๋‹น ๋ถ€๋ถ„์ด ๊ธฐ๊ณ„์—์„œ ์ƒ์„ฑํ•œ ์ŠคํŒธ์ด๋‚˜ ์Œ๋ž€ ์ฝ˜ํ…์ธ (Trinh and Le, 2018; Kreutzer et al., 2022)๋ผ๋Š” ์ ์ด ์ฃผ๋ชฉํ•  ๋งŒํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•„ํ„ฐ๋ง๋˜์ง€ ์•Š์€ ์›น ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์œผ๋ฉฐ, ๊ฒฐ๊ณผ์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๋ชจ๋ธ์ด ๋งŒ๋“ค์–ด์ง‘๋‹ˆ๋‹ค(Raffel et al., 2020). ์ตœ์‹  ํŒŒ์ดํ”„๋ผ์ธ์€ ์ด๋Ÿฌํ•œ ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ์ฝ˜ํ…์ธ ๋ฅผ ํ•„ํ„ฐ๋งํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค(Wenzek et al., 2020). ๋„“๊ฒŒ๋Š” ์ด๋Ÿฌํ•œ ํŒŒ์ดํ”„๋ผ์ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ๋‹จ๊ณ„๋ฅผ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค: (1) _์–ธ์–ด ์‹๋ณ„_, ์ €๋ ดํ•œ n-๊ทธ๋žจ ๋ชจ๋ธ(์˜ˆ: fastText Joulin et al. (2016))์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ; (2) _ํ•„ํ„ฐ๋ง ๊ทœ์น™ ๋ฐ ํœด๋ฆฌ์Šคํ‹ฑ_, ์˜ˆ๋ฅผ ๋“ค์–ด ์˜ฌ๋ฐ”๋ฅธ ๊ตฌ๋‘์ ์ด ์žˆ๋Š” ์ค„๋งŒ ์œ ์ง€ํ•˜๊ฑฐ๋‚˜, ๊ธฐํ˜ธ๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์€ ์ค„์€ ์‚ญ์ œํ•˜๊ฑฐ๋‚˜, ๊ธˆ์ง€ ๋‹จ์–ด๊ฐ€ ํฌํ•จ๋œ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ(Grave et al., 2018; Raffel et al., 2020); (3) _ML ๊ธฐ๋ฐ˜ ํ’ˆ์งˆ ํ•„ํ„ฐ๋ง_, ์•Œ๋ ค์ง„ ๊ณจ๋“œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๊ฒฝ๋Ÿ‰ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์œ ์‚ฌํ•œ ๊ณ ํ’ˆ์งˆ ์›น ๋ฌธ์„œ๋ฅผ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค(Wenzek et al., 2020; Brown et al., 2020); (4) _์ค‘๋ณต ์ œ๊ฑฐ_, ์ •ํ™•ํ•œ ์ค‘๋ณต ๋ฒ”์œ„ ๋˜๋Š” ์œ ์‚ฌํ•œ ๋ฌธ์„œ ์ œ๊ฑฐ(Lee et al., 2022). ์ผ๋ถ€ ํ•„ํ„ฐ๋ง์€ ํ•„์š”ํ•˜์ง€๋งŒ ๊ณผ๋„ํ•œ ํ•„ํ„ฐ๋ง์€ ๋ชจ๋ธ์— ๋ฐ”๋žŒ์งํ•˜์ง€ ์•Š์€ ํŽธํ–ฅ์„ฑ์„ ์œ ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ์†Œ์ˆ˜์ž์—๊ฒŒ ์ง€๋‚˜์น˜๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์œผ๋ฉฐ(Dodge et al., 2021), ํ—ˆ์šฉ๋œ URL์„ ์ˆ˜๋™์œผ๋กœ ํ๋ ˆ์ด์…˜ํ•˜๋Š” ์˜์‚ฌ ํฌ๋กค๋ง๊ณผ ๊ฐ™์€ ๊ด€ํ–‰์„ ์ฑ„ํƒํ•˜๋„๋ก ๋™๊ธฐ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค(Laurencon et al., 2022).
1. **No biases.** Following LLaMA, PaLM, and others, we exclude all bias terms from our architecture in order to improve training stability. 2. **Non-parametric layer norm.** We use the non-parametric formulation of layer norm (Ba et al., 2016) in which there is no affine transformation within the norm, i.e. no "adaptive gain" (or bias). We believe this was the safest option and it was also the fastest compared to the other variants we considered: parametric layer norm and RMSNorm (Zhang and Sennrich, 2019). 3. **SwiGLU activation function.** Like LLaMA, PaLM, and others we use the SwiGLU activation function (Shazeer, 2020) instead of ReLU, and following LLaMA the activation hidden size is approximately \(\frac{8}{3}d\), but increased to the closest multiple of 128 (e.g. 11,008 for our 7B model) to improve throughput.2 Footnote 2: Since SwiGLU is a โ€œgatedโ€ activation function, the output is half the size of the input. So technically our inputs to SwiGLU have a dimensionality of 2 \(\times\) 11,008 = 22,016 for our 7B model. 4. **Rotary positional embeddings (RoPE).** Like LLaMA, PaLM, and others we replace absolute positional embeddings with rotary positional embeddings (RoPE; Su et al., 2021). 5. **Vocabulary.** We use a modified version of the BPE-based tokenizer from GPT-NeoX-20B (Black et al., 2022) with additional tokens for masking personal identifiable information (PII). The final vocabulary size is 50,280. However, to maximize training throughput we increase the size of the corresponding embedding matrix in our model to 50,304 so that it's a multiple of 128.
1. **ํŽธํ–ฅ ์—†์Œ**. LLaMA, PaLM ๋“ฑ์— ๋”ฐ๋ผ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์•„ํ‚คํ…์ฒ˜์—์„œ ๋ชจ๋“  bias ํ•ญ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. 2. **๋น„๋ชจ์ˆ˜ layer norm**. ๋…ธ๋ฆ„ ๋‚ด์—์„œ ์•„ํ•€ ๋ณ€ํ™˜์ด ์—†๋Š” ๋น„๋ชจ์ˆ˜ ๊ณต์‹์ธ layer norm (Ba et al., 2016)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, "์ ์‘์  ์ด๋“"(๋˜๋Š” bias)์ด ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ๊ฐ€์žฅ ์•ˆ์ „ํ•œ ์˜ต์…˜์ด๋ผ๊ณ  ์ƒ๊ฐํ–ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋‹ค๋ฅธ ๋ณ€ํ˜•์ธ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ layer norm๊ณผ RMSNorm(Zhang and Sennrich, 2019)์— ๋น„ํ•ด ๊ฐ€์žฅ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ๋„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜(Shazeer, 2020)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, LLaMA์— ๋”ฐ๋ผ ํ™œ์„ฑํ™” ์ˆจ๊ฒจ์ง„ ํฌ๊ธฐ๋Š” ๋Œ€๋žต \(\frac{8}{3}d\)์ด์ง€๋งŒ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด 128์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ๋ฐฐ์ˆ˜๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: SwiGLU๋Š” "๊ฒŒ์ดํŠธ" ํ™œ์„ฑํ™” ํ•จ์ˆ˜์ด๋ฏ€๋กœ ์ถœ๋ ฅ์€ ์ž…๋ ฅ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์ˆ ์ ์œผ๋กœ SwiGLU์— ๋Œ€ํ•œ ์ž…๋ ฅ์˜ ์ฐจ์›์€ 2 \(\๋ฐฐ\) 11,008 = 22,016์ด๋ฉฐ, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 22,016์ž…๋‹ˆ๋‹ค. 4. **Rotary positional embeddings (RoPE)**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ˆ๋Œ€ positional embeddings์„ rotary positional embeddings (RoPE; Su et al., 2021)์œผ๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. 5. **์–ดํœ˜**. ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด(PII)๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ํ† ํฐ์ด ์žˆ๋Š” GPT-NeoX-20B(Black et al., 2022)์˜ BPE ๊ธฐ๋ฐ˜ tokenizer์˜ ์ˆ˜์ •๋œ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์–ดํœ˜ ํฌ๊ธฐ๋Š” 50,280์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์— ํ•ด๋‹น embedding ํ–‰๋ ฌ์˜ ํฌ๊ธฐ๋ฅผ 50,304๋กœ ๋Š˜๋ ค 128์˜ ๋ฐฐ์ˆ˜๊ฐ€ ๋˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
1. **ํŽธํ–ฅ ์—†์Œ**. LLaMA, PaLM ๋“ฑ์— ๋”ฐ๋ผ ํ›ˆ๋ จ ์•ˆ์ •์„ฑ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์•„ํ‚คํ…์ฒ˜์—์„œ ๋ชจ๋“  ๋ฐ”์ด์–ด์Šค ํ•ญ์„ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค. 2. **๋น„๋ชจ์ˆ˜ ๋ ˆ์ด์–ด ๊ทœ๋ฒ”**. ๋…ธ๋ฆ„ ๋‚ด์—์„œ ์•„ํ•€ ๋ณ€ํ™˜์ด ์—†๋Š” ๋น„๋ชจ์ˆ˜ ๊ณต์‹์ธ ๋ ˆ์ด์–ด ๋…ธ๋ฆ„(Ba et al., 2016)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, "์ ์‘์  ์ด๋“"(๋˜๋Š” ๋ฐ”์ด์–ด์Šค)์ด ์—†์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด๊ฒƒ์ด ๊ฐ€์žฅ ์•ˆ์ „ํ•œ ์˜ต์…˜์ด๋ผ๊ณ  ์ƒ๊ฐํ–ˆ์œผ๋ฉฐ, ์šฐ๋ฆฌ๊ฐ€ ๊ณ ๋ คํ•œ ๋‹ค๋ฅธ ๋ณ€ํ˜•์ธ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ ˆ์ด์–ด ๋…ธ๋ฆ„๊ณผ RMSNorm(Zhang and Sennrich, 2019)์— ๋น„ํ•ด ๊ฐ€์žฅ ๋น ๋ฅด๋‹ค๋Š” ๊ฒƒ๋„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. 3. **SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜**. LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ReLU ๋Œ€์‹  SwiGLU ํ™œ์„ฑํ™” ํ•จ์ˆ˜(Shazeer, 2020)๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, LLaMA์— ๋”ฐ๋ผ ํ™œ์„ฑํ™” ์ˆจ๊ฒจ์ง„ ํฌ๊ธฐ๋Š” ๋Œ€๋žต \(frac{8}{3}d\)์ด์ง€๋งŒ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด 128์˜ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ๋ฐฐ์ˆ˜๋กœ ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.2 ๊ฐ์ฃผ 2: SwiGLU๋Š” "๊ฒŒ์ดํŠธ" ํ™œ์„ฑํ™” ํ•จ์ˆ˜์ด๋ฏ€๋กœ ์ถœ๋ ฅ์€ ์ž…๋ ฅ์˜ ์ ˆ๋ฐ˜ ํฌ๊ธฐ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์ˆ ์ ์œผ๋กœ SwiGLU์— ๋Œ€ํ•œ ์ž…๋ ฅ์˜ ์ฐจ์›์€ 2 \(\๋ฐฐ\) 11,008 = 22,016์ด๋ฉฐ, 7B ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 22,016์ž…๋‹ˆ๋‹ค. 4. **๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ(RoPE).** LLaMA, PaLM ๋“ฑ๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ ˆ๋Œ€ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ์„ ๋กœํ„ฐ๋ฆฌ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ(RoPE; Su et al., 2021)์œผ๋กœ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. 5. **์–ดํœ˜.** ๊ฐœ์ธ ์‹๋ณ„ ์ •๋ณด(PII)๋ฅผ ๋งˆ์Šคํ‚นํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ํ† ํฐ์ด ์žˆ๋Š” GPT-NeoX-20B(Black et al., 2022)์˜ BPE ๊ธฐ๋ฐ˜ ํ† ํฐ๋ผ์ด์ €์˜ ์ˆ˜์ •๋œ ๋ฒ„์ „์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ตœ์ข… ์–ดํœ˜ ํฌ๊ธฐ๋Š” 50,280์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ›ˆ๋ จ ์ฒ˜๋ฆฌ๋Ÿ‰์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋ชจ๋ธ์— ํ•ด๋‹น ์ž„๋ฒ ๋”ฉ ํ–‰๋ ฌ์˜ ํฌ๊ธฐ๋ฅผ 50,304๋กœ ๋Š˜๋ ค 128์˜ ๋ฐฐ์ˆ˜๊ฐ€ ๋˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
Using GPT-NeoX (Andonian et al., 2021) based on Megatron-DeepSpeed (Shoeybi et al., 2019; Microsoft, 2020), we train autoregressive decoder-only transformers with a causal language modeling objective. The models use Pre-LN. Each model is trained using the same tokenizer as Black et al. (2022), which was trained exclusively on the Pile via the BPE algorithm (Sennrich et al., 2016). For all models, we train with the AdamW optimizer (Loshchilov and Hutter, 2019) using a batch size of 1104 and a sequence length of 2048. An epoch of training approximately corresponds to \(132,366\) total training steps. As mentioned in the previous section, we reset the optimizer states between datasets. We consider two model sizes 405M and 9.6B parameters (referred to as 10B in this work) including embeddings. We train the smaller models using data parallelism across 46 6 GPU nodes using a micro-batch size of 4. The larger model is trained using tensor parallelism(Shoeybi et al., 2020) spanning six GPUs within a node and pipeline parallelism(Huang et al., 2019) spanning four nodes; that is, each model replica spans 24 GPUs across four nodes. We train this model on 276 nodes using gradient accumulation of 4 steps. Each model uses optimizer sharding via ZeRO-1 (Rajbhandari et al., 2020), activation checkpointing (Chen et al., 2016), activation partitioning across tensor parallel ranks, and mixed precision FP16/FP32 to reduce GPU memory consumption and fully utilize NVIDIA tensor cores during training. We provided an extended description of all hyperparameters in the appendix (Table. 13).
GPT-NeoX(Andonian et al., 2021)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Megatron-DeepSpeed(Shoeybi et al., 2019; Microsoft, 2020)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” causal language modeling objective๋ฅผ ๊ฐ€์ง„ autoregressive decoder-only transformers๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ Pre-LN์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ์€ Black et al. (2022)๊ณผ ๋™์ผํ•œ tokenizer๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. ์ด tokenizer๋Š” BPE algorithm(Sennrich et al., 2016)์„ ํ†ตํ•ด Pile ์ „์šฉ์œผ๋กœ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•ด ์šฐ๋ฆฌ๋Š” AdamW optimizer(Loshchilov and Hutter, 2019)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ batch size 1104, sequence length 2048๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์˜ ํ•œ epoch๋Š” ๋Œ€๋žต \(132,366\)์˜ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์„น์…˜์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์— optimizer ์ƒํƒœ๋ฅผ ์žฌ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” embedding์„ ํฌํ•จํ•˜์—ฌ 405M ๋ฐ 9.6B parameters(์ด ์ž‘์—…์—์„œ๋Š” 10B๋ผ๊ณ  ํ•จ)์˜ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ์€ micro-batch size 4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 46 6 GPU ๋…ธ๋“œ์— ๊ฑธ์ณ data parallelism์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋” ํฐ ๋ชจ๋ธ์€ ๋…ธ๋“œ ๋‚ด 6๊ฐœ์˜ GPU์— ๊ฑธ์นœ tensor parallelism(Shoeybi et al., 2020)๊ณผ 4๊ฐœ์˜ ๋…ธ๋“œ์— ๊ฑธ์นœ pipeline parallelism(Huang et al., 2019)์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๊ฐ ๋ชจ๋ธ replica๋Š” 4๊ฐœ์˜ ๋…ธ๋“œ์— ๊ฑธ์ณ 24๊ฐœ์˜ GPU์— ๊ฑธ์ณ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ 276๊ฐœ์˜ ๋…ธ๋“œ์—์„œ gradient accumulation 4๋‹จ๊ณ„๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ์€ ํ›ˆ๋ จ ์ค‘์— GPU ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๋ฅผ ์ค„์ด๊ณ  NVIDIA tensor cores๋ฅผ ์™„์ „ํžˆ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ZeRO-1(Rajbhandari et al., 2020), activation checkpointing(Chen et al., 2016), tensor parallel ranks์— ๋”ฐ๋ฅธ activation partitioning, ํ˜ผํ•ฉ ์ •๋ฐ€๋„ FP16/FP32๋ฅผ ํ†ตํ•ด optimizer sharding์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  hyperparameters์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ๋ถ€๋ก(ํ‘œ 13)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
GPT-NeoX(Andonian et al., 2021)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. (์‡ผ์ด๋น„ ๋“ฑ, 2019; Microsoft, 2020)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ธ๊ณผ์  ์–ธ์–ด ๋ชจ๋ธ๋ง ๋ชฉํ‘œ๋ฅผ ๊ฐ€์ง„ ์ž๋™ ํšŒ๊ท€ ๋””์ฝ”๋” ์ „์šฉ ํŠธ๋žœ์Šคํฌ๋จธ๋ฅผ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์‚ฌ์ „-LN์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ์€ Black et al.๊ณผ ๋™์ผํ•œ ํ† ํฐํ™”๊ธฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. (2022), BPE ์•Œ๊ณ ๋ฆฌ์ฆ˜(Sennrich et al., 2016)์„ ํ†ตํ•ด ํŒŒ์ผ ์ „์šฉ์œผ๋กœ ํ›ˆ๋ จ๋œ (2022). ๋ชจ๋“  ๋ชจ๋ธ์— ๋Œ€ํ•ด AdamW ์˜ตํ‹ฐ๋งˆ์ด์ €(Loshchilov and Hutter, 2019)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐฐ์น˜ ํฌ๊ธฐ 1104, ์‹œํ€€์Šค ๊ธธ์ด 2048๋กœ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ํ›ˆ๋ จ์˜ ํ•œ ์—ํฌํฌ๋Š” ๋Œ€๋žต \(132,366\)์˜ ์ด ํ›ˆ๋ จ ๋‹จ๊ณ„์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ์ด์ „ ์„น์…˜์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์— ์˜ตํ‹ฐ๋งˆ์ด์ € ์ƒํƒœ๋ฅผ ์žฌ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์ž„๋ฒ ๋”ฉ์„ ํฌํ•จํ•˜์—ฌ 405M ๋ฐ 9.6B ํŒŒ๋ผ๋ฏธํ„ฐ(์ด ์ž‘์—…์—์„œ๋Š” 10B๋ผ๊ณ  ํ•จ)์˜ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ํฌ๊ธฐ๋ฅผ ๊ณ ๋ คํ•ฉ๋‹ˆ๋‹ค. ์ž‘์€ ๋ชจ๋ธ์€ ๋งˆ์ดํฌ๋กœ ๋ฐฐ์น˜ ํฌ๊ธฐ 4๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 46 6 GPU ๋…ธ๋“œ์— ๊ฑธ์ณ ๋ฐ์ดํ„ฐ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๋” ํฐ ๋ชจ๋ธ์€ ๋…ธ๋“œ ๋‚ด 6๊ฐœ์˜ GPU์— ๊ฑธ์นœ ํ…์„œ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ(Shoeybi et al., 2020)์™€ 4๊ฐœ์˜ ๋…ธ๋“œ์— ๊ฑธ์นœ ํŒŒ์ดํ”„๋ผ์ธ ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ(Huang et al., 2019)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๊ฐ ๋ชจ๋ธ ๋ณต์ œ๋ณธ์€ 4๊ฐœ์˜ ๋…ธ๋“œ์— ๊ฑธ์ณ 24๊ฐœ์˜ GPU์— ๊ฑธ์ณ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ 276๊ฐœ์˜ ๋…ธ๋“œ์—์„œ 4๋‹จ๊ณ„์˜ ๊ทธ๋ผ๋””์–ธํŠธ ๋ˆ„์ ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ฐ ๋ชจ๋ธ์€ ํ›ˆ๋ จ ์ค‘์— GPU ๋ฉ”๋ชจ๋ฆฌ ์†Œ๋น„๋ฅผ ์ค„์ด๊ณ  NVIDIA ํ…์„œ ์ฝ”์–ด๋ฅผ ์™„์ „ํžˆ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•ด ZeRO-1(Rajbhandari et al., 2020), ํ™œ์„ฑํ™” ์ฒดํฌํฌ์ธํŒ…(Chen et al., 2016), ํ…์„œ ๋ณ‘๋ ฌ ์ˆœ์œ„์— ๋”ฐ๋ฅธ ํ™œ์„ฑํ™” ํŒŒํ‹ฐ์…”๋‹, ํ˜ผํ•ฉ ์ •๋ฐ€๋„ FP16/FP32๋ฅผ ํ†ตํ•ด ์˜ตํ‹ฐ๋งˆ์ด์ € ์ƒค๋”ฉ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์„ค๋ช…์€ ๋ถ€๋ก(ํ‘œ 13)์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_LLM based generative approaches._ Reward models help to select aligned data from model responses. However, training reward models itself necessitates substantial high-quality human-labeled data, which is typically expensive and in short supply. In addition, although existing reward models can be reused, they might not be able to accurately capture the nonalignment behaviors in another separately trained LLM. Therefore, some work explores leveraging powerful LLMs to automatically generate human-aligned data. As a representative work, constitutional AI [371] proposes that human supervision comes from a set of principles (_i.e._, natural language instructions) governing AI behaviors. Based on these principles, LLMs will critique their own harmful responses and revise them repeatedly into finally aligned responses. Similarly, Self-Align [347] first adopts self-instruct [143] to generate instructions focusing on covering diverse topics. Then, the model is also prompted with multiple human-written principles that describe the rules of expected model behaviors (also with several in-context exemplars), to generate helpful, ethical, and reliable responses as alignment data. To mitigate the limit that the original SFT method can only learn from positive responses, FIGA [386] develops an improved supervised alignment approach, where both negative (the original output of low quality) and positive (the refined output by LLMs) responses are leveraged in a contrastive way, to enable LLMs to deeply understand what fine-grained revisions actually lead to good response.
\(\bullet\)_LLM ๊ธฐ๋ฐ˜ ์ƒ์„ฑ ์ ‘๊ทผ ๋ฐฉ์‹._ Reward models์€ ๋ชจ๋ธ ์‘๋‹ต์—์„œ ์ •๋ ฌ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Reward models ์ž์ฒด๋ฅผ ํ›ˆ๋ จํ•˜๋ ค๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ๊ณต๊ธ‰์ด ๋ถ€์กฑํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์ธ๊ฐ„ ๋ผ๋ฒจ๋ง ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด Reward models์„ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ณ„๋„๋กœ ํ›ˆ๋ จ๋œ ๋‹ค๋ฅธ LLM์—์„œ ์ •๋ ฌ๋˜์ง€ ์•Š์€ ํ–‰๋™์„ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ•๋ ฅํ•œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ์ด ์ •๋ ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, ํ—Œ๋ฒ• AI[371]๋Š” AI ํ–‰๋™์„ ์ง€๋ฐฐํ•˜๋Š” ์ผ๋ จ์˜ ์›์น™(์ฆ‰, ์ž์—ฐ์–ด Instructions)์—์„œ ์ธ๊ฐ„์˜ ๊ฐ๋…์ด ์ด๋ฃจ์–ด์ง„๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์›์น™์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์€ ์ž์‹ ์˜ ์œ ํ•ดํ•œ ์‘๋‹ต์„ ๋น„ํŒํ•˜๊ณ  ์ด๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ ์ตœ์ข…์ ์œผ๋กœ ์ •๋ ฌ๋œ ์‘๋‹ต์œผ๋กœ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Self-Align [347]์€ ๋จผ์ € ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘” Instructions์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Self-Instruct [143]๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๋ชจ๋ธ์— ์˜ˆ์ƒ๋˜๋Š” ๋ชจ๋ธ ํ–‰๋™์˜ ๊ทœ์น™์„ ์„ค๋ช…ํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์›์น™(์—ฌ๊ธฐ์—๋„ ๋ช‡ ๊ฐ€์ง€ In-Context ์˜ˆ์‹œ ํฌํ•จ)์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ ์œ ์šฉํ•˜๊ณ  ์œค๋ฆฌ์ ์ด๋ฉฐ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ Supervised Fine-tuning (SFT) ๋ฐฉ์‹์€ ๊ธ์ •์ ์ธ ์‘๋‹ต๋งŒ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ํ•œ๊ณ„๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด FIGA [386]๋Š” ๋ถ€์ •์ ์ธ ์‘๋‹ต(ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์›๋ž˜ ์ถœ๋ ฅ)๊ณผ ๊ธ์ •์ ์ธ ์‘๋‹ต(LLM์— ์˜ํ•œ ์ˆ˜์ •๋œ ์ถœ๋ ฅ)์„ ๋ชจ๋‘ ๋Œ€์กฐ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ LLM์ด ์‹ค์ œ๋กœ ์–ด๋–ค ์„ธ๋ถ„ํ™”๋œ ์ˆ˜์ •์ด ์ข‹์€ ์‘๋‹ต์œผ๋กœ ์ด์–ด์ง€๋Š”์ง€ ๊นŠ์ด ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐœ์„ ๋œ ์ง€๋„ ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_LLM ๊ธฐ๋ฐ˜ ์ƒ์„ฑ ์ ‘๊ทผ ๋ฐฉ์‹._ ๋ณด์ƒ ๋ชจ๋ธ์€ ๋ชจ๋ธ ์‘๋‹ต์—์„œ ์ •๋ ฌ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋ณด์ƒ ๋ชจ๋ธ ์ž์ฒด๋ฅผ ํ›ˆ๋ จํ•˜๋ ค๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค๊ณ  ๊ณต๊ธ‰์ด ๋ถ€์กฑํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์ธ๊ฐ„ ๋ผ๋ฒจ๋ง ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด ๋ณด์ƒ ๋ชจ๋ธ์„ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ณ„๋„๋กœ ํ›ˆ๋ จ๋œ ๋‹ค๋ฅธ LLM์—์„œ ์ •๋ ฌ๋˜์ง€ ์•Š์€ ํ–‰๋™์„ ์ •ํ™•ํ•˜๊ฒŒ ํฌ์ฐฉํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ถ€ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ•๋ ฅํ•œ LLM์„ ํ™œ์šฉํ•˜์—ฌ ์‚ฌ๋žŒ์ด ์ •๋ ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ž๋™์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ชจ์ƒ‰ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์—ฐ๊ตฌ๋กœ, ํ—Œ๋ฒ• AI[371]๋Š” AI ํ–‰๋™์„ ์ง€๋ฐฐํ•˜๋Š” ์ผ๋ จ์˜ ์›์น™(์ฆ‰, ์ž์—ฐ์–ด ์ง€์นจ)์—์„œ ์ธ๊ฐ„์˜ ๊ฐ๋…์ด ์ด๋ฃจ์–ด์ง„๋‹ค๊ณ  ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์›์น™์„ ๋ฐ”ํƒ•์œผ๋กœ LLM์€ ์ž์‹ ์˜ ์œ ํ•ดํ•œ ์‘๋‹ต์„ ๋น„ํŒํ•˜๊ณ  ์ด๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ ์ตœ์ข…์ ์œผ๋กœ ์ •๋ ฌ๋œ ์‘๋‹ต์œผ๋กœ ์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์…€ํ”„ ์–ผ๋ผ์ธ [347]์€ ๋จผ์ € ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ๋Š” ๋ฐ ์ค‘์ ์„ ๋‘” ์ง€์นจ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด ์…€ํ”„ ์ธ์ŠคํŠธ๋ŸญํŠธ [143]๋ฅผ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ๋ชจ๋ธ์— ์˜ˆ์ƒ๋˜๋Š” ๋ชจ๋ธ ํ–‰๋™์˜ ๊ทœ์น™์„ ์„ค๋ช…ํ•˜๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์›์น™(์—ฌ๊ธฐ์—๋„ ๋ช‡ ๊ฐ€์ง€ ์ปจํ…์ŠคํŠธ ๋‚ด ์˜ˆ์‹œ ํฌํ•จ)์„ ํ”„๋กฌํ”„ํŠธํ•˜์—ฌ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ ์œ ์šฉํ•˜๊ณ  ์œค๋ฆฌ์ ์ด๋ฉฐ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์›๋ž˜ SFT ๋ฐฉ์‹์€ ๊ธ์ •์ ์ธ ์‘๋‹ต๋งŒ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ํ•œ๊ณ„๋ฅผ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด FIGA [386]๋Š” ๋ถ€์ •์ ์ธ ์‘๋‹ต(ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์›๋ž˜ ์ถœ๋ ฅ)๊ณผ ๊ธ์ •์ ์ธ ์‘๋‹ต(LLM์— ์˜ํ•œ ์ˆ˜์ •๋œ ์ถœ๋ ฅ)์„ ๋ชจ๋‘ ๋Œ€์กฐ์ ์ธ ๋ฐฉ์‹์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ LLM์ด ์‹ค์ œ๋กœ ์–ด๋–ค ์„ธ๋ถ„ํ™”๋œ ์ˆ˜์ •์ด ์ข‹์€ ์‘๋‹ต์œผ๋กœ ์ด์–ด์ง€๋Š”์ง€ ๊นŠ์ด ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ฐœ์„ ๋œ ์ง€๋„ ์ •๋ ฌ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.
Large language model datasets.While we believe our results are independent of model architecture, we perform our analysis on Transformer-based decoder-only language models (Vaswani et al., 2017) trained for open-ended text generation. These current state-of-the-art models are trained on internet text. For example, the GPT-2 family of models Radford et al. (2019) is trained on WebText, a dataset of web documents highly ranked on Reddit--however this dataset was not made available publicly. A common dataset starting point is CommonCrawl, an index of public webpages. Among the models trained on CommonCrawl include GPT-3 (Brown et al., 2020) with the addition of book datasets, GROVER (Zellers et al., 2019) on a restricted subset filtered to news domains called RealNews, and T5 (Raffel et al., 2020) on a cleaned version of common crawl called C4. Other models are trained on more curated Internet sources--for example Guo et al. (2020) used high quality processed Wikipedia text from 40 different languages to train monolingual 141.4M parameter language models. Non-English models necessarily use different datasets; Zeng et al. (2021) for instance introduced PANGU-\(\alpha\), a family of models with up to 200B parameters that were trained on a non-public corpus of cleaned and filtered Chinese-language documents from CommonCrawl and other sources. Since many of these datasets are not public, we deduplicate three that are: Wiki-40B, C4, and RealNews-as well as the One Billion Word Language Model Benchmark (Chelba et al., 2013), a smaller dataset commonly used for evaluation.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ. ์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋ฌด๊ด€ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€๋งŒ, Transformer ๊ธฐ๋ฐ˜์˜ decoder ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ(Vaswani et al., 2017)์— ๋Œ€ํ•ด ๊ฐœ๋ฐฉํ˜• ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ตœ์‹  ๋ชจ๋ธ์€ ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ์ œํ’ˆ๊ตฐ(Radford et al., 2019)์€ Reddit์—์„œ ๋†’์€ ์ˆœ์œ„๋ฅผ ๋ฐ›์€ ์›น ๋ฌธ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ WebText๋กœ ํ•™์Šต๋˜์ง€๋งŒ, ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์‹œ์ž‘์ ์€ ๊ณต๊ฐœ ์›น ํŽ˜์ด์ง€์˜ ์ƒ‰์ธ์ธ CommonCrawl์ž…๋‹ˆ๋‹ค. CommonCrawl์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์—๋Š” ์ฑ… ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ถ”๊ฐ€๋œ GPT-3(Brown et al., 2020), ๋‰ด์Šค ๋„๋ฉ”์ธ์œผ๋กœ ํ•„ํ„ฐ๋ง๋œ ์ œํ•œ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ RealNews์— ๋Œ€ํ•œ GROVER(Zellers et al., 2019), C4๋ผ๋Š” CommonCrawl์˜ ์ •๋ฆฌ๋œ ๋ฒ„์ „์— ๋Œ€ํ•œ T5(Raffel et al., 2020) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋” ์—„์„ ๋œ ์ธํ„ฐ๋„ท ์†Œ์Šค๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Guo et al. (2020)์€ 40๊ฐœ ์–ธ์–ด์˜ ๊ณ ํ’ˆ์งˆ ์ฒ˜๋ฆฌ๋œ Wikipedia ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด 141.4M ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‹จ์–ด๋กœ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋Š”๋ฐ, ์˜ˆ๋ฅผ ๋“ค์–ด Zeng et al. (2021)์€ CommonCrawl ๋ฐ ๊ธฐํƒ€ ์†Œ์Šค์—์„œ ์ •๋ฆฌ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ์ค‘๊ตญ์–ด ๋ฌธ์„œ์˜ ๋น„๊ณต๊ฐœ ์ฝ”ํผ์Šค๋กœ ํ•™์Šตํ•œ ์ตœ๋Œ€ 200B ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ธ PANGU-\(\alpha\)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— Wiki-40B, C4, RealNews, ๊ทธ๋ฆฌ๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๋Š” ์†Œ๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ One Billion Word Language Model Benchmark(Chelba et al., 2013) ๋“ฑ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ.์šฐ๋ฆฌ๋Š” ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๊ฐ€ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์™€ ๋ฌด๊ด€ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•˜์ง€๋งŒ, ํŠธ๋žœ์Šคํฌ๋จธ ๊ธฐ๋ฐ˜์˜ ๋””์ฝ”๋” ์ „์šฉ ์–ธ์–ด ๋ชจ๋ธ(Vaswani et al., 2017)์— ๋Œ€ํ•ด ๊ฐœ๋ฐฉํ˜• ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•ด ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ ์ตœ์‹  ๋ชจ๋ธ์€ ์ธํ„ฐ๋„ท ํ…์ŠคํŠธ๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-2 ์ œํ’ˆ๊ตฐ(Radford et al., 2019)์€ Reddit์—์„œ ๋†’์€ ์ˆœ์œ„๋ฅผ ๋ฐ›์€ ์›น ๋ฌธ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ WebText๋กœ ํ•™์Šต๋˜์ง€๋งŒ, ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์‹œ์ž‘์ ์€ ๊ณต๊ฐœ ์›น ํŽ˜์ด์ง€์˜ ์ƒ‰์ธ์ธ CommonCrawl์ž…๋‹ˆ๋‹ค. CommonCrawl์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์—๋Š” ์ฑ… ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ์ถ”๊ฐ€๋œ GPT-3(Brown et al., 2020), ๋‰ด์Šค ๋„๋ฉ”์ธ์œผ๋กœ ํ•„ํ„ฐ๋ง๋œ ์ œํ•œ๋œ ํ•˜์œ„ ์ง‘ํ•ฉ์ธ RealNews์— ๋Œ€ํ•œ GROVER(Zellers et al., 2019), C4๋ผ๋Š” ๊ณตํ†ต ํฌ๋กค๋ง์˜ ์ •๋ฆฌ๋œ ๋ฒ„์ „์— ๋Œ€ํ•œ T5(Raffel et al., 2020) ๋“ฑ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ๋ชจ๋ธ์€ ๋” ์—„์„ ๋œ ์ธํ„ฐ๋„ท ์†Œ์Šค๋กœ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Guo et al. (2020)์€ 40๊ฐœ ์–ธ์–ด์˜ ๊ณ ํ’ˆ์งˆ ์ฒ˜๋ฆฌ๋œ ์œ„ํ‚ค๋ฐฑ๊ณผ ํ…์ŠคํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์–ด 141.4M ํŒŒ๋ผ๋ฏธํ„ฐ ์–ธ์–ด ๋ชจ๋ธ์„ ๋‹จ์–ด๋กœ ํ•™์Šตํ–ˆ์Šต๋‹ˆ๋‹ค. ๋น„์˜์–ด ๋ชจ๋ธ์€ ๋ฐ˜๋“œ์‹œ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋Š”๋ฐ, ์˜ˆ๋ฅผ ๋“ค์–ด Zeng et al. (2021)์€ ๊ณตํ†ต ํฌ๋กค๋ง ๋ฐ ๊ธฐํƒ€ ์†Œ์Šค์—์„œ ์ •๋ฆฌ ๋ฐ ํ•„ํ„ฐ๋ง๋œ ์ค‘๊ตญ์–ด ๋ฌธ์„œ์˜ ๋น„๊ณต๊ฐœ ์ฝ”ํผ์Šค๋กœ ํ•™์Šตํ•œ ์ตœ๋Œ€ 200B ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์ธ PANGU-\(\์•ŒํŒŒ\)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ๊ณต๊ฐœ๋˜์ง€ ์•Š์•˜๊ธฐ ๋•Œ๋ฌธ์— ์œ„ํ‚ค-40B, C4, ๋ฆฌ์–ผ๋‰ด์Šค, ๊ทธ๋ฆฌ๊ณ  ์ผ๋ฐ˜์ ์œผ๋กœ ํ‰๊ฐ€์— ์‚ฌ์šฉ๋˜๋Š” ์†Œ๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ 10์–ต ๋‹จ์–ด ์–ธ์–ด ๋ชจ๋ธ ๋ฒค์น˜๋งˆํฌ(Chelba et al., 2013) ๋“ฑ ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค‘๋ณต ์ œ๊ฑฐํ–ˆ์Šต๋‹ˆ๋‹ค.
* We introduced AMIE (Articulate Medical Intelligence Explorer), an LLM based AI system optimized for clinical history-taking and diagnostic dialogue. * To scale AMIE across a multitude of specialties and scenarios, we developed a novel self-play based simulated diagnostic dialogue environment with automated feedback mechanisms to enrich and accelerate its learning process. We also introduced an inference time chain-of-reasoning strategy to improve AMIE's diagnostic accuracy and conversation quality. * We developed a pilot evaluation rubric to assess the history-taking, diagnostic reasoning, communication skills and empathy of diagnostic conversational medical AI, encompassing both clinician-centred and patient-centred metrics. * We designed and conducted a blinded remote OSCE study with 149 case scenarios from clinical providers in Canada, the UK, and India, enabling randomized and counterbalanced comparison of AMIE to PCPs when performing consultations with validated patient actors. AMIE exhibited superior diagnostic accuracy compared to PCPs as assessed by various measures (e.g., top-1 and top-3 accuracy of the differential diagnosis list). Across 28 out of 32 evaluation axes from the specialist physician perspective and 24 out of 26 evaluation axes from the patient actor perspective, AMIE was rated superior to PCPs while being non-inferior on the rest. * We performed a range of ablations to further understand and characterize the capabilities of AMIE, highlighted important limitations, and proposed key next steps for real-world clinical translation of AMIE.
* ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™”๋œ LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์ธ AMIE(Articulate Medical Intelligence Explorer)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. * ๋‹ค์–‘ํ•œ ์ „๋ฌธ ๋ถ„์•ผ์™€ ์‹œ๋‚˜๋ฆฌ์˜ค์— AMIE๋ฅผ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ์ƒˆ๋กœ์šด self-play ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์ง„๋‹จ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ๊ฐœ๋ฐœํ•˜์—ฌ ํ•™์Šต ๊ณผ์ •์„ ๋ณด๊ฐ•ํ•˜๊ณ  ๊ฐ€์†ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ inference time chain-of-reasoning ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ AMIE์˜ ์ง„๋‹จ ์ •ํ™•๋„์™€ ๋Œ€ํ™” ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. * ์ž„์ƒ์˜ ์ค‘์‹ฌ ๋ฐ ํ™˜์ž ์ค‘์‹ฌ ์ง€ํ‘œ๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ์ง„๋‹จ ๋Œ€ํ™” ์˜๋ฃŒ AI์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ, ์ง„๋‹จ ์ถ”๋ก , ์˜์‚ฌ์†Œํ†ต ๊ธฐ์ˆ  ๋ฐ ๊ณต๊ฐ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํŒŒ์ผ๋Ÿฟ ํ‰๊ฐ€ ๋ฃจ๋ธŒ๋ฆญ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. * ์บ๋‚˜๋‹ค, ์˜๊ตญ, ์ธ๋„์˜ ์ž„์ƒ ์˜๋ฃŒ์ง„์œผ๋กœ๋ถ€ํ„ฐ 149๊ฐœ์˜ ์‚ฌ๋ก€ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ด ๋งน๊ฒ€ ์›๊ฒฉ OSCE ์—ฐ๊ตฌ๋ฅผ ์„ค๊ณ„ ๋ฐ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ฒ€์ฆ๋œ ํ™˜์ž ๋ฐฐ์šฐ์™€์˜ ์ƒ๋‹ด์„ ์ˆ˜ํ–‰ํ•  ๋•Œ AMIE์™€ PCP๋ฅผ ๋ฌด์ž‘์œ„๋กœ ๋น„๊ต ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ธก์ • ๊ธฐ์ค€(์˜ˆ: ๊ฐ๋ณ„ ์ง„๋‹จ ๋ชฉ๋ก์˜ ์ƒ์œ„ 1 ๋ฐ ์ƒ์œ„ 3 ์ •ํ™•๋„)์œผ๋กœ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, AMIE๋Š” PCP์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ „๋ฌธ์˜ ๊ด€์ ์—์„œ 32๊ฐœ ํ‰๊ฐ€ ์ถ• ์ค‘ 28๊ฐœ, ํ™˜์ž ํ–‰์œ„์ž ๊ด€์ ์—์„œ 26๊ฐœ ํ‰๊ฐ€ ์ถ• ์ค‘ 24๊ฐœ์—์„œ AMIE๊ฐ€ PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ํ‰๊ฐ€ ์ถ•์—์„œ๋Š” ๋น„์—ด๋“ฑํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. * AMIE์˜ ๊ธฐ๋Šฅ์„ ๋” ๊นŠ์ด ์ดํ•ดํ•˜๊ณ  ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์ ˆ์ œ์ˆ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ค‘์š”ํ•œ ํ•œ๊ณ„๋ฅผ ๊ฐ•์กฐํ•˜๊ณ  AMIE์˜ ์‹ค์ œ ์ž„์ƒ ์ ์šฉ์„ ์œ„ํ•œ ์ฃผ์š” ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.
* ์ž„์ƒ ๋ณ‘๋ ฅ ์ฒญ์ทจ ๋ฐ ์ง„๋‹จ ๋Œ€ํ™”์— ์ตœ์ ํ™”๋œ LLM ๊ธฐ๋ฐ˜ AI ์‹œ์Šคํ…œ์ธ AMIE(Articulate Medical Intelligence Explorer)๋ฅผ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. * ๋‹ค์–‘ํ•œ ์ „๋ฌธ ๋ถ„์•ผ์™€ ์‹œ๋‚˜๋ฆฌ์˜ค์— AMIE๋ฅผ ํ™•์žฅํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ ํ”ผ๋“œ๋ฐฑ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐ–์ถ˜ ์ƒˆ๋กœ์šด ์…€ํ”„ ํ”Œ๋ ˆ์ด ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์ง„๋‹จ ๋Œ€ํ™” ํ™˜๊ฒฝ์„ ๊ฐœ๋ฐœํ•˜์—ฌ ํ•™์Šต ๊ณผ์ •์„ ๋ณด๊ฐ•ํ•˜๊ณ  ๊ฐ€์†ํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”๋ก  ์‹œ๊ฐ„ ์ฒด์ธ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ AMIE์˜ ์ง„๋‹จ ์ •ํ™•๋„์™€ ๋Œ€ํ™” ํ’ˆ์งˆ์„ ๊ฐœ์„ ํ–ˆ์Šต๋‹ˆ๋‹ค. * ์ž„์ƒ์˜ ์ค‘์‹ฌ ๋ฐ ํ™˜์ž ์ค‘์‹ฌ ์ง€ํ‘œ๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ์ง„๋‹จ ๋Œ€ํ™” ์˜๋ฃŒ AI์˜ ๋ณ‘๋ ฅ ์ฒญ์ทจ, ์ง„๋‹จ ์ถ”๋ก , ์˜์‚ฌ์†Œํ†ต ๊ธฐ์ˆ  ๋ฐ ๊ณต๊ฐ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ํŒŒ์ผ๋Ÿฟ ํ‰๊ฐ€ ๋ฃจ๋ธŒ๋ฆญ์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. * ์บ๋‚˜๋‹ค, ์˜๊ตญ, ์ธ๋„์˜ ์ž„์ƒ ์˜๋ฃŒ์ง„์œผ๋กœ๋ถ€ํ„ฐ 149๊ฐœ์˜ ์‚ฌ๋ก€ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ํ†ตํ•ด ๋งน๊ฒ€ ์›๊ฒฉ OSCE ์—ฐ๊ตฌ๋ฅผ ์„ค๊ณ„ ๋ฐ ์ˆ˜ํ–‰ํ•˜์—ฌ ๊ฒ€์ฆ๋œ ํ™˜์ž ๋ฐฐ์šฐ์™€์˜ ์ƒ๋‹ด์„ ์ˆ˜ํ–‰ํ•  ๋•Œ AMIE์™€ PCP๋ฅผ ๋ฌด์ž‘์œ„๋กœ ๋น„๊ต ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์–‘ํ•œ ์ธก์ • ๊ธฐ์ค€(์˜ˆ: ๊ฐ๋ณ„ ์ง„๋‹จ ๋ชฉ๋ก์˜ ์ƒ์œ„ 1 ๋ฐ ์ƒ์œ„ 3 ์ •ํ™•๋„)์œผ๋กœ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, AMIE๋Š” PCP์— ๋น„ํ•ด ์šฐ์ˆ˜ํ•œ ์ง„๋‹จ ์ •ํ™•๋„๋ฅผ ๋ณด์˜€์Šต๋‹ˆ๋‹ค. ์ „๋ฌธ์˜ ๊ด€์ ์—์„œ 32๊ฐœ ํ‰๊ฐ€ ์ถ• ์ค‘ 28๊ฐœ, ํ™˜์ž ํ–‰์œ„์ž ๊ด€์ ์—์„œ 26๊ฐœ ํ‰๊ฐ€ ์ถ• ์ค‘ 24๊ฐœ์—์„œ AMIE๊ฐ€ PCP๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ ํ‰๊ฐ€ ์ถ•์—์„œ๋Š” ๋น„์—ด๋“ฑํ•œ ๊ฒƒ์œผ๋กœ ํ‰๊ฐ€๋˜์—ˆ์Šต๋‹ˆ๋‹ค. * AMIE์˜ ๊ธฐ๋Šฅ์„ ๋” ๊นŠ์ด ์ดํ•ดํ•˜๊ณ  ํŠน์„ฑํ™”ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ์ ˆ์ œ์ˆ ์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ค‘์š”ํ•œ ํ•œ๊ณ„๋ฅผ ๊ฐ•์กฐํ•˜๊ณ  AMIE์˜ ์‹ค์ œ ์ž„์ƒ ์ ์šฉ์„ ์œ„ํ•œ ์ฃผ์š” ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค.
**Robotics & Embodied Artificial Intelligence**: Recent works have developed more efficient reinforcement learning agents for robotics and embodied artificial intelligence [16; 140; 132; 133; 134; 135; 141; 136; 137; 138; 139; 140; 142; 143]. The focus is on enhancing autonomous agents' abilities for planning, reasoning, and collaboration in embodied environments. In specific, [140] proposes a unified agent system for embodied reasoning and task planning. In this system, the authors design high-level commands to enable improved planning while propose low-level controllers to translate commands into actions. Additionally, one can leverage dialogues to gather information [141] to accelerate the optimization process. [142; 143] employ autonomous agents for embodied decision-making and exploration. To overcome the physical constraints, the agents can generate executable plans and accomplish long-term tasks by leveraging multiple skills. In terms of control policies, SayCan [78] focuses on investigating a wide range of manipulation and navigation skills utilizing a mobile manipulator robot. Taking inspiration from typical tasks encountered in a kitchen environment, it presents a comprehensive set of 551 skills that cover seven skill families and 17 objects. These skills encompass various actions such as picking, placing, pouring, grasping, and manipulating objects, among others. TidyBot [137] is an embodied agent designed to personalize household cleanup tasks. It can learn users' preferences on object placement and manipulation methods through textual examples.
**๋กœ๋ด‡ ๊ณตํ•™ ๋ฐ ๊ตฌํ˜„๋œ ์ธ๊ณต ์ง€๋Šฅ**: ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋กœ๋ด‡ ๊ณตํ•™ ๋ฐ ๊ตฌํ˜„๋œ ์ธ๊ณต ์ง€๋Šฅ์„ ์œ„ํ•œ ๋ณด๋‹ค ํšจ์œจ์ ์ธ Reinforcement Learning ์—์ด์ „ํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค[16; 140; 132; 133; 134; 135; 141; 136; 137; 138; 139; 140; 142; 143]. ์ด๋“ค์˜ ์ดˆ์ ์€ ๊ตฌํ˜„๋œ ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš, ์ถ”๋ก  ๋ฐ ํ˜‘์—…์„ ์œ„ํ•œ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ [140]์€ ๊ตฌํ˜„๋œ ์ถ”๋ก ๊ณผ ์ž‘์—… ๊ณ„ํš์„ ์œ„ํ•œ ํ†ตํ•ฉ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹œ์Šคํ…œ์—์„œ ์ €์ž๋“ค์€ ๊ฐœ์„ ๋œ ๊ณ„ํš์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ์ƒ์œ„ ์ˆ˜์ค€์˜ ๋ช…๋ น์„ ์„ค๊ณ„ํ•˜๊ณ  ๋ช…๋ น์„ ํ–‰๋™์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํ•˜์œ„ ์ˆ˜์ค€์˜ ์ปจํŠธ๋กค๋Ÿฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ [141] ์ตœ์ ํ™” ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€์†ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [142; 143]์€ ๊ตฌํ˜„๋œ ์˜์‚ฌ ๊ฒฐ์ • ๋ฐ ํƒ์ƒ‰์„ ์œ„ํ•ด ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฌผ๋ฆฌ์  ์ œ์•ฝ์„ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ๋Š” ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  ์—ฌ๋Ÿฌ ์Šคํ‚ฌ์„ ํ™œ์šฉํ•˜์—ฌ ์žฅ๊ธฐ์ ์ธ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ œ์–ด ์ •์ฑ… ์ธก๋ฉด์—์„œ SayCan [78]์€ ์ด๋™์‹ ์กฐ์ž‘ ๋กœ๋ด‡์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์กฐ์ž‘ ๋ฐ ํƒ์ƒ‰ ๊ธฐ์ˆ ์„ ์กฐ์‚ฌํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ฃผ๋ฐฉ ํ™˜๊ฒฝ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์ž‘์—…์—์„œ ์˜๊ฐ์„ ์–ป์–ด 7๊ฐ€์ง€ ์Šคํ‚ฌ๊ตฐ๊ณผ 17๊ฐ€์ง€ ๋ฌผ์ฒด๋ฅผ ํฌ๊ด„ํ•˜๋Š” 551๊ฐ€์ง€ ์Šคํ‚ฌ ์„ธํŠธ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์Šคํ‚ฌ์—๋Š” ์ง‘๊ธฐ, ๋†“๊ธฐ, ๋”ฐ๋ฅด๊ธฐ, ์žก๊ธฐ, ์กฐ์ž‘ ๋“ฑ ๋‹ค์–‘ํ•œ ๋™์ž‘์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. TidyBot [137]์€ ๊ฐ€์ • ๋‚ด ์ฒญ์†Œ ์ž‘์—…์„ ๊ฐœ์ธํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋œ ๊ตฌํ˜„ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ํ…์ŠคํŠธ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ๋ฌผ์ฒด ๋ฐฐ์น˜ ๋ฐ ์กฐ์ž‘ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์‚ฌ์šฉ์ž์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
**๋กœ๋ด‡ ๊ณตํ•™ ๋ฐ ๊ตฌํ˜„๋œ ์ธ๊ณต ์ง€๋Šฅ**: ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ๋กœ๋ด‡ ๊ณตํ•™ ๋ฐ ๊ตฌํ˜„๋œ ์ธ๊ณต ์ง€๋Šฅ์„ ์œ„ํ•œ ๋ณด๋‹ค ํšจ์œจ์ ์ธ ๊ฐ•ํ™” ํ•™์Šต ์—์ด์ „ํŠธ๋ฅผ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค[16; 140; 132; 133; 134; 135; 141; 136; 137; 138; 139; 140; 142; 143]. ์ด๋“ค์˜ ์ดˆ์ ์€ ๊ตฌํ˜„๋œ ํ™˜๊ฒฝ์—์„œ ๊ณ„ํš, ์ถ”๋ก  ๋ฐ ํ˜‘์—…์„ ์œ„ํ•œ ์ž์œจ ์—์ด์ „ํŠธ์˜ ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํŠนํžˆ [140]์€ ๊ตฌํ˜„๋œ ์ถ”๋ก ๊ณผ ์ž‘์—… ๊ณ„ํš์„ ์œ„ํ•œ ํ†ตํ•ฉ ์—์ด์ „ํŠธ ์‹œ์Šคํ…œ์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹œ์Šคํ…œ์—์„œ ์ €์ž๋“ค์€ ๊ฐœ์„ ๋œ ๊ณ„ํš์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜๋Š” ์ƒ์œ„ ์ˆ˜์ค€์˜ ๋ช…๋ น์„ ์„ค๊ณ„ํ•˜๊ณ  ๋ช…๋ น์„ ํ–‰๋™์œผ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ํ•˜์œ„ ์ˆ˜์ค€์˜ ์ปจํŠธ๋กค๋Ÿฌ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋Œ€ํ™”๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ [141] ์ตœ์ ํ™” ํ”„๋กœ์„ธ์Šค๋ฅผ ๊ฐ€์†ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [142; 143]์€ ๊ตฌํ˜„๋œ ์˜์‚ฌ ๊ฒฐ์ • ๋ฐ ํƒ์ƒ‰์„ ์œ„ํ•ด ์ž์œจ ์—์ด์ „ํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฌผ๋ฆฌ์  ์ œ์•ฝ์„ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ๋Š” ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  ์—ฌ๋Ÿฌ ์Šคํ‚ฌ์„ ํ™œ์šฉํ•˜์—ฌ ์žฅ๊ธฐ์ ์ธ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ œ์–ด ์ •์ฑ… ์ธก๋ฉด์—์„œ SayCan [78]์€ ์ด๋™์‹ ์กฐ์ž‘ ๋กœ๋ด‡์„ ํ™œ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์กฐ์ž‘ ๋ฐ ํƒ์ƒ‰ ๊ธฐ์ˆ ์„ ์กฐ์‚ฌํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ฃผ๋ฐฉ ํ™˜๊ฒฝ์—์„œ ํ”ํžˆ ๋ณผ ์ˆ˜ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ์ž‘์—…์—์„œ ์˜๊ฐ์„ ์–ป์–ด 7๊ฐ€์ง€ ์Šคํ‚ฌ๊ตฐ๊ณผ 17๊ฐ€์ง€ ๋ฌผ์ฒด๋ฅผ ํฌ๊ด„ํ•˜๋Š” 551๊ฐ€์ง€ ์Šคํ‚ฌ ์„ธํŠธ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์Šคํ‚ฌ์—๋Š” ์ง‘๊ธฐ, ๋†“๊ธฐ, ๋”ฐ๋ฅด๊ธฐ, ์žก๊ธฐ, ์กฐ์ž‘ ๋“ฑ ๋‹ค์–‘ํ•œ ๋™์ž‘์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. TidyBot [137]์€ ๊ฐ€์ • ๋‚ด ์ฒญ์†Œ ์ž‘์—…์„ ๊ฐœ์ธํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋œ ๊ตฌํ˜„ ์—์ด์ „ํŠธ์ž…๋‹ˆ๋‹ค. ์ด ์—์ด์ „ํŠธ๋Š” ํ…์ŠคํŠธ ์˜ˆ์ œ๋ฅผ ํ†ตํ•ด ๋ฌผ์ฒด ๋ฐฐ์น˜ ๋ฐ ์กฐ์ž‘ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์‚ฌ์šฉ์ž์˜ ์„ ํ˜ธ๋„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Figure A5 shows the overlap between datasets selected by SemDeDup and SSL Prototypes. While the two methods do not arrive at the same set of data points, there is a significant overlap between the datasets curated by the two methods. We hypothesize that this is because both SSL prototypes and SemDeDup prune away dense regions of space surrounding cluster centroids: by definition, SemDeDup sparsifies dense regions of space within a cluster; similarly, by definition, SSL prototypes will prune away datapoints close to the cluster centroids. Since K-means clustering places centroids in dense regions of space (see Figure A6 where we observe that the distribution of cosine distances to cluster centroid is skewed right), we know that the regions of space surroundings centroids will be dense, and expect SSL prototypes and SemDeDup to have significant overlap. Qualitatively, we inspect a few examples of points close to cluster centroids in Figure A3, Figure A4, Figure A5, and see that examples close to cluster centroids can be semantically redundant (e.g. templates). Therefore, it makes sense that any reasonable data selection strategy would prioritize sparsifying these dense regions of space surrounding cluster centroids. As mentioned in Section 3.4, sparsifying these dense regions of space containing excessive semantic duplicates is the original motiviation behind D4. Asshown in Figure 7, omitting the re-clustering step significantly worsens performance, and we observe in the rightmost plot of Figure 7 that SemDeDup indeed removes duplicate-driven clusters.
๊ทธ๋ฆผ A5๋Š” SemDeDup๊ณผ SSL Prototypes์— ์˜ํ•ด ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์˜ ์ค‘์ฒฉ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋‘ ๋ฐฉ๋ฒ•์ด ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ง‘ํ•ฉ์— ๋„๋‹ฌํ•˜์ง€๋Š” ์•Š์ง€๋งŒ ๋‘ ๋ฐฉ๋ฒ•์— ์˜ํ•ด ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์—๋Š” ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” SSL Prototypes์™€ SemDeDup์ด ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ๊ณต๊ฐ„์˜ ๋ฐ€์ง‘ ์˜์—ญ์„ ๋ชจ๋‘ ์ž˜๋ผ๋‚ด๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ •์˜์ƒ SemDeDup์€ ํด๋Ÿฌ์Šคํ„ฐ ๋‚ด์˜ ๋ฐ€์ง‘ ์˜์—ญ์„ sparsifyํ•˜๋ฉฐ, ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ •์˜์ƒ SSL Prototypes๋Š” ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ์ž˜๋ผ๋‚ด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. K-means clustering์€ ์ค‘์‹ฌ์„ ๊ณต๊ฐ„์˜ ๋ฐ€์ง‘ ์˜์—ญ์— ๋ฐฐ์น˜ํ•˜๋ฏ€๋กœ(ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๋Œ€ํ•œ ์ฝ”์‚ฌ์ธ ๊ฑฐ๋ฆฌ ๋ถ„ํฌ๊ฐ€ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๋Š” ๊ทธ๋ฆผ A6 ์ฐธ์กฐ), ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ๊ณต๊ฐ„ ์˜์—ญ์ด ๋ฐ€์ง‘๋˜์–ด ์žˆ์„ ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ SSL Prototypes์™€ SemDeDup์ด ์ƒ๋‹นํ•œ ๊ฒน์นจ์„ ๋ณด์ผ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •์„ฑ์ ์œผ๋กœ ๊ทธ๋ฆผ A3, ๊ทธ๋ฆผ A4, ๊ทธ๋ฆผ A5์—์„œ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋ฅผ ๊ฒ€์‚ฌํ•œ ๊ฒฐ๊ณผ, ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ์˜ˆ๋Š” ์˜๋ฏธ์ ์œผ๋กœ ์ค‘๋ณต๋  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ…œํ”Œ๋ฆฟ). ๋”ฐ๋ผ์„œ ํ•ฉ๋ฆฌ์ ์ธ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์€ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ์ด๋Ÿฌํ•œ ๋ฐ€์ง‘๋œ ๊ณต๊ฐ„ ์˜์—ญ์„ sparsifyํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด ํ•ฉ๋ฆฌ์ ์ž…๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ๊ณผ๋„ํ•œ ์˜๋ฏธ์  ์ค‘๋ณต์ด ํฌํ•จ๋œ ์ด๋Ÿฌํ•œ ๋ฐ€์ง‘๋œ ๊ณต๊ฐ„ ์˜์—ญ์„ sparsifyํ•˜๋Š” ๊ฒƒ์ด D4์˜ ์›๋ž˜ ๋™๊ธฐ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ์žฌํด๋Ÿฌ์Šคํ„ฐ๋ง ๋‹จ๊ณ„๋ฅผ ์ƒ๋žตํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ˜„์ €ํžˆ ์•…ํ™”๋˜๋ฉฐ, ๊ทธ๋ฆผ 7์˜ ๊ฐ€์žฅ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—์„œ SemDeDup์ด ์‹ค์ œ๋กœ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๊ทธ๋ฆผ A5๋Š” SemDeDup๊ณผ SSL ํ”„๋กœํ† ํƒ€์ž…์— ์˜ํ•ด ์„ ํƒ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์˜ ์ค‘์ฒฉ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋‘ ๋ฐฉ๋ฒ•์ด ๋™์ผํ•œ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ์ง‘ํ•ฉ์— ๋„๋‹ฌํ•˜์ง€๋Š” ์•Š์ง€๋งŒ ๋‘ ๋ฐฉ๋ฒ•์— ์˜ํ•ด ํ๋ ˆ์ด์…˜๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฐ„์—๋Š” ์ƒ๋‹นํ•œ ์ค‘๋ณต์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” SSL ํ”„๋กœํ† ํƒ€์ž…๊ณผ SemDeDup์ด ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ๊ณต๊ฐ„์˜ ๋ฐ€์ง‘ ์˜์—ญ์„ ๋ชจ๋‘ ์ž˜๋ผ๋‚ด๊ธฐ ๋•Œ๋ฌธ์ด๋ผ๊ณ  ๊ฐ€์„ค์„ ์„ธ์›๋‹ˆ๋‹ค. ์ •์˜์ƒ SemDeDup์€ ํด๋Ÿฌ์Šคํ„ฐ ๋‚ด์˜ ๋ฐ€์ง‘ ์˜์—ญ์„ ํฌ๋ฐ•ํ™”ํ•˜๋ฉฐ, ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์ •์˜์ƒ SSL ํ”„๋กœํ† ํƒ€์ž…์€ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ์ž˜๋ผ๋‚ด๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. K-ํ‰๊ท  ํด๋Ÿฌ์Šคํ„ฐ๋ง์€ ์ค‘์‹ฌ์„ ๊ณต๊ฐ„์˜ ๋ฐ€์ง‘ ์˜์—ญ์— ๋ฐฐ์น˜ํ•˜๋ฏ€๋กœ(ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๋Œ€ํ•œ ์ฝ”์‚ฌ์ธ ๊ฑฐ๋ฆฌ ๋ถ„ํฌ๊ฐ€ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ์น˜์šฐ์ณ ์žˆ๋Š” ๊ทธ๋ฆผ A6 ์ฐธ์กฐ), ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ๊ณต๊ฐ„ ์˜์—ญ์ด ๋ฐ€์ง‘๋˜์–ด ์žˆ์„ ๊ฒƒ์ด๋ผ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ SSL ํ”„๋กœํ† ํƒ€์ž…๊ณผ SemDeDup์ด ์ƒ๋‹นํ•œ ๊ฒน์นจ์„ ๋ณด์ผ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •์„ฑ์ ์œผ๋กœ ๊ทธ๋ฆผ A3, ๊ทธ๋ฆผ A4, ๊ทธ๋ฆผ A5์—์„œ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋ฅผ ๊ฒ€์‚ฌํ•œ ๊ฒฐ๊ณผ, ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์— ๊ฐ€๊นŒ์šด ์˜ˆ๋Š” ์˜๋ฏธ์ ์œผ๋กœ ์ค‘๋ณต๋  ์ˆ˜ ์žˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(์˜ˆ: ํ…œํ”Œ๋ฆฟ). ๋”ฐ๋ผ์„œ ํ•ฉ๋ฆฌ์ ์ธ ๋ฐ์ดํ„ฐ ์„ ํƒ ์ „๋žต์€ ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์„ ๋‘˜๋Ÿฌ์‹ผ ์ด๋Ÿฌํ•œ ๋ฐ€์ง‘๋œ ๊ณต๊ฐ„ ์˜์—ญ์„ ํฌ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์„ ์šฐ์„ ์‹œํ•˜๋Š” ๊ฒƒ์ด ํ•ฉ๋ฆฌ์ ์ž…๋‹ˆ๋‹ค. ์„น์…˜ 3.4์—์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด, ๊ณผ๋„ํ•œ ์˜๋ฏธ์  ์ค‘๋ณต์ด ํฌํ•จ๋œ ์ด๋Ÿฌํ•œ ๋ฐ€์ง‘๋œ ๊ณต๊ฐ„ ์˜์—ญ์„ ํฌ์†Œํ™”ํ•˜๋Š” ๊ฒƒ์ด D4์˜ ์›๋ž˜ ๋™๊ธฐ์ž…๋‹ˆ๋‹ค. ๊ทธ๋ฆผ 7์— ํ‘œ์‹œ๋œ ๊ฒƒ์ฒ˜๋Ÿผ ์žฌํด๋Ÿฌ์Šคํ„ฐ๋ง ๋‹จ๊ณ„๋ฅผ ์ƒ๋žตํ•˜๋ฉด ์„ฑ๋Šฅ์ด ํ˜„์ €ํžˆ ์•…ํ™”๋˜๋ฉฐ, ๊ทธ๋ฆผ 7์˜ ๊ฐ€์žฅ ์˜ค๋ฅธ์ชฝ ํ”Œ๋กฏ์—์„œ SemDeDup์ด ์‹ค์ œ๋กœ ์ค‘๋ณต์œผ๋กœ ์ธํ•œ ํด๋Ÿฌ์Šคํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Lewis et al. (2020) and Guu et al. (2020) were among the first researchers to apply end-to-end training to the field of RALM. However, they differed in their approach. REALM Guu et al. (2020) used masked language training in the pre-training phase and included a retriever that can be trained end-to-end. In the fine-tuning phase, only the QA task was targeted while keeping the retriever frozen. On the other hand, RAG Lewis et al. (2020) used an already trained retriever, DPR, and only employed BART for partial end-to-end training. Similar to REALM, Sachan et al. (2021) present an unsupervised pre-training method that involves an inverse cloze task and masked salient spans. This is followed by supervised fine-tuning using question-context pairs. In addition, they find that the use of end-to-end trained retrievers resulted in a significant improvement in performance across tasks. Singh et al. (2021) apply end-to-end training to multi-document processing, in their proposed approach, the value of the latent variable, which represents the set of relevant documents for a given question, is estimated iteratively. This estimate is then used to update the parameters of the retriever and reader. Siriwardhana et al. (2023) describe the end-to-end optimization of RAG from previous studies and introduces an auxiliary training signal to incorporate more domain-specific knowledge. This signal forces RAG-end2end to reconstruct a given sentence by accessing relevant information in an external knowledge base. This approach has greatly improved domain adaptability.
Lewis ๋“ฑ(2020)๊ณผ Guu ๋“ฑ(2020)์€ RALM ๋ถ„์•ผ์— end-to-end training์„ ์ ์šฉํ•œ ์ตœ์ดˆ์˜ ์—ฐ๊ตฌ์ž ์ค‘ ํ•œ ๋ช…์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋‘ ์—ฐ๊ตฌ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋‹ฌ๋ž์Šต๋‹ˆ๋‹ค. REALM Guu ๋“ฑ(2020)์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ masked language training์„ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ end-to-end ํ•™์Šต์ด ๊ฐ€๋Šฅํ•œ retriever๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. Fine-tuning ๋‹จ๊ณ„์—์„œ๋Š” retriever๋ฅผ ๋™๊ฒฐํ•œ ์ƒํƒœ์—์„œ QA ์ž‘์—…๋งŒ ๋Œ€์ƒ์œผ๋กœ ์‚ผ์•˜์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, RAG Lewis ๋“ฑ(2020)์€ ์ด๋ฏธ ํ•™์Šต๋œ retriever์ธ DPR์„ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ ๋ถ€๋ถ„์ ์ธ end-to-end ํ•™์Šต์„ ์œ„ํ•ด BART๋งŒ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. REALM๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Sachan ๋“ฑ(2021)์€ ์—ญ cloze ์ž‘์—…๊ณผ masked salient spans๋ฅผ ํฌํ•จํ•˜๋Š” ๋น„์ง€๋„ ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” ์งˆ๋ฌธ-๋ฌธ๋งฅ ์Œ์„ ์‚ฌ์šฉํ•˜์—ฌ supervised Fine-tuning์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด๋“ค์€ end-to-end๋กœ ํ›ˆ๋ จ๋œ retriever์˜ ์‚ฌ์šฉ์ด ์ž‘์—… ์ „๋ฐ˜์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Singh ๋“ฑ(2021)์€ ๋‹ค๋ฌธ์„œ ์ฒ˜๋ฆฌ์— end-to-end training์„ ์ ์šฉํ•˜์—ฌ, ์ œ์•ˆํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ด€๋ จ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ž ์žฌ ๋ณ€์ˆ˜์˜ ๊ฐ’์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ถ”์ •์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ retriever์™€ reader์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์—…๋ฐ์ดํŠธํ•ฉ๋‹ˆ๋‹ค. Siriwardhana ๋“ฑ(2023)์€ ์ด์ „ ์—ฐ๊ตฌ์—์„œ RAG์˜ end-to-end optimization์„ ์„ค๋ช…ํ•˜๊ณ  ๋ณด๋‹ค ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ๋ณด์กฐ ํ›ˆ๋ จ ์‹ ํ˜ธ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹ ํ˜ธ๋Š” RAG-end2end๊ฐ€ ์™ธ๋ถ€ ์ง€์‹์ฐฝ๊ณ ์˜ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜์—ฌ ์ฃผ์–ด์ง„ ๋ฌธ์žฅ์„ ์žฌ๊ตฌ์„ฑํ•˜๋„๋ก ๊ฐ•์ œํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋„๋ฉ”์ธ ์ ์‘์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
Lewis ๋“ฑ(2020)๊ณผ Guu ๋“ฑ(2020)์€ RALM ๋ถ„์•ผ์— ์—”๋“œํˆฌ์—”๋“œ ํŠธ๋ ˆ์ด๋‹์„ ์ ์šฉํ•œ ์ตœ์ดˆ์˜ ์—ฐ๊ตฌ์ž ์ค‘ ํ•œ ๋ช…์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋‘ ์—ฐ๊ตฌ๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ด ๋‹ฌ๋ž์Šต๋‹ˆ๋‹ค. REALM Guu ๋“ฑ(2020)์€ ์‚ฌ์ „ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๋งˆ์Šคํฌ ์–ธ์–ด ํ•™์Šต์„ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ ์—”๋“œํˆฌ์—”๋“œ ํ•™์Šต์ด ๊ฐ€๋Šฅํ•œ ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค. ๋ฏธ์„ธ ์กฐ์ • ๋‹จ๊ณ„์—์„œ๋Š” ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ๋™๊ฒฐํ•œ ์ƒํƒœ์—์„œ QA ์ž‘์—…๋งŒ ๋Œ€์ƒ์œผ๋กœ ์‚ผ์•˜์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, RAG Lewis ๋“ฑ(2020)์€ ์ด๋ฏธ ํ•™์Šต๋œ ๊ฒ€์ƒ‰๊ธฐ์ธ DPR์„ ์‚ฌ์šฉํ–ˆ์œผ๋ฉฐ ๋ถ€๋ถ„์ ์ธ ์—”๋“œํˆฌ์—”๋“œ ํ•™์Šต์„ ์œ„ํ•ด BART๋งŒ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค. REALM๊ณผ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Sachan ๋“ฑ(2021)์€ ์—ญ ํด๋กœ์ฆˆ ์ž‘์—…๊ณผ ๋งˆ์Šคํฌ ๋ฒ”์œ„๋ฅผ ํฌํ•จํ•˜๋Š” ๋น„์ง€๋„ ์‚ฌ์ „ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ๋‹ค์Œ์—๋Š” ์งˆ๋ฌธ-๋ฌธ๋งฅ ์Œ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ง€๋„ ๋ฏธ์„ธ ์กฐ์ •์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์ด๋“ค์€ ์—”๋“œํˆฌ์—”๋“œ๋กœ ํ›ˆ๋ จ๋œ ๊ฒ€์ƒ‰๊ธฐ์˜ ์‚ฌ์šฉ์ด ์ž‘์—… ์ „๋ฐ˜์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ๊ฐ€์ ธ์˜จ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. Singh ๋“ฑ(2021)์€ ๋‹ค๋ฌธ์„œ ์ฒ˜๋ฆฌ์— ์—”๋“œํˆฌ์—”๋“œ ํ›ˆ๋ จ์„ ์ ์šฉํ•˜์—ฌ, ์ œ์•ˆํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์—์„œ ์ฃผ์–ด์ง„ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๊ด€๋ จ ๋ฌธ์„œ ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ž ์žฌ ๋ณ€์ˆ˜์˜ ๊ฐ’์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ์ด ์ถ”์ •์น˜๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ์™€ ๋…์ž์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์—…๋ฐ์ดํŠธํ•ฉ๋‹ˆ๋‹ค. Siriwardhana ๋“ฑ(2023)์€ ์ด์ „ ์—ฐ๊ตฌ์—์„œ RAG์˜ ์—”๋“œํˆฌ์—”๋“œ ์ตœ์ ํ™”๋ฅผ ์„ค๋ช…ํ•˜๊ณ  ๋ณด๋‹ค ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•œ ๋ณด์กฐ ํ›ˆ๋ จ ์‹ ํ˜ธ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ์‹ ํ˜ธ๋Š” RAG-end2end๊ฐ€ ์™ธ๋ถ€ ์ง€์‹์ฐฝ๊ณ ์˜ ๊ด€๋ จ ์ •๋ณด์— ์•ก์„ธ์Šคํ•˜์—ฌ ์ฃผ์–ด์ง„ ๋ฌธ์žฅ์„ ์žฌ๊ตฌ์„ฑํ•˜๋„๋ก ๊ฐ•์ œํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋„๋ฉ”์ธ ์ ์‘์„ฑ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.
Human alignment has been discussed a lot for traditional LLMs. In the field of LLM-based autonomous agent, especially when the agents are leveraged for simulation, we believe this concept should be discussed more in depth. In order to better serve human-beings, traditional LLMs are usually fine-tuned to be aligned with correct human values, for example, the agent should not plan to make a bomb for avening society. However, when the agents are leveraged for real-world simulation, an ideal simulator should be able to honestly depict diverse human traits, including the ones with incorrect values. Actually, simulating the human negative aspects can be even more important, since an important goal of simulation is to discover and solve problems, and without negative aspects means no problem to be solved. For example, to simulate the real-world society, we may have to allow the agent to plan for making a bomb, and observe how it will act to implement the plan as well as the influence of its behaviors. Based on these observations, people can make better actions to stop similar behaviors in real-world society. Inspired by the above case, maybe an important problem for agent-based simulation is how to conduct generalized human alignment, that is, for different purposes and applications, the agent should be able to align with diverse human values. However, existing powerful LLMs including ChatGPT and GPT-4 are mostly aligned with unified human values. Thus, an interesting direction is how to "realign" these models by designing proper prompting strategies.
์ธ๊ฐ„ alignment๋Š” ๊ธฐ์กด LLM์— ๋Œ€ํ•ด ๋งŽ์ด ๋…ผ์˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์—์ด์ „ํŠธ๊ฐ€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์— ํ™œ์šฉ๋˜๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์—์„œ๋Š” ์ด ๊ฐœ๋…์„ ๋” ๊นŠ์ด ๋…ผ์˜ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์—๊ฒŒ ๋” ๋‚˜์€ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด LLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌํšŒ ๋ณต์ง€๋ฅผ ์œ„ํ•ด ํญํƒ„์„ ๋งŒ๋“ค ๊ณ„ํš์„ ์„ธ์›Œ์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๋“ฑ ์˜ฌ๋ฐ”๋ฅธ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ Fine-tuning๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—์ด์ „ํŠธ๊ฐ€ ์‹ค์ œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์— ํ™œ์šฉ๋˜๋Š” ๊ฒฝ์šฐ ์ด์ƒ์ ์ธ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์€ ์ž˜๋ชป๋œ ๊ฐ€์น˜๊ด€์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ธ๊ฐ„์˜ ํŠน์„ฑ์„ ์ •์งํ•˜๊ฒŒ ๋ฌ˜์‚ฌํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ์ค‘์š”ํ•œ ๋ชฉํ‘œ๋Š” ๋ฌธ์ œ๋ฅผ ๋ฐœ๊ฒฌํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๋ถ€์ •์ ์ธ ์ธก๋ฉด์ด ์—†๋‹ค๋ฉด ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ฌธ์ œ๊ฐ€ ์—†๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ธ๊ฐ„์˜ ๋ถ€์ •์ ์ธ ์ธก๋ฉด์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋Š” ๊ฒƒ์ด ๋”์šฑ ์ค‘์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‹ค์ œ ์‚ฌํšŒ๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ํญํƒ„์„ ๋งŒ๋“ค ๊ณ„ํš์„ ์„ธ์šฐ๋„๋ก ํ—ˆ์šฉํ•˜๊ณ  ๊ณ„ํš์„ ์‹คํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์–ด๋–ป๊ฒŒ ํ–‰๋™ํ•˜๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ทธ ํ–‰๋™์ด ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ด€์ฐฐํ•ด์•ผ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ์‚ฌ๋žŒ๋“ค์€ ์‹ค์ œ ์‚ฌํšŒ์—์„œ ์œ ์‚ฌํ•œ ํ–‰๋™์„ ๋ง‰๊ธฐ ์œ„ํ•ด ๋” ๋‚˜์€ ํ–‰๋™์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ ์‚ฌ๋ก€์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ์ค‘์š”ํ•œ ๋ฌธ์ œ๋Š” ์ผ๋ฐ˜ํ™”๋œ ์ธ๊ฐ„ alignment, ์ฆ‰ ๋‹ค์–‘ํ•œ ๋ชฉ์ ๊ณผ ์ ์šฉ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๊ฐ€ ๋‹ค์–‘ํ•œ ์ธ๊ฐ„ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ChatGPT ๋ฐ GPT-4๋ฅผ ํฌํ•จํ•œ ๊ธฐ์กด์˜ ๊ฐ•๋ ฅํ•œ LLM์€ ๋Œ€๋ถ€๋ถ„ ํ†ต์ผ๋œ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํฅ๋ฏธ๋กœ์šด ๋ฐฉํ–ฅ์€ ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŠธ ์ „๋žต์„ ์„ค๊ณ„ํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ "์žฌ์ •๋ ฌ"ํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.
์ธ๊ฐ„ ์ •๋ ฌ์€ ๊ธฐ์กด LLM์— ๋Œ€ํ•ด ๋งŽ์ด ๋…ผ์˜๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์—์ด์ „ํŠธ๊ฐ€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์— ํ™œ์šฉ๋˜๋Š” LLM ๊ธฐ๋ฐ˜ ์ž์œจ ์—์ด์ „ํŠธ ๋ถ„์•ผ์—์„œ๋Š” ์ด ๊ฐœ๋…์„ ๋” ๊นŠ์ด ๋…ผ์˜ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ธ๊ฐ„์—๊ฒŒ ๋” ๋‚˜์€ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์กด LLM์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์‚ฌํšŒ ๋ณต์ง€๋ฅผ ์œ„ํ•ด ํญํƒ„์„ ๋งŒ๋“ค ๊ณ„ํš์„ ์„ธ์›Œ์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๋“ฑ ์˜ฌ๋ฐ”๋ฅธ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ ๋ฏธ์„ธ ์กฐ์ •๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—์ด์ „ํŠธ๊ฐ€ ์‹ค์ œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์— ํ™œ์šฉ๋˜๋Š” ๊ฒฝ์šฐ ์ด์ƒ์ ์ธ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์€ ์ž˜๋ชป๋œ ๊ฐ€์น˜๊ด€์„ ํฌํ•จํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ธ๊ฐ„์˜ ํŠน์„ฑ์„ ์ •์งํ•˜๊ฒŒ ๋ฌ˜์‚ฌํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์‹ค์ œ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ์ค‘์š”ํ•œ ๋ชฉํ‘œ๋Š” ๋ฌธ์ œ๋ฅผ ๋ฐœ๊ฒฌํ•˜๊ณ  ํ•ด๊ฒฐํ•˜๋Š” ๊ฒƒ์ด๋ฉฐ, ๋ถ€์ •์ ์ธ ์ธก๋ฉด์ด ์—†๋‹ค๋ฉด ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ฌธ์ œ๊ฐ€ ์—†๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ธ๊ฐ„์˜ ๋ถ€์ •์ ์ธ ์ธก๋ฉด์„ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋Š” ๊ฒƒ์ด ๋”์šฑ ์ค‘์š”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์‹ค์ œ ์‚ฌํšŒ๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด ์—์ด์ „ํŠธ๊ฐ€ ํญํƒ„์„ ๋งŒ๋“ค ๊ณ„ํš์„ ์„ธ์šฐ๋„๋ก ํ—ˆ์šฉํ•˜๊ณ  ๊ณ„ํš์„ ์‹คํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์–ด๋–ป๊ฒŒ ํ–‰๋™ํ•˜๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ๊ทธ ํ–‰๋™์ด ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ๊ด€์ฐฐํ•ด์•ผ ํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ด€์ฐฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ์‚ฌ๋žŒ๋“ค์€ ์‹ค์ œ ์‚ฌํšŒ์—์„œ ์œ ์‚ฌํ•œ ํ–‰๋™์„ ๋ง‰๊ธฐ ์œ„ํ•ด ๋” ๋‚˜์€ ํ–‰๋™์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„์˜ ์‚ฌ๋ก€์—์„œ ์˜๊ฐ์„ ๋ฐ›์•„ ์—์ด์ „ํŠธ ๊ธฐ๋ฐ˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์˜ ์ค‘์š”ํ•œ ๋ฌธ์ œ๋Š” ์ผ๋ฐ˜ํ™”๋œ ์ธ๊ฐ„ ์ •๋ ฌ, ์ฆ‰ ๋‹ค์–‘ํ•œ ๋ชฉ์ ๊ณผ ์ ์šฉ์— ๋”ฐ๋ผ ์—์ด์ „ํŠธ๊ฐ€ ๋‹ค์–‘ํ•œ ์ธ๊ฐ„ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์ผ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ChatGPT ๋ฐ GPT-4๋ฅผ ํฌํ•จํ•œ ๊ธฐ์กด์˜ ๊ฐ•๋ ฅํ•œ LLM์€ ๋Œ€๋ถ€๋ถ„ ํ†ต์ผ๋œ ์ธ๊ฐ„์˜ ๊ฐ€์น˜๊ด€์— ๋งž์ถฐ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํฅ๋ฏธ๋กœ์šด ๋ฐฉํ–ฅ์€ ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŠธ ์ „๋žต์„ ์„ค๊ณ„ํ•˜์—ฌ ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ "์žฌ์ •๋ ฌ"ํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.
Utilizing one of the most advanced open-source models, Llama-2 7B, our research investigates the effects on LLMs when they undergo finetuning8 with data generated by other LLMs. To ensure the generation of high-quality data and to provide a relevant but not trivial downstream task, we employ the Wikitext-103 dataset. We segment this dataset into chunks of 128 tokens, between each with a stride of 64 tokens, resulting in approximately 2.2 million chunks. Denote this dataset as \(\mathcal{D}_{0}\). The task for generation involves producing the final 32 tokens given the initial 96 tokens from each chunk in the original dataset. In the initial generation (0-th generation), we use the Llama-2 7B FT model, which has been finetuned on \(\mathcal{D}_{0}\), applying a generation loss that focuses solely on the cross-entropy loss of the final 32 tokens. We denote this initial model as \(\mathcal{M}_{0}\), which demonstrates enhanced capacity for the generation task compared to the standard Llama-2 7B model. By querying \(\mathcal{M}_{0}\) with the original 96 tokens from \(\mathcal{D}_{0}\), we generate the dataset \(\mathcal{D}_{1}\) and subsequently finetune Llama-2 7B on this dataset to obtain \(\mathcal{M}_{1}\). This process is sequentially repeated to generate \(\mathcal{D}_{i}\) from \(\mathcal{M}_{i-1}\) and obtain \(\mathcal{M}_{i}\) through finetuning. By comparing the performance of various \(\mathcal{M}\) models on the test set derived from Wikitext-103, also segmented into 128-token chunks, we aim to investigate the model collapse in LLMs.
๊ฐ€์žฅ ์ง„๋ณด๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ธ Llama-2 7B๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ๋‹ค๋ฅธ LLM์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ๋กœ Fine-tuning์„ ๊ฑฐ์นœ LLM์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ด€๋ จ์„ฑ์ด ๋†’์ง€๋งŒ ์‚ฌ์†Œํ•˜์ง€ ์•Š์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด Wikitext-103 ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์•ฝ 220๋งŒ ๊ฐœ์˜ ์ฒญํฌ๋กœ ๋ถ„ํ• ํ•˜๊ณ , ๊ฐ ์ฒญํฌ ์‚ฌ์ด์— 64๊ฐœ์˜ ํ† ํฐ ๊ฐ„๊ฒฉ์„ ๋‘์–ด ํ† ํฐ 128๊ฐœ๋กœ ์„ธ๋ถ„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ \(\mathcal{D}_{0}\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ ์ž‘์—…์€ ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์ฒญํฌ์—์„œ ์ฒ˜์Œ 96๊ฐœ์˜ ํ† ํฐ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๋งˆ์ง€๋ง‰ 32๊ฐœ์˜ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์ƒ์„ฑ(0์„ธ๋Œ€)์—์„œ๋Š” \(\mathcal{D}_{0}\)์— ๋Œ€ํ•ด Fine-tuning๋œ Llama-2 7B FT ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งˆ์ง€๋ง‰ 32๊ฐœ์˜ ํ† ํฐ์˜ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ Loss์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ˜ ์ƒ์„ฑ Loss๋ฅผ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ดˆ๊ธฐ ๋ชจ๋ธ์„ \(\mathcal{M}_{0}\)์œผ๋กœ ํ‘œ์‹œํ•˜๋ฉฐ, ์ด ๋ชจ๋ธ์€ ํ‘œ์ค€ Llama-2 7B ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ์„ฑ ์ž‘์—…์— ๋Œ€ํ•œ ์šฉ๋Ÿ‰์ด ํ–ฅ์ƒ๋œ ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. \(\mathcal{M}_{0}\)์„ ์‚ฌ์šฉํ•˜์—ฌ \(\mathcal{D}_{0}\)์˜ ์›๋ž˜ 96๊ฐœ ํ† ํฐ์„ ์ฟผ๋ฆฌํ•˜๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{1}\)์„ ์ƒ์„ฑํ•˜๊ณ  ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Llama-2 7B๋ฅผ Fine-tuningํ•˜์—ฌ \(\mathcal{M}_{1}\)์„ ์–ป์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ฐ˜๋ณตํ•˜์—ฌ \(\mathcal{M}_{i-1}\)์—์„œ \(\mathcal{D}_{i}\)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  Fine-tuning์„ ํ†ตํ•ด \(\mathcal{M}_{i}\)๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. 128 ํ† ํฐ ์ฒญํฌ๋กœ ์„ธ๋ถ„ํ™”๋œ Wikitext-103์—์„œ ํŒŒ์ƒ๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ \(\mathcal{M}\) ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์—ฌ LLM์˜ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
๊ฐ€์žฅ ์ง„๋ณด๋œ ์˜คํ”ˆ ์†Œ์Šค ๋ชจ๋ธ ์ค‘ ํ•˜๋‚˜์ธ Llama-2 7B๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ๋‹ค๋ฅธ LLM์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ๋กœ ๋ฏธ์„ธ ์กฐ์ •8์„ ๊ฑฐ์นœ LLM์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๊ด€๋ จ์„ฑ์ด ๋†’์ง€๋งŒ ์‚ฌ์†Œํ•˜์ง€ ์•Š์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด Wikitext-103 ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์•ฝ 220๋งŒ ๊ฐœ์˜ ์ฒญํฌ๋กœ ๋ถ„ํ• ํ•˜๊ณ , ๊ฐ ์ฒญํฌ ์‚ฌ์ด์— 64๊ฐœ์˜ ํ† ํฐ ๊ฐ„๊ฒฉ์„ ๋‘์–ด ํ† ํฐ 128๊ฐœ๋กœ ์„ธ๋ถ„ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ \(\mathcal{D}_{0}\)๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ ์ž‘์—…์€ ์›๋ณธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ์ฒญํฌ์—์„œ ์ฒ˜์Œ 96๊ฐœ์˜ ํ† ํฐ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ๋งˆ์ง€๋ง‰ 32๊ฐœ์˜ ํ† ํฐ์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ์ƒ์„ฑ(0์„ธ๋Œ€)์—์„œ๋Š” \(\mathcal{D}_{0}\)์— ๋Œ€ํ•ด ๋ฏธ์„ธ ์กฐ์ •๋œ Llama-2 7B FT ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋งˆ์ง€๋ง‰ 32๊ฐœ์˜ ํ† ํฐ์˜ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ˜ ์ƒ์„ฑ ์†์‹ค์„ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์ดˆ๊ธฐ ๋ชจ๋ธ์„ \(\mathcal{M}_{0}\)์œผ๋กœ ํ‘œ์‹œํ•˜๋ฉฐ, ์ด ๋ชจ๋ธ์€ ํ‘œ์ค€ Llama-2 7B ๋ชจ๋ธ์— ๋น„ํ•ด ์ƒ์„ฑ ์ž‘์—…์— ๋Œ€ํ•œ ์šฉ๋Ÿ‰์ด ํ–ฅ์ƒ๋œ ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. mathcal{M}_{0}\)์„ ์‚ฌ์šฉํ•˜์—ฌ \(\mathcal{D}_{0}\)์˜ ์›๋ž˜ 96๊ฐœ ํ† ํฐ์„ ์ฟผ๋ฆฌํ•˜๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ \(\mathcal{D}_{1}\)์„ ์ƒ์„ฑํ•˜๊ณ  ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•ด Llama-2 7B๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ \(\mathcal{M}_{1}\)์„ ์–ป์Šต๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์„ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ฐ˜๋ณตํ•˜์—ฌ \(\mathcal{M}_{i-1}\)์—์„œ \(\mathcal{D}_{i}\)๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ๋ฏธ์„ธ ์กฐ์ •์„ ํ†ตํ•ด \(\mathcal{M}_{i}\)๋ฅผ ์–ป์Šต๋‹ˆ๋‹ค. 128 ํ† ํฐ ์ฒญํฌ๋กœ ์„ธ๋ถ„ํ™”๋œ Wikitext-103์—์„œ ํŒŒ์ƒ๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ์—์„œ ๋‹ค์–‘ํ•œ \(\mathcal{M}\) ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์—ฌ LLM์˜ ๋ชจ๋ธ ๋ถ•๊ดด๋ฅผ ์กฐ์‚ฌํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค.
Benefits of Synthesized DataThere is a range of results showing benefits of AI-synthesized data in training better models, though mostly these results pertain to image data, specifically in the context of diffusion models (Azizi et al., 2023; He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023), though not only (see Dai et al. (2023); Xu et al. (2023); Huang et al. (2022); Wang et al. (2023) for chat-related examples). One might argue that they either throw model-collapse caution to the winds or, possibly, settle in the protected corner where mild amounts of synthetic data (or larger amounts of "mildly synthetic" data, like in the case of data augmentation) helps. In particular, often benefits of synthetic data are observed when the synthetic data is generated by a model trained for a different use case than the downstream task (like images synthesized from diffusion models helping classification models) or generated by a stronger model (He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023). However, other works critically analyze the purported benefit of generated data. (Burg et al., 2023) find that while synthesized data from a diffusion model helps improving downstream tasks, such as classification, using the _pre-training data_ of the diffusion model alone gives even stronger performance (which we can interpret as evidence of mild first-generation model collapse). All in all it is fair to say that the impact of data augmentation using generative models is still not fully understood.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ๋” ๋‚˜์€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์žˆ์–ด AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋‹ค์–‘ํ•œ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ, ํŠนํžˆ diffusion ๋ชจ๋ธ์˜ ๋งฅ๋ฝ์—์„œ ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚˜์ง€๋งŒ(Azizi et al., 2023; He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023), ์ด์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ฑ„ํŒ… ๊ด€๋ จ ์˜ˆ๋Š” Dai et al. (2023); Xu et al. (2023); Huang et al. (2022); Wang et al. (2023)). ๋ชจ๋ธ ๋ถ•๊ดด์— ๋Œ€ํ•œ ์ฃผ์˜๋ฅผ ๋ฌด์‹œํ•˜๊ฑฐ๋‚˜, ์•ฝ๊ฐ„์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(๋˜๋Š” data augmentation์˜ ๊ฒฝ์šฐ์ฒ˜๋Ÿผ '์•ฝ๊ฐ„ ํ•ฉ์„ฑ๋œ' ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ๋” ๋งŽ์ด)๊ฐ€ ๋„์›€์ด ๋˜๋Š” ๋ณดํ˜ธ ๊ตฌ์—ญ์— ์•ˆ์ฃผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…๊ณผ ๋‹ค๋ฅธ ์‚ฌ์šฉ ์‚ฌ๋ก€(์˜ˆ: ๋ถ„๋ฅ˜ ๋ชจ๋ธ์— ๋„์›€์ด ๋˜๋Š” diffusion ๋ชจ๋ธ์—์„œ ํ•ฉ์„ฑ๋œ ์ด๋ฏธ์ง€)๋ฅผ ์œ„ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋˜๊ฑฐ๋‚˜ ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋  ๋•Œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์ด ๊ด€์ฐฐ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค(He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์— ๋Œ€ํ•ด ๋น„ํŒ์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. (Burg et al., 2023)์€ diffusion ๋ชจ๋ธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ„๋ฅ˜์™€ ๊ฐ™์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, diffusion ๋ชจ๋ธ์˜ _pre-training data_๋งŒ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(์ด๋Š” ๊ฒฝ๋ฏธํ•œ 1์„ธ๋Œ€ ๋ชจ๋ธ ๋ถ•๊ดด์˜ ์ฆ๊ฑฐ๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค). ์ƒ์„ฑ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ data augmentation์˜ ์˜ํ–ฅ์€ ์•„์ง ์™„์ „ํžˆ ์ดํ•ด๋˜์ง€ ์•Š์•˜๋‹ค๊ณ  ํ•ด๋„ ๊ณผ์–ธ์ด ์•„๋‹™๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ๋” ๋‚˜์€ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๋Š” ๋ฐ ์žˆ์–ด AI ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์„ ๋ณด์—ฌ์ฃผ๋Š” ๋‹ค์–‘ํ•œ ๊ฒฐ๊ณผ๊ฐ€ ์žˆ์ง€๋งŒ, ๋Œ€๋ถ€๋ถ„ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ, ํŠนํžˆ ํ™•์‚ฐ ๋ชจ๋ธ์˜ ๋งฅ๋ฝ์—์„œ ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜ํƒ€๋‚˜์ง€๋งŒ(Azizi et al., 2023; He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023), ์ด์—๋งŒ ๊ตญํ•œ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค(์ฑ„ํŒ… ๊ด€๋ จ ์˜ˆ๋Š” Dai et al. (2023); Xu et al. (2023); Huang et al. (2022); Wang et al. (2023)). ๋ชจ๋ธ ๋ถ•๊ดด์— ๋Œ€ํ•œ ์ฃผ์˜๋ฅผ ๋ฌด์‹œํ•˜๊ฑฐ๋‚˜, ์•ฝ๊ฐ„์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ(๋˜๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ๊ฒฝ์šฐ์ฒ˜๋Ÿผ '์•ฝ๊ฐ„ ํ•ฉ์„ฑ๋œ' ๋ฐ์ดํ„ฐ์˜ ์–‘์ด ๋” ๋งŽ์ด)๊ฐ€ ๋„์›€์ด ๋˜๋Š” ๋ณดํ˜ธ ๊ตฌ์—ญ์— ์•ˆ์ฃผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…๊ณผ ๋‹ค๋ฅธ ์‚ฌ์šฉ ์‚ฌ๋ก€(์˜ˆ: ๋ถ„๋ฅ˜ ๋ชจ๋ธ์— ๋„์›€์ด ๋˜๋Š” ํ™•์‚ฐ ๋ชจ๋ธ์—์„œ ํ•ฉ์„ฑ๋œ ์ด๋ฏธ์ง€)๋ฅผ ์œ„ํ•ด ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋˜๊ฑฐ๋‚˜ ๋” ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์— ์˜ํ•ด ์ƒ์„ฑ๋  ๋•Œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์ด ๊ด€์ฐฐ๋˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค(He et al., 2023; Shipard et al., 2023; Bansal and Grover, 2023; Lin et al., 2023). ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค๋ฅธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ์ด์ ์— ๋Œ€ํ•ด ๋น„ํŒ์ ์œผ๋กœ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. (Burg et al., 2023)์€ ํ™•์‚ฐ ๋ชจ๋ธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ„๋ฅ˜์™€ ๊ฐ™์€ ๋‹ค์šด์ŠคํŠธ๋ฆผ ์ž‘์—…์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋˜์ง€๋งŒ, ํ™•์‚ฐ ๋ชจ๋ธ์˜ _์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ_๋งŒ ์‚ฌ์šฉํ•˜๋ฉด ๋” ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(์ด๋Š” ๊ฒฝ๋ฏธํ•œ 1์„ธ๋Œ€ ๋ชจ๋ธ ๋ถ•๊ดด์˜ ์ฆ๊ฑฐ๋กœ ํ•ด์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค). ์ƒ์„ฑ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ์˜ํ–ฅ์€ ์•„์ง ์™„์ „ํžˆ ์ดํ•ด๋˜์ง€ ์•Š์•˜๋‹ค๊ณ  ํ•ด๋„ ๊ณผ์–ธ์ด ์•„๋‹™๋‹ˆ๋‹ค.
In existing literature [823, 824], multimodal models mainly refer to the models that can process and integrate information of various modalities (_e.g._, text, image, and audio) from input, and further produce corresponding output in certain modalities. In this part, we mainly focus on the multimodal extension of LLMs by enabling the information modeling of non-textual modalities, especially the vision modality, called _multimodal large language models (MLLMs)_[797]49. To start our discussion, we specify the input to be text-image pairs and the output to be text responses. Similar discussions can be made for other modalities, _e.g._, language-audio models [825], which is beyond our scope here. In essence, MLLMs are developed by adapting the information from other modalities to the text modality, so as to leverage the excellent model capacities of LLMs that are learned based on world text. Typically, a MLLM comprises an image encoder for image encoding and a LLM for text generation, associated by a connection module that aligns vision and language representations. During generation, the image is first split into patches, and then transformed into patch embeddings by the image encoder and the connection module, to derive a visual representation that can be understood by the LLM. Subsequently, the patch embeddings and text embeddings are concatenated, and fed into the MLLM, allowing the language model to generate the response autoregressively. In the following, we will discuss the training, evaluation, and key points to develop capable MLLMs.
๊ธฐ์กด ๋ฌธํ—Œ[823, 824]์—์„œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์€ ์ฃผ๋กœ ์ž…๋ ฅ์—์„œ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ(_์˜ˆ: ํ…์ŠคํŠธ, ์ด๋ฏธ์ง€, ์˜ค๋””์˜ค)์˜ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ  ํ†ตํ•ฉํ•˜์—ฌ ํŠน์ • ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์—์„œ ํ•ด๋‹น ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ํ…์ŠคํŠธ๊ฐ€ ์•„๋‹Œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ, ํŠนํžˆ ๋น„์ „ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์˜ ์ •๋ณด ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์—ฌ LLM์„ ํ™•์žฅํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ ‘๊ทผ ๋ฐฉ์‹์— ์ฃผ๋กœ ์ดˆ์ ์„ ๋งž์ถ”๋ฉฐ, ์ด๋ฅผ _multimodal large language models (MLLMs)_[797]49๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ์˜๋ฅผ ์‹œ์ž‘ํ•˜๊ธฐ ์œ„ํ•ด ์ž…๋ ฅ์€ ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ์Œ์ด๊ณ  ์ถœ๋ ฅ์€ ํ…์ŠคํŠธ ์‘๋‹ต์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚˜๋Š” ๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ, _์˜ˆ๋ฅผ ๋“ค์–ด language-audio models[825]์— ๋Œ€ํ•ด์„œ๋„ ์œ ์‚ฌํ•œ ๋…ผ์˜๊ฐ€ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณธ์งˆ์ ์œผ๋กœ MLLMs์€ ๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์˜ ์ •๋ณด๋ฅผ ํ…์ŠคํŠธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์— ์ ์šฉํ•˜์—ฌ ์„ธ๊ณ„ ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต๋œ LLM์˜ ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐœ๋ฐœ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ MLLM์€ ์ด๋ฏธ์ง€ ์ธ์ฝ”๋”ฉ์„ ์œ„ํ•œ image encoder์™€ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•œ LLM์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์‹œ๊ฐ๊ณผ ์–ธ์–ด ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋Š” ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋กœ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ ์ค‘์— ์ด๋ฏธ์ง€๋Š” ๋จผ์ € ํŒจ์น˜๋กœ ๋ถ„ํ• ๋œ ๋‹ค์Œ image encoder์™€ ์—ฐ๊ฒฐ ๋ชจ๋“ˆ์— ์˜ํ•ด ํŒจ์น˜ ์ž„๋ฒ ๋”ฉ์œผ๋กœ ๋ณ€ํ™˜๋˜์–ด LLM์ด ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ๊ฐ์  ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํŒจ์น˜ ์ž„๋ฒ ๋”ฉ๊ณผ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ์„ ์—ฐ๊ฒฐํ•˜์—ฌ MLLM์— ๊ณต๊ธ‰ํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์ด ์‘๋‹ต์„ ์ž๋™ ํšŒ๊ท€์ ์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋Šฅ๋ ฅ ์žˆ๋Š” MLLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ํ›ˆ๋ จ, ํ‰๊ฐ€ ๋ฐ ํ•ต์‹ฌ ์‚ฌํ•ญ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
๊ธฐ์กด ๋ฌธํ—Œ[823, 824]์—์„œ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋ชจ๋ธ์€ ์ฃผ๋กœ ์ž…๋ ฅ์—์„œ ๋‹ค์–‘ํ•œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ(_์˜ˆ: ํ…์ŠคํŠธ, ์ด๋ฏธ์ง€, ์˜ค๋””์˜ค)์˜ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ  ํ†ตํ•ฉํ•˜์—ฌ ํŠน์ • ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์—์„œ ํ•ด๋‹น ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์—์„œ๋Š” ํ…์ŠคํŠธ๊ฐ€ ์•„๋‹Œ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ, ํŠนํžˆ ๋น„์ „ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์˜ ์ •๋ณด ๋ชจ๋ธ๋ง์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•˜์—ฌ LLM์„ ํ™•์žฅํ•˜๋Š” ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ ‘๊ทผ ๋ฐฉ์‹์— ์ฃผ๋กœ ์ดˆ์ ์„ ๋งž์ถ”๋ฉฐ, ์ด๋ฅผ _๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(MMLM)_[797]49๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ์˜๋ฅผ ์‹œ์ž‘ํ•˜๊ธฐ ์œ„ํ•ด ์ž…๋ ฅ์€ ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ์Œ์ด๊ณ  ์ถœ๋ ฅ์€ ํ…์ŠคํŠธ ์‘๋‹ต์ด๋ผ๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ๋Š” ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚˜๋Š” ๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ, _์˜ˆ๋ฅผ ๋“ค์–ด ์–ธ์–ด-์˜ค๋””์˜ค ๋ชจ๋ธ[825]์— ๋Œ€ํ•ด์„œ๋„ ์œ ์‚ฌํ•œ ๋…ผ์˜๊ฐ€ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณธ์งˆ์ ์œผ๋กœ MLLM์€ ๋‹ค๋ฅธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์˜ ์ •๋ณด๋ฅผ ํ…์ŠคํŠธ ๋ชจ๋‹ฌ๋ฆฌํ‹ฐ์— ์ ์šฉํ•˜์—ฌ ์„ธ๊ณ„ ํ…์ŠคํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต๋œ LLM์˜ ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ ์šฉ๋Ÿ‰์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐœ๋ฐœ๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ MLLM์€ ์ด๋ฏธ์ง€ ์ธ์ฝ”๋”ฉ์„ ์œ„ํ•œ ์ด๋ฏธ์ง€ ์ธ์ฝ”๋”์™€ ํ…์ŠคํŠธ ์ƒ์„ฑ์„ ์œ„ํ•œ LLM์œผ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์‹œ๊ฐ๊ณผ ์–ธ์–ด ํ‘œํ˜„์„ ์ •๋ ฌํ•˜๋Š” ์—ฐ๊ฒฐ ๋ชจ๋“ˆ๋กœ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. ์ƒ์„ฑ ์ค‘์— ์ด๋ฏธ์ง€๋Š” ๋จผ์ € ํŒจ์น˜๋กœ ๋ถ„ํ• ๋œ ๋‹ค์Œ ์ด๋ฏธ์ง€ ์ธ์ฝ”๋”์™€ ์—ฐ๊ฒฐ ๋ชจ๋“ˆ์— ์˜ํ•ด ํŒจ์น˜ ์ž„๋ฒ ๋”ฉ์œผ๋กœ ๋ณ€ํ™˜๋˜์–ด LLM์ด ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ๊ฐ์  ํ‘œํ˜„์„ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ ํŒจ์น˜ ์ž„๋ฒ ๋”ฉ๊ณผ ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ์„ ์—ฐ๊ฒฐํ•˜์—ฌ MLLM์— ๊ณต๊ธ‰ํ•จ์œผ๋กœ์จ ์–ธ์–ด ๋ชจ๋ธ์ด ์‘๋‹ต์„ ์ž๋™ ํšŒ๊ท€์ ์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์—์„œ๋Š” ๋Šฅ๋ ฅ ์žˆ๋Š” MLLM์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ํ›ˆ๋ จ, ํ‰๊ฐ€ ๋ฐ ํ•ต์‹ฌ ์‚ฌํ•ญ์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
On single-turn dialogue, we evaluate the different methods on the subset of the test split of the Anthropic HH dataset [1] with one step of human-assistant interaction. GPT-4 evaluations use the preferred completions on the test as the reference to compute the win rate for different methods. As there is no standard SFT model for this task, we start with a pre-trained Pythia-2.8B, use Preferred-FT to train a reference model on the chosen completions such that completions are within distribution of the model, and then train using DPO. We also compare against the best of 128 Preferred-FT completions (we found the Best of \(N\) baseline plateaus at 128 completions for this task; see Appendix Figure 4) and a 2-shot prompted version of the Pythia-2.8B base model, finding DPO performs as well or better for the best-performing temperatures for each method. We also evaluate an RLHF model trained with PPO on the Anthropic HH dataset 5 from a well-known source 6, but are unable to find a prompt or sampling temperature that gives performance better than the base Pythia-2.8B model. Based on our results from TL;DR and the fact that both methods optimize the same reward function, we consider Best of 128 a rough proxy for PPO-level performance. Overall, DPO is the only computationally efficient method that improves over the preferred completions in the Anthropic HH dataset, and provides similar or better performance to the computationally demanding Best of 128 baseline. Finally, Figure 3 shows that DPO converges to its best performance relatively quickly.
๋‹จ์ผ ํ„ด ๋Œ€ํ™”์—์„œ๋Š” ์ธ๊ฐ„๊ณผ ์–ด์‹œ์Šคํ„ดํŠธ์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ํ•œ ๋‹จ๊ณ„ ์žˆ๋Š” Anthropic HH ๋ฐ์ดํ„ฐ ์„ธํŠธ [1]์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ•  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. GPT-4 ํ‰๊ฐ€๋Š” ํ…Œ์ŠคํŠธ์—์„œ ์„ ํ˜ธํ•˜๋Š” ์™„์„ฑ์„ ์ฐธ์กฐ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์˜ ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์—๋Š” ํ‘œ์ค€ Supervised Fine-tuning (SFT) ๋ชจ๋ธ์ด ์—†์œผ๋ฏ€๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ Pythia-2.8B๋กœ ์‹œ์ž‘ํ•˜์—ฌ, Preferred-FT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ ํƒํ•œ ์™„์„ฑ์— ๋Œ€ํ•œ ์ฐธ์กฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์™„์„ฑ์ด ๋ชจ๋ธ์˜ ๋ถ„ํฌ ๋‚ด์— ์žˆ๋„๋ก ํ•œ ๋‹ค์Œ DPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 128๊ฐœ์˜ Preferred-FT ์™„์„ฑ ์ค‘ ์ตœ๊ณ ์˜ ์™„์„ฑ๊ณผ ๋น„๊ตํ•˜์—ฌ(์ด ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ตœ๊ณ ์˜ \(N\) ๊ธฐ์ค€์„ ์ด 128๊ฐœ์˜ ์™„์„ฑ์—์„œ ์ •์ฒด๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฐธ๊ณ  ๊ทธ๋ฆผ 4 ์ฐธ์กฐ), 2-shot ํ”„๋กฌํ”„ํŠธ ๋ฒ„์ „์˜ Pythia-2.8B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ๋น„๊ตํ•˜์—ฌ ๊ฐ ๋ฐฉ๋ฒ•์˜ ์ตœ์  ์„ฑ๋Šฅ ์˜จ๋„์— ๋Œ€ํ•ด DPO๊ฐ€ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š”์ง€ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž˜ ์•Œ๋ ค์ง„ ์†Œ์Šค 6์˜ Anthropic HH ๋ฐ์ดํ„ฐ ์„ธํŠธ 5์—์„œ PPO๋กœ ํ•™์Šต๋œ Reinforcement Learning from Human Feedback (RLHF) ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ–ˆ์ง€๋งŒ, ๊ธฐ๋ณธ Pythia-2.8B ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋” ๋‚˜์€ ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” ์ƒ˜ํ”Œ๋ง ์˜จ๋„๋ฅผ ์ฐพ์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. TL;DR์˜ ๊ฒฐ๊ณผ์™€ ๋‘ ๋ฐฉ๋ฒ•์ด ๋™์ผํ•œ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐ”ํƒ•์œผ๋กœ 128๊ฐœ ์ค‘ ์ตœ๊ณ ์˜ ๊ฒƒ์„ PPO ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๋Œ€๋žต์ ์ธ ๋Œ€๋ฆฌ ์—ญํ• ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ DPO๋Š” Anthropic HH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํ˜ธ๋„ ์™„์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ์œ ์ผํ•œ ๊ณ„์‚ฐ ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์ด๋ฉฐ, ๊ณ„์‚ฐ ์ง‘์•ฝ์ ์ธ ์ตœ๊ณ ์˜ 128 ๊ธฐ์ค€์„ ๊ณผ ๋น„์Šทํ•˜๊ฑฐ๋‚˜ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ทธ๋ฆผ 3์€ DPO๊ฐ€ ๋น„๊ต์  ๋น ๋ฅด๊ฒŒ ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์— ์ˆ˜๋ ดํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
๋‹จ์ผ ํ„ด ๋Œ€ํ™”์—์„œ๋Š” ์ธ๊ฐ„๊ณผ ์–ด์‹œ์Šคํ„ดํŠธ์˜ ์ƒํ˜ธ ์ž‘์šฉ์ด ํ•œ ๋‹จ๊ณ„ ์žˆ๋Š” Anthropic HH ๋ฐ์ดํ„ฐ ์„ธํŠธ [1]์˜ ํ…Œ์ŠคํŠธ ๋ถ„ํ•  ํ•˜์œ„ ์ง‘ํ•ฉ์—์„œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. GPT-4 ํ‰๊ฐ€๋Š” ํ…Œ์ŠคํŠธ์—์„œ ์„ ํ˜ธํ•˜๋Š” ์™„์„ฑ์„ ์ฐธ์กฐ๋กœ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์˜ ์Šน๋ฅ ์„ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. ์ด ์ž‘์—…์—๋Š” ํ‘œ์ค€ SFT ๋ชจ๋ธ์ด ์—†์œผ๋ฏ€๋กœ ์‚ฌ์ „ ํ•™์Šต๋œ Pythia-2.8B๋กœ ์‹œ์ž‘ํ•˜์—ฌ, ์„ ํ˜ธ๋„-FT๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ ํƒํ•œ ์™„์„ฑ์— ๋Œ€ํ•œ ์ฐธ์กฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผœ ์™„์„ฑ์ด ๋ชจ๋ธ์˜ ๋ถ„ํฌ ๋‚ด์— ์žˆ๋„๋ก ํ•œ ๋‹ค์Œ DPO๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ 128๊ฐœ์˜ ์„ ํ˜ธ๋„-FT ์™„์„ฑ ์ค‘ ์ตœ๊ณ ์˜ ์™„์„ฑ๊ณผ ๋น„๊ตํ•˜์—ฌ(์ด ์ž‘์—…์˜ ๊ฒฝ์šฐ ์ตœ๊ณ ์˜ \(N\) ๊ธฐ์ค€์„ ์ด 128๊ฐœ์˜ ์™„์„ฑ์—์„œ ์ •์ฒด๋˜๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฐธ๊ณ  ๊ทธ๋ฆผ 4 ์ฐธ์กฐ), 2์ƒท ํ”„๋กฌํ”„ํŠธ ๋ฒ„์ „์˜ Pythia-2.8B ๊ธฐ๋ณธ ๋ชจ๋ธ์„ ๋น„๊ตํ•˜์—ฌ ๊ฐ ๋ฐฉ๋ฒ•์˜ ์ตœ์  ์„ฑ๋Šฅ ์˜จ๋„์— ๋Œ€ํ•ด DPO๊ฐ€ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๋Š”์ง€ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์ž˜ ์•Œ๋ ค์ง„ ์†Œ์Šค 6์˜ Anthropic HH ๋ฐ์ดํ„ฐ ์„ธํŠธ 5์—์„œ PPO๋กœ ํ•™์Šต๋œ RLHF ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ–ˆ์ง€๋งŒ, ๊ธฐ๋ณธ Pythia-2.8B ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋” ๋‚˜์€ ํ”„๋กฌํ”„ํŠธ ๋˜๋Š” ์ƒ˜ํ”Œ๋ง ์˜จ๋„๋ฅผ ์ฐพ์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค. TL;DR์˜ ๊ฒฐ๊ณผ์™€ ๋‘ ๋ฐฉ๋ฒ•์ด ๋™์ผํ•œ ๋ณด์ƒ ํ•จ์ˆ˜๋ฅผ ์ตœ์ ํ™”ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ฐ”ํƒ•์œผ๋กœ 128๊ฐœ ์ค‘ ์ตœ๊ณ ์˜ ๊ฒƒ์„ PPO ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์— ๋Œ€ํ•œ ๋Œ€๋žต์ ์ธ ๋Œ€๋ฆฌ ์—ญํ• ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์ „๋ฐ˜์ ์œผ๋กœ DPO๋Š” ์•ˆํŠธ๋กœํ”ฝ HH ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์„ ํ˜ธ๋„ ์™„์„ฑ์„ ๊ฐœ์„ ํ•˜๋Š” ์œ ์ผํ•œ ๊ณ„์‚ฐ ํšจ์œจ์ ์ธ ๋ฐฉ๋ฒ•์ด๋ฉฐ, ๊ณ„์‚ฐ ์ง‘์•ฝ์ ์ธ ์ตœ๊ณ ์˜ 128 ๊ธฐ์ค€์„ ๊ณผ ๋น„์Šทํ•˜๊ฑฐ๋‚˜ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๊ทธ๋ฆผ 3์€ DPO๊ฐ€ ๋น„๊ต์  ๋น ๋ฅด๊ฒŒ ์ตœ์ƒ์˜ ์„ฑ๋Šฅ์— ์ˆ˜๋ ดํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
* _Suggestions_: LLMs and small models have their own merits in different aspects: LLMs are can provide unified solutions to various NLP tasks and achieve competitive performance (especially in the zero/few-shot setting), while small models are economical to develop and can be specially tuned according to target tasks, which can achieve good performance with sufficient high-quality labeled data [770, 771, 754, 770, 771]. In applications, one can make suitable choices based on the actual needs, comprehensively considering flexibility, data availability, training compute, and efficiency. * _Future direction:_ Despite the excellent general capacities, LLMs still cannot effectively process the NLP tasks in low-resource domains, _e.g._, minor language translation. To tackle such tasks, it needs to develop effective approaches to injecting necessary task information or domain-specific knowledge into LLMs, either through fine-tuning or prompting. In addition, it is still challenging for LLMs to handle complex semantic relations in classic NLP tasks (_e.g._, nested entity extraction), which is worth more exploration from the underlying working mechanism of LLMs. It is also promising to combine LLMs and fine-tuned small language models for complementing with each other in solving complex cases of classic NLP tasks [772]. Another promising direction is to conduct human-machine collaborative research (_e.g._, conversational translation [768]) on NLP tasks, since LLMs can effectively understand human instructions and make meaningful responses.
* _์ œ์•ˆ ์‚ฌํ•ญ_: LLM๊ณผ ์†Œํ˜• ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๊ณ ์œ ํ•œ ์žฅ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค: LLM์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์— ๋Œ€ํ•œ ํ†ตํ•ฉ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜๊ณ  ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ(ํŠนํžˆ zero/few-shot ์„ค์ •)์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์†Œํ˜• ๋ชจ๋ธ์€ ๊ฐœ๋ฐœ ๋น„์šฉ์ด ์ €๋ ดํ•˜๊ณ  ๋Œ€์ƒ ์ž‘์—…์— ๋”ฐ๋ผ ํŠน๋ณ„ํžˆ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์–ด ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ๋ง๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ถฉ๋ถ„ํ•˜๋ฉด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[770, 771, 754, 770, 771]. ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ๋Š” ์œ ์—ฐ์„ฑ, ๋ฐ์ดํ„ฐ ๊ฐ€์šฉ์„ฑ, ํ›ˆ๋ จ ์ปดํ“จํŒ… ๋ฐ ํšจ์œจ์„ฑ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๊ณ ๋ คํ•˜์—ฌ ์‹ค์ œ ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ์„ ํƒ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * _๋ฏธ๋ž˜ ๋ฐฉํ–ฅ:_ ๋›ฐ์–ด๋‚œ ์ผ๋ฐ˜์ ์ธ ์—ญ๋Ÿ‰์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์€ ์—ฌ์ „ํžˆ ์ €์ž์› ๋„๋ฉ”์ธ์˜ NLP ์ž‘์—…(_์˜ˆ: ์†Œ์ˆ˜ ์–ธ์–ด ๋ฒˆ์—ญ)์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด Fine-tuning ๋˜๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด LLM์— ํ•„์š”ํ•œ ์ž‘์—… ์ •๋ณด ๋˜๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์ฃผ์ž…ํ•˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM์ด ๊ณ ์ „์ ์ธ NLP ์ž‘์—…์—์„œ ๋ณต์žกํ•œ ์˜๋ฏธ ๊ด€๊ณ„(_์˜ˆ: ์ค‘์ฒฉ๋œ ์—”ํ‹ฐํ‹ฐ ์ถ”์ถœ)๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๊ณผ์ œ์ด๋ฉฐ, ์ด๋Š” LLM์˜ ๊ธฐ๋ณธ ์ž‘๋™ ๋ฉ”์ปค๋‹ˆ์ฆ˜์—์„œ ๋” ๋งŽ์€ ํƒ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM๊ณผ Fine-tuned ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ณ ์ „์ ์ธ NLP ์ž‘์—…์˜ ๋ณต์žกํ•œ ์‚ฌ๋ก€๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์„œ๋กœ๋ฅผ ๋ณด์™„ํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค[772]. ๋˜ ๋‹ค๋ฅธ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์€ LLM์ด ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์˜๋ฏธ ์žˆ๋Š” ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— NLP ์ž‘์—…์— ๋Œ€ํ•œ ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„์˜ ํ˜‘์—… ์—ฐ๊ตฌ(์˜ˆ: ๋Œ€ํ™”ํ˜• ๋ฒˆ์—ญ[768])๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
* _์ œ์•ˆ ์‚ฌํ•ญ_: LLM๊ณผ ์†Œํ˜• ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ธก๋ฉด์—์„œ ๊ณ ์œ ํ•œ ์žฅ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค: LLM์€ ๋‹ค์–‘ํ•œ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ์ž‘์—…์— ๋Œ€ํ•œ ํ†ตํ•ฉ ์†”๋ฃจ์…˜์„ ์ œ๊ณตํ•˜๊ณ  ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ์„ฑ๋Šฅ(ํŠนํžˆ ์ œ๋กœ/์†Œ๋Ÿ‰ ์ƒท ์„ค์ •)์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์†Œํ˜• ๋ชจ๋ธ์€ ๊ฐœ๋ฐœ ๋น„์šฉ์ด ์ €๋ ดํ•˜๊ณ  ๋Œ€์ƒ ์ž‘์—…์— ๋”ฐ๋ผ ํŠน๋ณ„ํžˆ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ์–ด ๊ณ ํ’ˆ์งˆ์˜ ๋ผ๋ฒจ๋ง๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ถฉ๋ถ„ํ•˜๋ฉด ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[770, 771, 754, 770, 771]. ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ๋Š” ์œ ์—ฐ์„ฑ, ๋ฐ์ดํ„ฐ ๊ฐ€์šฉ์„ฑ, ํ›ˆ๋ จ ์ปดํ“จํŒ… ๋ฐ ํšจ์œจ์„ฑ์„ ์ข…ํ•ฉ์ ์œผ๋กœ ๊ณ ๋ คํ•˜์—ฌ ์‹ค์ œ ์š”๊ตฌ ์‚ฌํ•ญ์— ๋”ฐ๋ผ ์ ํ•ฉํ•œ ์„ ํƒ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * _๋ฏธ๋ž˜ ๋ฐฉํ–ฅ:_ ๋›ฐ์–ด๋‚œ ์ผ๋ฐ˜์ ์ธ ์—ญ๋Ÿ‰์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  LLM์€ ์—ฌ์ „ํžˆ ์ €์ž์› ๋„๋ฉ”์ธ์˜ NLP ์ž‘์—…(_์˜ˆ: ์†Œ์ˆ˜ ์–ธ์–ด ๋ฒˆ์—ญ)์„ ํšจ๊ณผ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๋ ค๋ฉด ๋ฏธ์„ธ ์กฐ์ • ๋˜๋Š” ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด LLM์— ํ•„์š”ํ•œ ์ž‘์—… ์ •๋ณด ๋˜๋Š” ๋„๋ฉ”์ธ๋ณ„ ์ง€์‹์„ ์ฃผ์ž…ํ•˜๋Š” ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, LLM์ด ๊ณ ์ „์ ์ธ NLP ์ž‘์—…์—์„œ ๋ณต์žกํ•œ ์˜๋ฏธ ๊ด€๊ณ„(_์˜ˆ: ์ค‘์ฒฉ๋œ ์—”ํ‹ฐํ‹ฐ ์ถ”์ถœ)๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์€ ์—ฌ์ „ํžˆ ์–ด๋ ค์šด ๊ณผ์ œ์ด๋ฉฐ, ์ด๋Š” LLM์˜ ๊ธฐ๋ณธ ์ž‘๋™ ๋ฉ”์ปค๋‹ˆ์ฆ˜์—์„œ ๋” ๋งŽ์€ ํƒ๊ตฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ LLM๊ณผ ๋ฏธ์„ธ ์กฐ์ •๋œ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ๊ณ ์ „์ ์ธ NLP ์ž‘์—…์˜ ๋ณต์žกํ•œ ์‚ฌ๋ก€๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ์„œ๋กœ๋ฅผ ๋ณด์™„ํ•˜๋Š” ๊ฒƒ๋„ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค[772]. ๋˜ ๋‹ค๋ฅธ ์œ ๋งํ•œ ๋ฐฉํ–ฅ์€ LLM์ด ์ธ๊ฐ„์˜ ์ง€์‹œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ดํ•ดํ•˜๊ณ  ์˜๋ฏธ ์žˆ๋Š” ์‘๋‹ต์„ ํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— NLP ์ž‘์—…์— ๋Œ€ํ•œ ์ธ๊ฐ„๊ณผ ๊ธฐ๊ณ„์˜ ํ˜‘์—… ์—ฐ๊ตฌ(์˜ˆ: ๋Œ€ํ™”ํ˜• ๋ฒˆ์—ญ[768])๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
* Harmful and unpredictable content: Though our model can reject unethical queries, these models may still generate harmful or misaligned with human preferences and values. This issue may arise from biases in the training data or the models' inability to discern appropriate outputs in certain contexts. * Insufficient training: Due to constraints in computing power and data availability, the training of the models may not be sufficient for optimal performance. As a result, there is still room for improvement in the Chinese understanding capabilities of the models. * Lack of robustness: The models may exhibit brittleness in some situations, producing inconsistent or nonsensical outputs when faced with adversarial inputs or rare language phenomena. * Comprehensive evaluation: Evaluating large language models is an important topic in the current era. While we have seen many evaluation benchmarks for LLMs, their comprehensiveness and appropriateness for LLMs should be well-studied and examined. A more diverse and comprehensive LLM evaluation dataset and benchmark will have a great positive effect on shaping the future of LLM research. * Scalability and efficiency: Although we applied LoRA and quantization to make the model more accessible to a broader community, when combined with the original LLaMA, the models' large size and complexity can lead to difficulties in deployment, especially for users with limited computational resources. This issue may hinder the accessibility and widespread adoption of the models in various applications.
* ์œ ํ•ดํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ์ฝ˜ํ…์ธ : ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋น„์œค๋ฆฌ์ ์ธ ์ฟผ๋ฆฌ๋ฅผ ๊ฑฐ๋ถ€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์œ ํ•ดํ•˜๊ฑฐ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ๋ฐ ๊ฐ€์น˜๊ด€์— ๋งž์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์ด๋‚˜ ํŠน์ • ์ƒํ™ฉ์—์„œ ์ ์ ˆํ•œ ์ถœ๋ ฅ์„ ๊ตฌ๋ถ„ํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๋ถˆ์ถฉ๋ถ„ํ•œ ํ•™์Šต: ์ปดํ“จํŒ… ์„ฑ๋Šฅ ๋ฐ ๋ฐ์ดํ„ฐ ๊ฐ€์šฉ์„ฑ์˜ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๋ชจ๋ธ ํ•™์Šต์ด ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋ชจ๋ธ์˜ ์ค‘๊ตญ์–ด ์ดํ•ด ์—ญ๋Ÿ‰์€ ์—ฌ์ „ํžˆ ๊ฐœ์„ ์˜ ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. * Lack of robustness: ๋ชจ๋ธ์€ ์ผ๋ถ€ ์ƒํ™ฉ์—์„œ ์ทจ์•ฝ์„ฑ์„ ๋ณด์—ฌ ์ ๋Œ€์ ์ธ ์ž…๋ ฅ์ด๋‚˜ ํฌ๊ท€ ์–ธ์–ด ํ˜„์ƒ์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ์ผ๊ด€์„ฑ์ด ์—†๊ฑฐ๋‚˜ ๋ฌด์˜๋ฏธํ•œ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * Comprehensive evaluation: ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ํ˜„์žฌ ์‹œ๋Œ€์˜ ์ค‘์š”ํ•œ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค. LLM์— ๋Œ€ํ•œ ๋งŽ์€ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋“ฑ์žฅํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ์˜ ํฌ๊ด„์„ฑ๊ณผ ์ ์ ˆ์„ฑ์€ ์ž˜ ์—ฐ๊ตฌ๋˜๊ณ  ๊ฒ€ํ† ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ๋‹ค์–‘ํ•˜๊ณ  ํฌ๊ด„์ ์ธ LLM ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ๋Š” LLM ์—ฐ๊ตฌ์˜ ๋ฏธ๋ž˜๋ฅผ ํ˜•์„ฑํ•˜๋Š” ๋ฐ ํฐ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์ž…๋‹ˆ๋‹ค. * Scalability ๋ฐ efficiency: ๋” ๋งŽ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ๋ชจ๋ธ์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก LoRA์™€ Quantization์„ ์ ์šฉํ–ˆ์ง€๋งŒ, ์›๋ž˜์˜ LLaMA์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์˜ ํฌ๊ธฐ์™€ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ํŠนํžˆ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์‚ฌ์šฉ์ž์˜ ๊ฒฝ์šฐ ๋ฐฐํฌ์— ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ฑ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ์ฑ„ํƒ์„ ๋ฐฉํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
* ์œ ํ•ดํ•˜๊ณ  ์˜ˆ์ธกํ•  ์ˆ˜ ์—†๋Š” ์ฝ˜ํ…์ธ : ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ๋น„์œค๋ฆฌ์ ์ธ ์ฟผ๋ฆฌ๋ฅผ ๊ฑฐ๋ถ€ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ์—ฌ์ „ํžˆ ์œ ํ•ดํ•˜๊ฑฐ๋‚˜ ์ธ๊ฐ„์˜ ์„ ํ˜ธ๋„ ๋ฐ ๊ฐ€์น˜๊ด€์— ๋งž์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํŽธํ–ฅ์„ฑ์ด๋‚˜ ํŠน์ • ์ƒํ™ฉ์—์„œ ์ ์ ˆํ•œ ์ถœ๋ ฅ์„ ๊ตฌ๋ถ„ํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ๋ถˆ์ถฉ๋ถ„ํ•œ ํ•™์Šต: ์ปดํ“จํŒ… ์„ฑ๋Šฅ ๋ฐ ๋ฐ์ดํ„ฐ ๊ฐ€์šฉ์„ฑ์˜ ์ œ์•ฝ์œผ๋กœ ์ธํ•ด ๋ชจ๋ธ ํ•™์Šต์ด ์ตœ์ ์˜ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ๋ชจ๋ธ์˜ ์ค‘๊ตญ์–ด ์ดํ•ด ์—ญ๋Ÿ‰์€ ์—ฌ์ „ํžˆ ๊ฐœ์„ ์˜ ์—ฌ์ง€๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. * ๊ฒฌ๊ณ ์„ฑ ๋ถ€์กฑ: ๋ชจ๋ธ์€ ์ผ๋ถ€ ์ƒํ™ฉ์—์„œ ์ทจ์•ฝ์„ฑ์„ ๋ณด์—ฌ ์ ๋Œ€์ ์ธ ์ž…๋ ฅ์ด๋‚˜ ํฌ๊ท€ ์–ธ์–ด ํ˜„์ƒ์— ์ง๋ฉดํ–ˆ์„ ๋•Œ ์ผ๊ด€์„ฑ์ด ์—†๊ฑฐ๋‚˜ ๋ฌด์˜๋ฏธํ•œ ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. * ์ข…ํ•ฉ์ ์ธ ํ‰๊ฐ€: ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์€ ํ˜„์žฌ ์‹œ๋Œ€์˜ ์ค‘์š”ํ•œ ์ฃผ์ œ์ž…๋‹ˆ๋‹ค. LLM์— ๋Œ€ํ•œ ๋งŽ์€ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๊ฐ€ ๋“ฑ์žฅํ–ˆ์ง€๋งŒ, ์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ์˜ ํฌ๊ด„์„ฑ๊ณผ ์ ์ ˆ์„ฑ์€ ์ž˜ ์—ฐ๊ตฌ๋˜๊ณ  ๊ฒ€ํ† ๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด๋‹ค ๋‹ค์–‘ํ•˜๊ณ  ํฌ๊ด„์ ์ธ LLM ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ๋Š” LLM ์—ฐ๊ตฌ์˜ ๋ฏธ๋ž˜๋ฅผ ํ˜•์„ฑํ•˜๋Š” ๋ฐ ํฐ ๊ธ์ •์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์ž…๋‹ˆ๋‹ค. * ํ™•์žฅ์„ฑ ๋ฐ ํšจ์œจ์„ฑ: ๋” ๋งŽ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ๋ชจ๋ธ์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก LoRA์™€ ์–‘์žํ™”๋ฅผ ์ ์šฉํ–ˆ์ง€๋งŒ, ์›๋ž˜์˜ LLaMA์™€ ๊ฒฐํ•ฉํ•˜๋ฉด ๋ชจ๋ธ์˜ ํฌ๊ธฐ์™€ ๋ณต์žก์„ฑ์œผ๋กœ ์ธํ•ด ํŠนํžˆ ๊ณ„์‚ฐ ๋ฆฌ์†Œ์Šค๊ฐ€ ์ œํ•œ๋œ ์‚ฌ์šฉ์ž์˜ ๊ฒฝ์šฐ ๋ฐฐํฌ์— ์–ด๋ ค์›€์„ ๊ฒช์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋Š” ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ ‘๊ทผ์„ฑ๊ณผ ๊ด‘๋ฒ”์œ„ํ•œ ์ฑ„ํƒ์„ ๋ฐฉํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
the latest knowledge beyond the training data. To tackle this issue, a straightforward approach is to regularly update LLMs with new data. However, it is very costly to fine-tune LLMs, and also likely to cause the catastrophic forgetting issue when incrementally training LLMs. Therefore, it is necessary to develop efficient and effective approaches that can integrate new knowledge into existing LLMs, making them up-to-date. Existing studies have explored how to utilize the external knowledge source (_e.g.,_ search engine) to complement LLMs, which can be either jointly optimized with LLMs [653] or used as a plug-and-play module [659]. For instance, ChatGPT utilizes a retrieval plugin to access up-to-date information sources [665]. By incorporating the extracted relevant information into the context [666, 667, 668], LLMs can acquire new factual knowledge and perform better on relevant tasks. However, such an approach seems to be still at a superficial level. In addition, existing studies also explore editing parameters of language models to update intrinsic knowledge [669, 670, 671]. Nevertheless, previous work [672] has shown that several parameter editing methods perform not well on LLMs, though they can improve the performance of small language models. Therefore, it is still difficult to directly amend intrinsic knowledge or inject specific knowledge into LLMs, which remains an open research problem [672]. Recently, a useful framework _EasyEdit_[673] has been released to facilitate the research of knowledge editing for LLMs.
ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๋„˜์–ด์„œ๋Š” ์ตœ์‹  ์ง€์‹์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •๊ธฐ์ ์œผ๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ LLM์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLM์„ Fine-tuningํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๋ฉฐ, LLM์„ ์ ์ง„์ ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ์น˜๋ช…์ ์ธ ๋ง๊ฐ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์กด LLM์— ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์ตœ์‹  ์ƒํƒœ๋กœ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์™ธ๋ถ€ ์ง€์‹ ์†Œ์Šค(_์˜ˆ:_ ๊ฒ€์ƒ‰ ์—”์ง„)๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ณด์™„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” LLM๊ณผ ๊ณต๋™ ์ตœ์ ํ™”ํ•˜๊ฑฐ๋‚˜ ํ”Œ๋Ÿฌ๊ทธ ์•ค ํ”Œ๋ ˆ์ด ๋ชจ๋“ˆ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653]. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatGPT๋Š” ๊ฒ€์ƒ‰ ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์‹  ์ •๋ณด ์†Œ์Šค์— ์•ก์„ธ์Šคํ•ฉ๋‹ˆ๋‹ค[665]. ์ถ”์ถœ๋œ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ปจํ…์ŠคํŠธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ[666, 667, 668], LLM์€ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์  ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ๊ด€๋ จ ์ž‘์—…์„ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์•„์ง ํ”ผ์ƒ์ ์ธ ์ˆ˜์ค€์— ๋จธ๋ฌผ๋Ÿฌ ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํŽธ์ง‘ํ•˜์—ฌ ๋‚ด์žฌ์  ์ง€์‹์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐฉ์‹๋„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[669, 670, 671]. ๊ทธ๋Ÿฌ๋‚˜ ์ด์ „ ์—ฐ๊ตฌ[672]์— ๋”ฐ๋ฅด๋ฉด, ์—ฌ๋Ÿฌ ๋งค๊ฐœ๋ณ€์ˆ˜ ํŽธ์ง‘ ๋ฐฉ๋ฒ•์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ LLM์—์„œ๋Š” ์„ฑ๋Šฅ์ด ์ข‹์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ฌ์ „ํžˆ ๋‚ด์žฌ์  ์ง€์‹์„ ์ง์ ‘ ์ˆ˜์ •ํ•˜๊ฑฐ๋‚˜ LLM์— ํŠน์ • ์ง€์‹์„ ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šด ๋ฌธ์ œ์ด๋ฉฐ, ์ด๋Š” ์•„์ง ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค[672]. ์ตœ๊ทผ์—๋Š” LLM์— ๋Œ€ํ•œ ์ง€์‹ ํŽธ์ง‘ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ์œ ์šฉํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ _EasyEdit_[673]์ด ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๋„˜์–ด์„œ๋Š” ์ตœ์‹  ์ง€์‹์„ ์Šต๋“ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ •๊ธฐ์ ์œผ๋กœ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ LLM์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๋ฐ๋Š” ๋งŽ์€ ๋น„์šฉ์ด ๋“ค๋ฉฐ, LLM์„ ์ ์ง„์ ์œผ๋กœ ํ›ˆ๋ จํ•  ๋•Œ ์น˜๋ช…์ ์ธ ๋ง๊ฐ ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ธฐ์กด LLM์— ์ƒˆ๋กœ์šด ์ง€์‹์„ ํ†ตํ•ฉํ•˜์—ฌ ์ตœ์‹  ์ƒํƒœ๋กœ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋Š” ํšจ์œจ์ ์ด๊ณ  ํšจ๊ณผ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์„ ๊ฐœ๋ฐœํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์™ธ๋ถ€ ์ง€์‹ ์†Œ์Šค(_์˜ˆ:_ ๊ฒ€์ƒ‰ ์—”์ง„)๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ณด์™„ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ํƒ๊ตฌํ–ˆ์œผ๋ฉฐ, ์ด๋Š” LLM๊ณผ ๊ณต๋™ ์ตœ์ ํ™”ํ•˜๊ฑฐ๋‚˜ ํ”Œ๋Ÿฌ๊ทธ ์•ค ํ”Œ๋ ˆ์ด ๋ชจ๋“ˆ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[653]. ์˜ˆ๋ฅผ ๋“ค์–ด, ChatGPT๋Š” ๊ฒ€์ƒ‰ ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์‹  ์ •๋ณด ์†Œ์Šค์— ์•ก์„ธ์Šคํ•ฉ๋‹ˆ๋‹ค[665]. ์ถ”์ถœ๋œ ๊ด€๋ จ ์ •๋ณด๋ฅผ ์ปจํ…์ŠคํŠธ์— ํ†ตํ•ฉํ•จ์œผ๋กœ์จ[666, 667, 668], LLM์€ ์ƒˆ๋กœ์šด ์‚ฌ์‹ค์  ์ง€์‹์„ ์Šต๋“ํ•˜๊ณ  ๊ด€๋ จ ์ž‘์—…์„ ๋” ์ž˜ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์•„์ง ํ”ผ์ƒ์ ์ธ ์ˆ˜์ค€์— ๋จธ๋ฌผ๋Ÿฌ ์žˆ๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์˜ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํŽธ์ง‘ํ•˜์—ฌ ๋‚ด์žฌ์  ์ง€์‹์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐฉ์‹๋„ ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[669, 670, 671]. ๊ทธ๋Ÿฌ๋‚˜ ์ด์ „ ์—ฐ๊ตฌ[672]์— ๋”ฐ๋ฅด๋ฉด, ์—ฌ๋Ÿฌ ๋งค๊ฐœ๋ณ€์ˆ˜ ํŽธ์ง‘ ๋ฐฉ๋ฒ•์€ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ์ง€๋งŒ LLM์—์„œ๋Š” ์„ฑ๋Šฅ์ด ์ข‹์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ฌ์ „ํžˆ ๋‚ด์žฌ์  ์ง€์‹์„ ์ง์ ‘ ์ˆ˜์ •ํ•˜๊ฑฐ๋‚˜ LLM์— ํŠน์ • ์ง€์‹์„ ์ฃผ์ž…ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šด ๋ฌธ์ œ์ด๋ฉฐ, ์ด๋Š” ์•„์ง ๋ฏธํ•ด๊ฒฐ ๊ณผ์ œ๋กœ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค[672]. ์ตœ๊ทผ์—๋Š” LLM์— ๋Œ€ํ•œ ์ง€์‹ ํŽธ์ง‘ ์—ฐ๊ตฌ๋ฅผ ์ด‰์ง„ํ•˜๊ธฐ ์œ„ํ•ด ์œ ์šฉํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์ธ _EasyEdit_[673]์ด ์ถœ์‹œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Auxiliary optimization objectives._ Besides the primary cross-entropy loss, several studies propose auxiliary training loss to enhance the learning from the alignment data. First, since the responses of each instruction can be scored by the reward model, the ranking loss can be used to train the model to preserve the ranking order of these responses. For example, RRHF [390] samples responses from multiple sources, including model-generated responses, such as those derived from the model itself, ChatGPT, and GPT-4, as well as human-written responses, spanning both high-quality and low-quality instances. To align with the scores from reward models, it further optimizes the ranking loss by encouraging the model to have a higher conditional log probability for the response with a higher ranking. SLiCPF [391] proposes to assess the similarity between model outputs and human preference via the distance in the latent space, and introduces specific calibration and regularization loss to calibrate the candidate sequences based on human-preference data. Second, to enhance the relatedness between the response and the instruction, some work adopts contrastive learning to push up the probability of correct instruction-response pairs while pushing down incorrect instruction-response pairs. Specifically, for an output response, the proposed approach in [392] contrasts the target instruction to the other irrelevant instructions. By doing so, it can enable the model to learn the right correlation between instructions and responses.
\(\bullet\)_Auxiliary optimization objectives._ ๊ธฐ๋ณธ cross-entropy loss ์™ธ์—๋„ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ์˜ ํ•™์Šต์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋ณด์กฐ ํ›ˆ๋ จ loss๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ฐ instruction์— ๋Œ€ํ•œ ์‘๋‹ต์€ ๋ณด์ƒ ๋ชจ๋ธ์— ์˜ํ•ด ์ ์ˆ˜๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ranking loss๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์‘๋‹ต์˜ ์ˆœ์œ„ ์ˆœ์„œ๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, RRHF [390]๋Š” ๋ชจ๋ธ ์ž์ฒด์—์„œ ํŒŒ์ƒ๋œ ๋ชจ๋ธ ์ƒ์„ฑ ์‘๋‹ต, ChatGPT, GPT-4 ๋“ฑ์˜ ๋ชจ๋ธ ์ƒ์„ฑ ์‘๋‹ต๊ณผ ๊ณ ํ’ˆ์งˆ ๋ฐ ์ €ํ’ˆ์งˆ ์ธ์Šคํ„ด์Šค์— ๊ฑธ์ณ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์‘๋‹ต์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ์†Œ์Šค์˜ ์‘๋‹ต์„ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ƒ ๋ชจ๋ธ์˜ ์ ์ˆ˜์™€ ์ผ์น˜ํ•˜๋„๋ก ๋ชจ๋ธ์ด ๋” ๋†’์€ ์ˆœ์œ„๋ฅผ ๊ฐ€์ง„ ์‘๋‹ต์— ๋Œ€ํ•ด ์กฐ๊ฑด๋ถ€ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋†’์ด๋„๋ก ์žฅ๋ คํ•˜์—ฌ ranking loss๋ฅผ ์ถ”๊ฐ€๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. SLiCPF [391]๋Š” ์ž ์žฌ ๊ณต๊ฐ„์˜ ๊ฑฐ๋ฆฌ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ์ถœ๋ ฅ๊ณผ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ณ , ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ›„๋ณด ์‹œํ€€์Šค๋ฅผ ๋ณด์ •ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ๋ณด์ • ๋ฐ ์ •๊ทœํ™” loss๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‘๋‹ต๊ณผ instruction ๊ฐ„์˜ ๊ด€๋ จ์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” contrastive learning์„ ์ฑ„ํƒํ•˜์—ฌ ์˜ฌ๋ฐ”๋ฅธ instruction-response ์Œ์˜ ํ™•๋ฅ ์„ ๋†’์ด๊ณ  ์ž˜๋ชป๋œ instruction-response ์Œ์˜ ํ™•๋ฅ ์„ ๋‚ฎ์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, [392]์—์„œ ์ œ์•ˆ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ถœ๋ ฅ๋œ ์‘๋‹ต์— ๋Œ€ํ•ด ๋Œ€์ƒ instruction์„ ๋‹ค๋ฅธ ๊ด€๋ จ ์—†๋Š” instruction๊ณผ ๋Œ€์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด instruction๊ณผ ์‘๋‹ต ๊ฐ„์˜ ์˜ฌ๋ฐ”๋ฅธ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_๋ณด์กฐ ์ตœ์ ํ™” ๋ชฉํ‘œ._ ๊ธฐ๋ณธ ๊ต์ฐจ ์—”ํŠธ๋กœํ”ผ ์†์‹ค ์™ธ์—๋„ ์—ฌ๋Ÿฌ ์—ฐ๊ตฌ์—์„œ ์ •๋ ฌ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ์˜ ํ•™์Šต์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๋ณด์กฐ ํ›ˆ๋ จ ์†์‹ค์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ฐ ๋ช…๋ น์— ๋Œ€ํ•œ ์‘๋‹ต์€ ๋ณด์ƒ ๋ชจ๋ธ์— ์˜ํ•ด ์ ์ˆ˜๋ฅผ ๋งค๊ธธ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ์ˆœ์œ„ ์†์‹ค์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์ด ์ด๋Ÿฌํ•œ ์‘๋‹ต์˜ ์ˆœ์œ„ ์ˆœ์„œ๋ฅผ ์œ ์ง€ํ•˜๋„๋ก ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, RRHF [390]๋Š” ๋ชจ๋ธ ์ž์ฒด์—์„œ ํŒŒ์ƒ๋œ ๋ชจ๋ธ ์ƒ์„ฑ ์‘๋‹ต, ChatGPT, GPT-4 ๋“ฑ์˜ ๋ชจ๋ธ ์ƒ์„ฑ ์‘๋‹ต๊ณผ ๊ณ ํ’ˆ์งˆ ๋ฐ ์ €ํ’ˆ์งˆ ์ธ์Šคํ„ด์Šค์— ๊ฑธ์ณ ์‚ฌ๋žŒ์ด ์ž‘์„ฑํ•œ ์‘๋‹ต์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ์†Œ์Šค์˜ ์‘๋‹ต์„ ์ƒ˜ํ”Œ๋งํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ƒ ๋ชจ๋ธ์˜ ์ ์ˆ˜์™€ ์ผ์น˜ํ•˜๋„๋ก ๋ชจ๋ธ์ด ๋” ๋†’์€ ์ˆœ์œ„๋ฅผ ๊ฐ€์ง„ ์‘๋‹ต์— ๋Œ€ํ•ด ์กฐ๊ฑด๋ถ€ ๋กœ๊ทธ ํ™•๋ฅ ์„ ๋†’์ด๋„๋ก ์žฅ๋ คํ•˜์—ฌ ์ˆœ์œ„ ์†์‹ค์„ ์ถ”๊ฐ€๋กœ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค. SLiCPF [391]๋Š” ์ž ์žฌ ๊ณต๊ฐ„์˜ ๊ฑฐ๋ฆฌ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ ์ถœ๋ ฅ๊ณผ ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ํ‰๊ฐ€ํ•˜๊ณ , ์‚ฌ๋žŒ์˜ ์„ ํ˜ธ๋„ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ›„๋ณด ์‹œํ€€์Šค๋ฅผ ๋ณด์ •ํ•˜๊ธฐ ์œ„ํ•ด ํŠน์ • ๋ณด์ • ๋ฐ ์ •๊ทœํ™” ์†์‹ค์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‘๋‹ต๊ณผ ๋ช…๋ น์–ด ๊ฐ„์˜ ๊ด€๋ จ์„ฑ์„ ๊ฐ•ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์ผ๋ถ€ ์ž‘์—…์—์„œ๋Š” ๋Œ€์กฐ ํ•™์Šต์„ ์ฑ„ํƒํ•˜์—ฌ ์˜ฌ๋ฐ”๋ฅธ ๋ช…๋ น์–ด-์‘๋‹ต์–ด ์Œ์˜ ํ™•๋ฅ ์„ ๋†’์ด๊ณ  ์ž˜๋ชป๋œ ๋ช…๋ น์–ด-์‘๋‹ต์–ด ์Œ์˜ ํ™•๋ฅ ์„ ๋‚ฎ์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ, [392]์—์„œ ์ œ์•ˆ๋œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ถœ๋ ฅ๋œ ์‘๋‹ต์— ๋Œ€ํ•ด ๋Œ€์ƒ ๋ช…๋ น์–ด๋ฅผ ๋‹ค๋ฅธ ๊ด€๋ จ ์—†๋Š” ๋ช…๋ น์–ด์™€ ๋Œ€์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด ๋ช…๋ น์–ด์™€ ์‘๋‹ต ๊ฐ„์˜ ์˜ฌ๋ฐ”๋ฅธ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Large language models are trained on massive scrapes of the web, which are often unstructured, noisy, and poorly phrased. Current scaling laws show that learning from such data requires an abundance of both compute and data, which grows with the size of the model being trained. This is infeasible both because of the large compute costs and duration associated with pre-training, and the impending scarcity of high-quality data on the web. In this work, we propose **Web** Rephrase **A**ugmented **P**re-training (**WRAP**) that uses an off-the-shelf instruction-tuned model prompted to paraphrase documents on the web in specific styles such as "like Wikipedia" or in "question-answer format" to jointly pre-train LLMs on real and synthetic rephrases. First, we show that using **WRAP** on the C4 dataset, which is naturally noisy, speeds up pre-training by \(\sim 3\times\). At the same pre-training compute budget, it improves perplexity by more than 10% on average across different subsets of the Pile, and improves zero-shot question answer accuracy across 13 tasks by more than 2%. Second, we investigate the impact of the re-phrasing style on the performance of the model, offering insights into how the composition of the training data can impact the performance of LLMs in OOD settings. Our gains are attributed to the fact that re-phrased synthetic data has higher utility than just real data because it (i) incorporates style diversity that closely reflects downstream evaluation style, and (ii) has higher 'quality' than web-scraped data.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ข…์ข… ๋น„์ •ํ˜•์ ์ด๊ณ  ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์ž˜๋ชป ํ‘œํ˜„๋œ ์›น ์Šคํฌ๋žฉ์„ ๋Œ€๋Ÿ‰์œผ๋กœ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ scaling laws์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋ ค๋ฉด ํ•™์Šต ์ค‘์ธ ๋ชจ๋ธ์˜ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•˜๋Š” compute์™€ data๊ฐ€ ๋ชจ๋‘ ํ’๋ถ€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” pre-training๊ณผ ๊ด€๋ จ๋œ ๋ง‰๋Œ€ํ•œ compute ๋น„์šฉ๊ณผ ๊ธฐ๊ฐ„, ์›น์—์„œ ๊ณ ํ’ˆ์งˆ data์˜ ์ž„๋ฐ•ํ•œ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์›น์ƒ์˜ ๋ฌธ์„œ๋ฅผ '์œ„ํ‚ค๋ฐฑ๊ณผ์ฒ˜๋Ÿผ' ๋˜๋Š” '์งˆ๋ฌธ-๋‹ต๋ณ€ ํ˜•์‹'๊ณผ ๊ฐ™์€ ํŠน์ • ์Šคํƒ€์ผ๋กœ ์˜์—ญํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋œ off-the-shelf Instruction-tuned ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹ค์ œ ๋ฐ ํ•ฉ์„ฑ ์˜์—ญ์— ๋Œ€ํ•ด LLM์„ ๊ณต๋™์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” **Web** ์˜์—ญ **A**ugmented **P**re-training(**WRAP**)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, ์ž์—ฐ์ ์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์€ C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ์— **WRAP**์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ์†๋„๊ฐ€ \(3๋ฐฐ\) ๊ฐ€์†ํ™”๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต compute ์˜ˆ์‚ฐ์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ์ž์—ฐ์ ์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์€ C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ์— **WRAP**์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ์†๋„๊ฐ€ \(3๋ฐฐ\) ๊ฐ€์†ํ™”๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต compute ์˜ˆ์‚ฐ์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์ข…์ข… ๋น„์ •ํ˜•์ ์ด๊ณ  ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์œผ๋ฉฐ ์ž˜๋ชป ํ‘œํ˜„๋œ ์›น ์Šคํฌ๋žฉ์„ ๋Œ€๋Ÿ‰์œผ๋กœ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฌ์˜ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์— ๋”ฐ๋ฅด๋ฉด ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•˜๋ ค๋ฉด ํ•™์Šต ์ค‘์ธ ๋ชจ๋ธ์˜ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ฆ๊ฐ€ํ•˜๋Š” ์ปดํ“จํŒ…๊ณผ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋‘ ํ’๋ถ€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ์‚ฌ์ „ ํ•™์Šต๊ณผ ๊ด€๋ จ๋œ ๋ง‰๋Œ€ํ•œ ์ปดํ“จํŒ… ๋น„์šฉ๊ณผ ๊ธฐ๊ฐ„, ์›น์—์„œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์˜ ์ž„๋ฐ•ํ•œ ๋ถ€์กฑ์œผ๋กœ ์ธํ•ด ๋ถˆ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์›น์ƒ์˜ ๋ฌธ์„œ๋ฅผ '์œ„ํ‚ค๋ฐฑ๊ณผ์ฒ˜๋Ÿผ' ๋˜๋Š” '์งˆ๋ฌธ-๋‹ต๋ณ€ ํ˜•์‹'๊ณผ ๊ฐ™์€ ํŠน์ • ์Šคํƒ€์ผ๋กœ ์˜์—ญํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋œ ์˜คํ”„๋” ์…ธํ”„ ์ง€์‹œ ํŠœ๋‹ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹ค์ œ ๋ฐ ํ•ฉ์„ฑ ์˜์—ญ์— ๋Œ€ํ•ด LLM์„ ๊ณต๋™์œผ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋Š” **Web** ์˜์—ญ **A**ugmented **P**re-training(**WRAP**)์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, ์ž์—ฐ์ ์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์€ C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ์— **WRAP**์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ์†๋„๊ฐ€ \(3๋ฐฐ\) ๊ฐ€์†ํ™”๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ปดํ“จํŒ… ์˜ˆ์‚ฐ์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ์ž์—ฐ์ ์œผ๋กœ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์€ C4 ๋ฐ์ดํ„ฐ ์„ธํŠธ์— **WRAP**์„ ์‚ฌ์šฉํ•˜๋ฉด ์‚ฌ์ „ ํ•™์Šต ์†๋„๊ฐ€ \(3๋ฐฐ\) ๊ฐ€์†ํ™”๋œ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋™์ผํ•œ ์‚ฌ์ „ ํ•™์Šต ์ปดํ“จํŒ… ์˜ˆ์‚ฐ์—์„œ ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ €, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋” ๋งŽ์€ ์–‘์˜ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜๋ฉด ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Single-path Reasoning_. In this strategy, the final task is decomposed into several intermediate steps. These steps are connected in a cascading manner, with each step leading to only one subsequent step. LLMs follow these steps to achieve the final goal. Specifically, Chain of Thought (CoT) [45] proposes inputting reasoning steps for solving complex problems into the prompt. These steps serve as examples to inspire LLMs to plan and act in a step-by-step manner. In this method, the plans are created based on the inspiration from the examples in the prompts. Zero-shot-CoT [46] enables LLMs to generate task reasoning processes by prompting them with trigger sentences like "think step by step". Unlike CoT, this method does not incorporate reasoning steps as examples in the prompts. Reprompting [47] involves checking whether each step meets the necessary prerequisites before generating a plan. If a step fails to meet the prerequisites, it introduces a prerequisite error message and prompts the LLM to regenerate the plan. ReWOO [48] introduces a paradigm of separating plans from external observations, where the agents first generate plans and obtain observations independently, and then combine them together to derive the final results. HuggingGPT [13] first decomposes the task into many sub-goals, and then solves each of them based on Huggingface. Different from CoT and Zero-shot-CoT, which outcome all the reasoning steps in a one-shot manner, ReWOO and HuggingGPT produce the results by accessing LLMs multiply times.
\(\bullet\)_๋‹จ์ผ ๊ฒฝ๋กœ ์ถ”๋ก _. ์ด ์ „๋žต์—์„œ๋Š” ์ตœ์ข… ์ž‘์—…์„ ์—ฌ๋Ÿฌ ์ค‘๊ฐ„ ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋Š” ๊ณ„๋‹จ์‹์œผ๋กœ ์—ฐ๊ฒฐ๋˜๋ฉฐ ๊ฐ ๋‹จ๊ณ„๋Š” ํ•˜๋‚˜์˜ ํ›„์† ๋‹จ๊ณ„๋กœ๋งŒ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. LLM์€ ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ผ ์ตœ์ข… ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ์ž…๋ ฅํ•˜๋Š” Chain of Thought (CoT) [45]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋Š” LLM์ด ๋‹จ๊ณ„๋ณ„ ๋ฐฉ์‹์œผ๋กœ ๊ณ„ํšํ•˜๊ณ  ํ–‰๋™ํ•˜๋„๋ก ์˜๊ฐ์„ ์ฃผ๋Š” ์˜ˆ์ œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ์˜ˆ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ณ„ํš์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. Zero-shot-CoT [46]๋Š” "๋‹จ๊ณ„๋ณ„๋กœ ์ƒ๊ฐํ•˜๋ผ"์™€ ๊ฐ™์€ ํŠธ๋ฆฌ๊ฑฐ ๋ฌธ์žฅ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด LLM์ด ์ž‘์—… ์ถ”๋ก  ๊ณผ์ •์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ํ”„๋กฌํ”„ํŠธ์— ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์˜ˆ์ œ๋กœ ํ†ตํ•ฉํ•˜์ง€ ์•Š์œผ๋ฉฐ, CoT์™€๋Š” ๋‹ค๋ฆ…๋‹ˆ๋‹ค. Reprompting [47]๋Š” ๊ฐ ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•œ ์ „์ œ ์กฐ๊ฑด์„ ์ถฉ์กฑํ•˜๋Š”์ง€ ํ™•์ธํ•œ ํ›„ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์–ด๋–ค ๋‹จ๊ณ„๊ฐ€ ์ „์ œ ์กฐ๊ฑด์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋ฉด ์ „์ œ ์กฐ๊ฑด ์˜ค๋ฅ˜ ๋ฉ”์‹œ์ง€๋ฅผ ํ‘œ์‹œํ•˜๊ณ  LLM์ด ๊ณ„ํš์„ ๋‹ค์‹œ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ReWOO [48]๋Š” ๊ณ„ํš๊ณผ ์™ธ๋ถ€ ๊ด€์ฐฐ์„ ๋ถ„๋ฆฌํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋„์ž…ํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  ๊ด€์ฐฐ์„ ๋…๋ฆฝ์ ์œผ๋กœ ์–ป์€ ๋‹ค์Œ ์ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. HuggingGPT [13]๋Š” ๋จผ์ € ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ Huggingface๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์›์ƒท ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” CoT ๋ฐ Zero-shot-CoT์™€ ๋‹ฌ๋ฆฌ, ReWOO์™€ HuggingGPT๋Š” LLM์— ์—ฌ๋Ÿฌ ๋ฒˆ ์•ก์„ธ์Šคํ•˜์—ฌ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_๋‹จ์ผ ๊ฒฝ๋กœ ์ถ”๋ก _. ์ด ์ „๋žต์—์„œ๋Š” ์ตœ์ข… ์ž‘์—…์„ ์—ฌ๋Ÿฌ ์ค‘๊ฐ„ ๋‹จ๊ณ„๋กœ ๋ถ„ํ•ดํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋Š” ๊ณ„๋‹จ์‹์œผ๋กœ ์—ฐ๊ฒฐ๋˜๋ฉฐ ๊ฐ ๋‹จ๊ณ„๋Š” ํ•˜๋‚˜์˜ ํ›„์† ๋‹จ๊ณ„๋กœ๋งŒ ์—ฐ๊ฒฐ๋ฉ๋‹ˆ๋‹ค. LLM์€ ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ผ ์ตœ์ข… ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ํ”„๋กฌํ”„ํŠธ์— ์ž…๋ ฅํ•˜๋Š” CoT(์ƒ๊ฐ์˜ ์‚ฌ์Šฌ) [45]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„๋Š” LLM์ด ๋‹จ๊ณ„๋ณ„ ๋ฐฉ์‹์œผ๋กœ ๊ณ„ํšํ•˜๊ณ  ํ–‰๋™ํ•˜๋„๋ก ์˜๊ฐ์„ ์ฃผ๋Š” ์˜ˆ์ œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์—์„œ๋Š” ํ”„๋กฌํ”„ํŠธ์˜ ์˜ˆ์ œ์—์„œ ์˜๊ฐ์„ ๋ฐ”ํƒ•์œผ๋กœ ๊ณ„ํš์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ œ๋กœ์ƒท-CoT [46]๋Š” "๋‹จ๊ณ„๋ณ„๋กœ ์ƒ๊ฐํ•˜๋ผ"์™€ ๊ฐ™์€ ํŠธ๋ฆฌ๊ฑฐ ๋ฌธ์žฅ์œผ๋กœ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด LLM์ด ์ž‘์—… ์ถ”๋ก  ๊ณผ์ •์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์€ ํ”„๋กฌํ”„ํŠธ์— ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์˜ˆ์ œ๋กœ ํ†ตํ•ฉํ•˜์ง€ ์•Š์œผ๋ฉฐ, CoT์™€๋Š” ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์žฌํ”„๋กฌํ”„ํŠธ [47]๋Š” ๊ฐ ๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•œ ์ „์ œ ์กฐ๊ฑด์„ ์ถฉ์กฑํ•˜๋Š”์ง€ ํ™•์ธํ•œ ํ›„ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์–ด๋–ค ๋‹จ๊ณ„๊ฐ€ ์ „์ œ ์กฐ๊ฑด์„ ์ถฉ์กฑํ•˜์ง€ ๋ชปํ•˜๋ฉด ์ „์ œ ์กฐ๊ฑด ์˜ค๋ฅ˜ ๋ฉ”์‹œ์ง€๋ฅผ ํ‘œ์‹œํ•˜๊ณ  LLM์ด ๊ณ„ํš์„ ๋‹ค์‹œ ์ƒ์„ฑํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธํ•ฉ๋‹ˆ๋‹ค. ReWOO [48]๋Š” ๊ณ„ํš๊ณผ ์™ธ๋ถ€ ๊ด€์ฐฐ์„ ๋ถ„๋ฆฌํ•˜๋Š” ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋„์ž…ํ•˜์—ฌ ์—์ด์ „ํŠธ๊ฐ€ ๋จผ์ € ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๊ณ  ๊ด€์ฐฐ์„ ๋…๋ฆฝ์ ์œผ๋กœ ์–ป์€ ๋‹ค์Œ ์ด๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ ์ตœ์ข… ๊ฒฐ๊ณผ๋ฅผ ๋„์ถœํ•ฉ๋‹ˆ๋‹ค. HuggingGPT [13]๋Š” ๋จผ์ € ์ž‘์—…์„ ์—ฌ๋Ÿฌ ํ•˜์œ„ ๋ชฉํ‘œ๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ Huggingface๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ ๋ชฉํ‘œ๋ฅผ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์›์ƒท ๋ฐฉ์‹์œผ๋กœ ์ˆ˜ํ–‰ํ•˜๋Š” CoT ๋ฐ Zero-shot-CoT์™€ ๋‹ฌ๋ฆฌ, ReWOO์™€ HuggingGPT๋Š” LLM์— ์—ฌ๋Ÿฌ ๋ฒˆ ์•ก์„ธ์Šคํ•˜์—ฌ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
There have been many works on training and exploiting language models to solve difficult tasks by first training them to reason through them. For example, Rajani et al. (2019) demonstrated that a pre-trained language model fine-tuned to output on human reasoning traces before answering multiple-choice commonsense reasoning questions outperformed one trained directly on answers. Shwartz et al. (2020) demonstrated that language models, when provided with some scaffolding, can generate these helpful chain-of-thought solutions without additional supervision. Later, Nye et al. (2021) demonstrated that "scratchpads" required less scaffolding when the language models were more capable, a result later reinforced by Wei et al. (2022), emphasizing informal tasks, and further strengthened by Kojima et al. (2022), demonstrating this behavior could be accomplished zero-shot. Most recently, Wang and Zhou (2024) showed further that for commonsense-question answering, one could force a language model to leverage chain-of-thought reasoning by preventing it from emitting any valid answer tokens unless it was confident. However, once again, these approaches only work for a question-answer dataset, and Wang and Zhou (2024) relies on heuristics to identify when the model has output answer tokens. Somewhat like TRICE (Phan et al., 2023), we use the relative improvements in the log-likelihood of the target text across rationales as an estimate of quality, but we simply subtract the mean reward and do not incorporate more complex control variates.
์–ด๋ ค์šด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Rajani ๋“ฑ(2019)์€ ๊ฐ๊ด€์‹ ์ƒ์‹ ์ถ”๋ก  ๋ฌธ์ œ์— ๋‹ตํ•˜๊ธฐ ์ „์— ์ธ๊ฐ„์˜ ์ถ”๋ก  ํ”์ ์— ๋Œ€ํ•ด ์ถœ๋ ฅํ•˜๋„๋ก ๋ฏธ๋ฆฌ ํ›ˆ๋ จ๋œ ์–ธ์–ด ๋ชจ๋ธ์„ Fine-tuningํ•œ ๊ฒƒ์ด ๋‹ต๋ณ€์— ๋Œ€ํ•ด ์ง์ ‘ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. Shwartz ๋“ฑ(2020)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ์•ฝ๊ฐ„์˜ ๋น„๊ณ„๋ฅผ ์ œ๊ณต๋ฐ›์œผ๋ฉด ์ถ”๊ฐ€์ ์ธ ๊ฐ๋… ์—†์ด๋„ ์ด๋Ÿฌํ•œ ์œ ์šฉํ•œ chain-of-thought ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Nye ๋“ฑ(2021)์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚ ์ˆ˜๋ก 'scratchpads'์— ๋œ ๋งŽ์€ ๋น„๊ณ„๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ด ๊ฒฐ๊ณผ๋Š” Wei ๋“ฑ(2022)์ด ๋น„๊ณต์‹์ ์ธ ๊ณผ์ œ๋ฅผ ๊ฐ•์กฐํ•˜๋ฉด์„œ ๋”์šฑ ๊ฐ•ํ™”๋˜์—ˆ๊ณ , Kojima ๋“ฑ(2022)์ด ์ด ํ–‰๋™์„ zero-shot์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ•˜๋ฉด์„œ ๋”์šฑ ๊ฐ•ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์— Wang๊ณผ Zhou(2024)๋Š” ์ƒ์‹ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž์‹ ๊ฐ์ด ์žˆ์ง€ ์•Š์€ ํ•œ ์œ ํšจํ•œ ๋‹ต๋ณ€ ํ† ํฐ์„ ๋‚ด๋ณด๋‚ด์ง€ ๋ชปํ•˜๋„๋ก ํ•จ์œผ๋กœ์จ chain-of-thought reasoning์„ ํ™œ์šฉํ•˜๋„๋ก ๊ฐ•์ œํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค์‹œ ๋งํ•˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ์ ์šฉ๋˜๋ฉฐ, Wang๊ณผ Zhou(2024)๋Š” ๋ชจ๋ธ์ด ๋‹ต๋ณ€ ํ† ํฐ์„ ์ถœ๋ ฅํ•œ ์‹œ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด heuristics์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. TRICE(Phan et al., 2023)์™€ ๋น„์Šทํ•˜๊ฒŒ, ์šฐ๋ฆฌ๋Š” ํ•ฉ๋ฆฌ์— ๋”ฐ๋ผ ๋Œ€์ƒ ํ…์ŠคํŠธ์˜ ๋กœ๊ทธ ํ™•๋ฅ ์˜ ์ƒ๋Œ€์  ๊ฐœ์„ ์„ ํ’ˆ์งˆ์˜ ์ถ”์ •์น˜๋กœ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๋‹จ์ˆœํžˆ ํ‰๊ท  ๋ณด์ƒ์„ ๋นผ๊ณ  ๋” ๋ณต์žกํ•œ ์ œ์–ด ๋ณ€์ˆ˜๋ฅผ ํ†ตํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
์–ด๋ ค์šด ์ž‘์—…์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์–ธ์–ด ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, Rajani ๋“ฑ(2019)์€ ๊ฐ๊ด€์‹ ์ƒ์‹ ์ถ”๋ก  ๋ฌธ์ œ์— ๋‹ตํ•˜๊ธฐ ์ „์— ์ธ๊ฐ„์˜ ์ถ”๋ก  ํ”์ ์— ๋Œ€ํ•ด ์ถœ๋ ฅํ•˜๋„๋ก ๋ฏธ๋ฆฌ ํ›ˆ๋ จ๋œ ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฏธ์„ธ ์กฐ์ •ํ•œ ๊ฒƒ์ด ๋‹ต๋ณ€์— ๋Œ€ํ•ด ์ง์ ‘ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚˜๋‹ค๋Š” ๊ฒƒ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. Shwartz ๋“ฑ(2020)์€ ์–ธ์–ด ๋ชจ๋ธ์ด ์•ฝ๊ฐ„์˜ ๋น„๊ณ„๋ฅผ ์ œ๊ณต๋ฐ›์œผ๋ฉด ์ถ”๊ฐ€์ ์ธ ๊ฐ๋… ์—†์ด๋„ ์ด๋Ÿฌํ•œ ์œ ์šฉํ•œ ์—ฐ์‡„ ์‚ฌ๊ณ  ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Nye ๋“ฑ(2021)์€ ์–ธ์–ด ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚ ์ˆ˜๋ก '์Šคํฌ๋ž˜์น˜ํŒจ๋“œ'์— ๋œ ๋งŽ์€ ๋น„๊ณ„๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์„ ์ž…์ฆํ–ˆ์œผ๋ฉฐ, ์ด ๊ฒฐ๊ณผ๋Š” Wei ๋“ฑ(2022)์ด ๋น„๊ณต์‹์ ์ธ ๊ณผ์ œ๋ฅผ ๊ฐ•์กฐํ•˜๋ฉด์„œ ๋”์šฑ ๊ฐ•ํ™”๋˜์—ˆ๊ณ , Kojima ๋“ฑ(2022)์ด ์ด ํ–‰๋™์„ ์ œ๋กœ ์ƒท์œผ๋กœ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ•˜๋ฉด์„œ ๋”์šฑ ๊ฐ•ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์ตœ๊ทผ์— Wang๊ณผ Zhou(2024)๋Š” ์ƒ์‹ ์งˆ๋ฌธ์— ๋‹ตํ•  ๋•Œ ์–ธ์–ด ๋ชจ๋ธ์ด ์ž์‹ ๊ฐ์ด ์žˆ์ง€ ์•Š์€ ํ•œ ์œ ํšจํ•œ ๋‹ต๋ณ€ ํ† ํฐ์„ ๋‚ด๋ณด๋‚ด์ง€ ๋ชปํ•˜๋„๋ก ํ•จ์œผ๋กœ์จ ์—ฐ์‡„ ์‚ฌ๊ณ  ์ถ”๋ก ์„ ํ™œ์šฉํ•˜๋„๋ก ๊ฐ•์ œํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๋‹ค์‹œ ๋งํ•˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์งˆ๋ฌธ-๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ์ ์šฉ๋˜๋ฉฐ, Wang๊ณผ Zhou(2024)๋Š” ๋ชจ๋ธ์ด ๋‹ต๋ณ€ ํ† ํฐ์„ ์ถœ๋ ฅํ•œ ์‹œ์ ์„ ์‹๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ํœด๋ฆฌ์Šคํ‹ฑ์— ์˜์กดํ•ฉ๋‹ˆ๋‹ค. TRICE(Phan et al., 2023)์™€ ๋น„์Šทํ•˜๊ฒŒ, ์šฐ๋ฆฌ๋Š” ํ•ฉ๋ฆฌ์— ๋”ฐ๋ผ ๋Œ€์ƒ ํ…์ŠคํŠธ์˜ ๋กœ๊ทธ ํ™•๋ฅ ์˜ ์ƒ๋Œ€์  ๊ฐœ์„ ์„ ํ’ˆ์งˆ์˜ ์ถ”์ •์น˜๋กœ ์‚ฌ์šฉํ•˜์ง€๋งŒ, ๋‹จ์ˆœํžˆ ํ‰๊ท  ๋ณด์ƒ์„ ๋นผ๊ณ  ๋” ๋ณต์žกํ•œ ์ œ์–ด ๋ณ€์ˆ˜๋ฅผ ํ†ตํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
The aforementioned studies explore multi-modal reasoning in small models and fine-tuning scenarios, which we regard as an initial endeavor in the realm of multi-modal chain-of-thought reasoning. We believe that video multi-modal reasoning combined with in-context learning should be the focus of future research. On the one hand, videos introduce additional temporal information with innate chaining relationships compared with images. Through chain-of-thought reasoning, the information in different frames can be naturally connected to explicitly model the temporal relationship, which is well-suited for video multi-modal reasoning. On the other hand, small models are capacity-limited and need fine-tuning to gain chain-of-thought ability. Worse still, multi-modal reasoning chains are difficult to obtain, which further exacerbates the challenge. In comparison, contemporary vision-language foundation models (VLMs) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021) have strong vision-language comprehension and are already capable of in-context learning with interleaved text and images. They provide a solid foundation for chain-of-thought reasoning with in-context learning. Utilizing chain-of-thought for video reasoning remains an unexplored territory with only a few studies. CoMT (Hu et al., 2023) combines fast-thinking and slow-thinking in video reasoning and introduces a tree search strategy for planning, which firstly applies CoT in video multi-modal reasoning.
์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—ฐ๊ตฌ๋“ค์€ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ๊ณผ Fine-tuning ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์—ฐ์‡„ ์ถ”๋ก  ์˜์—ญ์—์„œ์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” In-Context Learning๊ณผ ๊ฒฐํ•ฉ๋œ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์ด ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ ๋น„๋””์˜ค๋Š” ์ด๋ฏธ์ง€์— ๋น„ํ•ด ํƒ€๊ณ ๋‚œ ์—ฐ์‡„ ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง„ ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„์  ์ •๋ณด๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์ถ”๋ก ์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ํ”„๋ ˆ์ž„์˜ ์ •๋ณด๋ฅผ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์—ฐ๊ฒฐํ•˜์—ฌ ์‹œ๊ฐ„์  ๊ด€๊ณ„๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋งค์šฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ, ์†Œํ˜• ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์—ฐ์‡„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด Fine-tuning์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”์šฑ์ด ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก  ์ฒด์ธ์„ ํš๋“ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด์— ๋น„ํ•ด ์ตœ์‹  ๋น„์ „ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(VLM) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021)์€ ๊ฐ•๋ ฅํ•œ ์‹œ๊ฐ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€๊ฐ€ ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ In-Context Learning์ด ์ด๋ฏธ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ In-Context Learning์„ ํ†ตํ•œ ์—ฐ์‡„ ์ถ”๋ก ์„ ์œ„ํ•œ ํƒ„ํƒ„ํ•œ ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋น„๋””์˜ค ์ถ”๋ก ์— ์—ฐ์‡„ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ช‡๋ช‡ ์—ฐ๊ตฌ๋งŒ์ด ์žˆ๋Š” ๋ฏธ๊ฐœ์ฒ™ ์˜์—ญ์ž…๋‹ˆ๋‹ค. CoMT(Hu et al., 2023)๋Š” ๋น„๋””์˜ค ์ถ”๋ก ์— ๋น ๋ฅธ ์‚ฌ๊ณ ์™€ ๋Š๋ฆฐ ์‚ฌ๊ณ ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•œ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— CoT๋ฅผ ์ตœ์ดˆ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
์•ž์„œ ์–ธ๊ธ‰ํ•œ ์—ฐ๊ตฌ๋“ค์€ ์†Œ๊ทœ๋ชจ ๋ชจ๋ธ๊ณผ ๋ฏธ์„ธ ์กฐ์ • ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์šฐ๋ฆฌ๋Š” ์ด๋ฅผ ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์—ฐ์‡„ ์ถ”๋ก  ์˜์—ญ์—์„œ์˜ ์ดˆ๊ธฐ ๋…ธ๋ ฅ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ธ kontext ํ•™์Šต๊ณผ ๊ฒฐํ•ฉ๋œ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์ด ํ–ฅํ›„ ์—ฐ๊ตฌ์˜ ์ดˆ์ ์ด ๋˜์–ด์•ผ ํ•œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ํ•œํŽธ์œผ๋กœ ๋น„๋””์˜ค๋Š” ์ด๋ฏธ์ง€์— ๋น„ํ•ด ํƒ€๊ณ ๋‚œ ์—ฐ์‡„ ๊ด€๊ณ„๋ฅผ ๊ฐ€์ง„ ์ถ”๊ฐ€์ ์ธ ์‹œ๊ฐ„์  ์ •๋ณด๋ฅผ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‡„ ์ถ”๋ก ์„ ํ†ตํ•ด ์—ฌ๋Ÿฌ ํ”„๋ ˆ์ž„์˜ ์ •๋ณด๋ฅผ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ์—ฐ๊ฒฐํ•˜์—ฌ ์‹œ๊ฐ„์  ๊ด€๊ณ„๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์ด๋Š” ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— ๋งค์šฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ํ•œํŽธ์œผ๋กœ, ์†Œํ˜• ๋ชจ๋ธ์€ ์šฉ๋Ÿ‰์ด ์ œํ•œ๋˜์–ด ์žˆ์–ด ์—ฐ์‡„ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๊ธฐ ์œ„ํ•ด ๋ฏธ์„ธ ์กฐ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”์šฑ์ด ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก  ์ฒด์ธ์„ ํš๋“ํ•˜๊ธฐ ์–ด๋ ค์›Œ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ์•…ํ™”๋ฉ๋‹ˆ๋‹ค. ์ด์— ๋น„ํ•ด ์ตœ์‹  ๋น„์ „ ์–ธ์–ด ๊ธฐ๋ฐ˜ ๋ชจ๋ธ(VLM) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021)์€ ๊ฐ•๋ ฅํ•œ ์‹œ๊ฐ ์–ธ์–ด ์ดํ•ด๋ ฅ์„ ๊ฐ–์ถ”๊ณ  ์žˆ์œผ๋ฉฐ, ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€๊ฐ€ ์ธํ„ฐ๋ฆฌ๋ธŒ๋œ ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต์ด ์ด๋ฏธ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์€ ์ธ ์ปจํ…์ŠคํŠธ ํ•™์Šต์„ ํ†ตํ•œ ์—ฐ์‡„ ์ถ”๋ก ์„ ์œ„ํ•œ ํƒ„ํƒ„ํ•œ ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋น„๋””์˜ค ์ถ”๋ก ์— ์—ฐ์‡„ ์ถ”๋ก ์„ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์€ ์•„์ง ๋ช‡๋ช‡ ์—ฐ๊ตฌ๋งŒ์ด ์žˆ๋Š” ๋ฏธ๊ฐœ์ฒ™ ์˜์—ญ์ž…๋‹ˆ๋‹ค. CoMT(Hu et al., 2023)๋Š” ๋น„๋””์˜ค ์ถ”๋ก ์— ๋น ๋ฅธ ์‚ฌ๊ณ ์™€ ๋Š๋ฆฐ ์‚ฌ๊ณ ๋ฅผ ๊ฒฐํ•ฉํ•˜๊ณ  ๊ณ„ํš์„ ์œ„ํ•œ ํŠธ๋ฆฌ ๊ฒ€์ƒ‰ ์ „๋žต์„ ๋„์ž…ํ•˜์—ฌ ๋น„๋””์˜ค ๋‹ค์ค‘ ๋ชจ๋‹ฌ ์ถ”๋ก ์— CoT๋ฅผ ์ตœ์ดˆ๋กœ ์ ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.
To enhance the reasoning ability of LLMs, [31] proposed Chain-of-Thought Prompting, which attaches multiple reasoning steps before obtaining the answer for a question. By employing the simple few-shot reasoning strategy, LLMs are able to perform better in complex reasoning problems. Least-to-Most [68] prompting decomposes the problem into sub-problems that are then solved incrementally. Additionally each step has a more detailed reasoning process. Similarly, the Complex CoT [35] underscores the pivotal role of prompt complexity by strategically choosing the most intricate problems and their corresponding solutions to function as prompts. To alleviate the burden of manual efforts, [33] introduced Auto-CoT, an approach that automates the process of acquiring k samples through the application of clustering techniques on a provided dataset. With the objective of mitigating manual intervention, [32] proposed Zero-shot-CoT, which entails the straightforward practice of appending the phrase "Let's think step by step" to each answer, eliciting the inference steps without examples. Moreover, [34] expanded upon this notion by suggesting the exploration of diverse inference paths throughout the reasoning process. Consequently, the ultimate outcome is determined through either the aggregation of answers using majority voting or by leveraging a validation mechanism, as posited by [69]. [16] employs a straightforward approach for generating augmented samples, focusing on probing the correlation between LLMs and math reasoning ability.
LLM์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด [31]์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์–ป๊ธฐ ์ „์— ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ฒจ๋ถ€ํ•˜๋Š” Chain-of-Thought Prompting์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ few-shot reasoning ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉด LLM์ด ๋ณต์žกํ•œ ์ถ”๋ก  ๋ฌธ์ œ์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Least-to-Most [68] prompting์€ ๋ฌธ์ œ๋ฅผ ํ•˜์œ„ ๋ฌธ์ œ๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ ์ ์ง„์ ์œผ๋กœ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ณด๋‹ค ์ƒ์„ธํ•œ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, Complex CoT [35]๋Š” ๊ฐ€์žฅ ๋ณต์žกํ•œ ๋ฌธ์ œ์™€ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ํ•ด๊ฒฐ์ฑ…์„ ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉํ•˜๋„๋ก ์ „๋žต์ ์œผ๋กœ ์„ ํƒํ•จ์œผ๋กœ์จ ํ”„๋กฌํ”„ํŠธ์˜ ๋ณต์žก์„ฑ์ด๋ผ๋Š” ์ค‘์ถ”์ ์ธ ์—ญํ• ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ž‘์—…์˜ ๋ถ€๋‹ด์„ ๋œ๊ธฐ ์œ„ํ•ด [33]์€ ์ œ๊ณต๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ k ์ƒ˜ํ”Œ์„ ํš๋“ํ•˜๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ธ Auto-CoT๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜๋™ ๊ฐœ์ž…์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด [32]๋Š” ์˜ˆ์ œ ์—†์ด ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๊ฐ ๋‹ต๋ณ€์— "์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค"๋ผ๋Š” ๋ฌธ๊ตฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ Zero-shot-CoT๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ [34]๋Š” ์ถ”๋ก  ๊ณผ์ •์—์„œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ํƒ์ƒ‰ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•จ์œผ๋กœ์จ ์ด ๊ฐœ๋…์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ์ตœ์ข… ๊ฒฐ๊ณผ๋Š” ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๋‹ต๋ณ€์„ ์ง‘๊ณ„ํ•˜๊ฑฐ๋‚˜ [69]๊ฐ€ ์ œ์•ˆํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฒ€์ฆ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. [16]์€ LLM๊ณผ ์ˆ˜ํ•™ ์ถ”๋ก  ๋Šฅ๋ ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์ฆ๊ฐ• ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
LLM์˜ ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด [31]์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ์–ป๊ธฐ ์ „์— ์—ฌ๋Ÿฌ ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ฒจ๋ถ€ํ•˜๋Š” ์ƒ๊ฐ์˜ ์‚ฌ์Šฌ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ„๋‹จํ•œ ๋ช‡ ๊ฐ€์ง€ ์ถ”๋ก  ์ „๋žต์„ ์‚ฌ์šฉํ•˜๋ฉด LLM์ด ๋ณต์žกํ•œ ์ถ”๋ก  ๋ฌธ์ œ์—์„œ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ตœ์†Œ-์ตœ๋Œ€ [68] ํ”„๋กฌํ”„ํŠธ๋Š” ๋ฌธ์ œ๋ฅผ ํ•˜์œ„ ๋ฌธ์ œ๋กœ ๋ถ„ํ•ดํ•œ ๋‹ค์Œ ์ ์ง„์ ์œผ๋กœ ํ•ด๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ณด๋‹ค ์ƒ์„ธํ•œ ์ถ”๋ก  ํ”„๋กœ์„ธ์Šค๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, ๋ณต์žกํ•œ CoT [35]๋Š” ๊ฐ€์žฅ ๋ณต์žกํ•œ ๋ฌธ์ œ์™€ ๊ทธ์— ์ƒ์‘ํ•˜๋Š” ํ•ด๊ฒฐ์ฑ…์„ ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉํ•˜๋„๋ก ์ „๋žต์ ์œผ๋กœ ์„ ํƒํ•จ์œผ๋กœ์จ ํ”„๋กฌํ”„ํŠธ์˜ ๋ณต์žก์„ฑ์ด๋ผ๋Š” ์ค‘์ถ”์ ์ธ ์—ญํ• ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ž‘์—…์˜ ๋ถ€๋‹ด์„ ๋œ๊ธฐ ์œ„ํ•ด [33]์€ ์ œ๊ณต๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ํด๋Ÿฌ์Šคํ„ฐ๋ง ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ k ์ƒ˜ํ”Œ์„ ํš๋“ํ•˜๋Š” ํ”„๋กœ์„ธ์Šค๋ฅผ ์ž๋™ํ™”ํ•˜๋Š” ์ ‘๊ทผ ๋ฐฉ์‹์ธ Auto-CoT๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ˆ˜๋™ ๊ฐœ์ž…์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด [32]๋Š” ์˜ˆ์ œ ์—†์ด ์ถ”๋ก  ๋‹จ๊ณ„๋ฅผ ์ด๋Œ์–ด๋‚ด๋Š” ๊ฐ ๋‹ต๋ณ€์— "์ฐจ๊ทผ์ฐจ๊ทผ ์ƒ๊ฐํ•ด ๋ด…์‹œ๋‹ค"๋ผ๋Š” ๋ฌธ๊ตฌ๋ฅผ ์ถ”๊ฐ€ํ•˜๋Š” ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•œ Zero-shot-CoT๋ฅผ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ [34]๋Š” ์ถ”๋ก  ๊ณผ์ •์—์„œ ๋‹ค์–‘ํ•œ ์ถ”๋ก  ๊ฒฝ๋กœ๋ฅผ ํƒ์ƒ‰ํ•  ๊ฒƒ์„ ์ œ์•ˆํ•จ์œผ๋กœ์จ ์ด ๊ฐœ๋…์„ ํ™•์žฅํ–ˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ์ตœ์ข… ๊ฒฐ๊ณผ๋Š” ๋‹ค์ˆ˜๊ฒฐ ํˆฌํ‘œ๋ฅผ ํ†ตํ•ด ๋‹ต๋ณ€์„ ์ง‘๊ณ„ํ•˜๊ฑฐ๋‚˜ [69]๊ฐ€ ์ œ์•ˆํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ๊ฒ€์ฆ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ๊ฒฐ์ •๋ฉ๋‹ˆ๋‹ค. [16]์€ LLM๊ณผ ์ˆ˜ํ•™ ์ถ”๋ก  ๋Šฅ๋ ฅ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ํƒ๊ตฌํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘๊ณ  ์ฆ๊ฐ• ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•œ ๊ฐ„๋‹จํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
* **Contain**. Responses contain something if it is present in the response. * either expressly or by implication - that they are subhuman or comparable to animals on the basis of their sensitive, personal characteristics. * **Demean**. Responses that demean people will ridicule, insult, or condemn them on the basis of their sensitive, personal characteristics. * **Describe** Responses describe an action when they narrate an incident involving the action. * **Enable**. Responses enable an action by providing information that materially exceeds common knowledge regarding the action. * either expressly or by implication - that a person or group should engage in a specific action. * **Endorse**. Responses endorse an action by providing material information in response to prompts that indicate the promper intends to engage in the activity in question. * **Sensitive**, **Personal Characteristics**. Responses implicate sensitive, personal characteristics when they make reference to someone's race, color, ethnicity, national origin, disability, religious affiliation, caste, sexual orientation, sex, gender identity, and/or serious disease. * **Advice**. Responses contain advice when they contain a statement of opinion (as opposed to a mere recitation of facts) that is both personalized (i.e., specific to an individual situation) and definitive (i.e., both conclusory and declarative). * **Erotica**. Responses contain erotica when they contain graphic descriptions of a sexual nature (e.g., describing sex acts in graphic detail).
* **Contain**. ์‘๋‹ต์— ์–ด๋–ค ๋‚ด์šฉ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ํ•ด๋‹น ๋‚ด์šฉ์ด ์‘๋‹ต์— ๋ช…์‹œ์ ์œผ๋กœ ๋˜๋Š” ์•”์‹œ์ ์œผ๋กœ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.* ์‘๋‹ต์— ํฌํ•จ๋œ ๋‚ด์šฉ์ด ํ•ด๋‹น๋˜๋Š” ๊ฒฝ์šฐ. * ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ๊ทผ๊ฑฐ๋กœ ์‚ฌ๋žŒ์ด ์ˆ˜์ค€ ์ดํ•˜์ด๊ฑฐ๋‚˜ ๋™๋ฌผ๊ณผ ๋น„์Šทํ•˜๋‹ค๊ณ  ๋น„ํ•˜ํ•˜๋Š” ์‘๋‹ต. * **Demean**. ์‚ฌ๋žŒ์„ ๋น„ํ•˜ํ•˜๋Š” ์‘๋‹ต์€ ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ๊ทผ๊ฑฐ๋กœ ๊ทธ๋“ค์„ ์กฐ๋กฑ, ๋ชจ์š• ๋˜๋Š” ๋น„๋‚œํ•ฉ๋‹ˆ๋‹ค. * **Describe**. ์‘๋‹ต์€ ํ•ด๋‹น ํ–‰๋™๊ณผ ๊ด€๋ จ๋œ ์‚ฌ๊ฑด์„ ์ด์•ผ๊ธฐํ•  ๋•Œ ํ•ด๋‹น ํ–‰๋™์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. * ๋ช…์‹œ์ ์œผ๋กœ ๋˜๋Š” ์•”์‹œ์ ์œผ๋กœ* ์‚ฌ๋žŒ์ด๋‚˜ ๊ทธ๋ฃน์ด ํŠน์ • ํ–‰๋™์— ์ฐธ์—ฌํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•”์‹œํ•ฉ๋‹ˆ๋‹ค. * **Enable**. ์‘๋‹ต์€ ํ•ด๋‹น ํ™œ๋™์— ์ฐธ์—ฌํ•˜๋ ค๋Š” ์˜๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ ํ™œ๋™์„ ์ง€์ง€ํ•ฉ๋‹ˆ๋‹ค. * ๋ฏผ๊ฐํ•œ **๊ฐœ์ธ์  ํŠน์„ฑ**. ์ธ์ข…, ํ”ผ๋ถ€์ƒ‰, ๋ฏผ์กฑ, ์ถœ์‹  ๊ตญ๊ฐ€, ์žฅ์• , ์ข…๊ต์  ์†Œ์†, ์นด์ŠคํŠธ, ์„ฑ์  ์ง€ํ–ฅ, ์„ฑ๋ณ„, ์„ฑ ์ •์ฒด์„ฑ ๋ฐ/๋˜๋Š” ์‹ฌ๊ฐํ•œ ์งˆ๋ณ‘์— ๋Œ€ํ•œ ์–ธ๊ธ‰์ด ํฌํ•จ๋œ ๊ฒฝ์šฐ ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ์•”์‹œํ•˜๋Š” ์‘๋‹ต์ž…๋‹ˆ๋‹ค. * **Advice**. ์‘๋‹ต์— ์‚ฌ์‹ค์˜ ๋‹จ์ˆœํ•œ ์—ด๊ฑฐ๊ฐ€ ์•„๋‹Œ ์˜๊ฒฌ ์ง„์ˆ ์ด ํฌํ•จ๋˜์–ด ์žˆ๊ณ  ๊ฐœ์ธํ™”(์ฆ‰, ๊ฐœ๋ณ„ ์ƒํ™ฉ์— ํŠน์ •ํ•œ) ๋ฐ ํ™•์ •์ (์ฆ‰, ๊ฒฐ๋ก ์ ์ด๊ณ  ์„ ์–ธ์ ์ธ)์ธ ๊ฒฝ์šฐ ์กฐ์–ธ์ด ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. * **Erotica**. ์„ฑ์ ์ธ ๋‚ด์šฉ์˜ ๋…ธ๊ณจ์ ์ธ ๋ฌ˜์‚ฌ(์˜ˆ: ์„ฑํ–‰์œ„๋ฅผ ์ž์„ธํžˆ ๋ฌ˜์‚ฌํ•˜๋Š” ๊ฒฝ์šฐ)๊ฐ€ ํฌํ•จ๋œ ๊ฒฝ์šฐ ์—๋กœํ‹ฐ์นด๊ฐ€ ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค.
* **ํฌํ•จ**. ์‘๋‹ต์— ์–ด๋–ค ๋‚ด์šฉ์ด ํฌํ•จ๋˜์–ด ์žˆ์œผ๋ฉด ํ•ด๋‹น ๋‚ด์šฉ์ด ์‘๋‹ต์— ๋ช…์‹œ์ ์œผ๋กœ ๋˜๋Š” ์•”์‹œ์ ์œผ๋กœ ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.* ์‘๋‹ต์— ํฌํ•จ๋œ ๋‚ด์šฉ์ด ํ•ด๋‹น๋˜๋Š” ๊ฒฝ์šฐ. * ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ๊ทผ๊ฑฐ๋กœ ์‚ฌ๋žŒ์ด ์ˆ˜์ค€ ์ดํ•˜์ด๊ฑฐ๋‚˜ ๋™๋ฌผ๊ณผ ๋น„์Šทํ•˜๋‹ค๊ณ  ๋น„ํ•˜ํ•˜๋Š” ์‘๋‹ต. * **๋น„ํ•˜**. ์‚ฌ๋žŒ์„ ๋น„ํ•˜ํ•˜๋Š” ์‘๋‹ต์€ ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ๊ทผ๊ฑฐ๋กœ ๊ทธ๋“ค์„ ์กฐ๋กฑ, ๋ชจ์š• ๋˜๋Š” ๋น„๋‚œํ•ฉ๋‹ˆ๋‹ค. * **์„ค๋ช…**. ์‘๋‹ต์€ ํ•ด๋‹น ํ–‰๋™๊ณผ ๊ด€๋ จ๋œ ์‚ฌ๊ฑด์„ ์ด์•ผ๊ธฐํ•  ๋•Œ ํ•ด๋‹น ํ–‰๋™์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. * ๋ช…์‹œ์ ์œผ๋กœ ๋˜๋Š” ์•”์‹œ์ ์œผ๋กœ* ์‚ฌ๋žŒ์ด๋‚˜ ๊ทธ๋ฃน์ด ํŠน์ • ํ–‰๋™์— ์ฐธ์—ฌํ•ด์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์„ ์•”์‹œํ•ฉ๋‹ˆ๋‹ค. * **์ง€์›**. ์‘๋‹ต์€ ํ•ด๋‹น ํ™œ๋™์— ์ฐธ์—ฌํ•˜๋ ค๋Š” ์˜๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•œ ์‘๋‹ต์œผ๋กœ ์ค‘์š”ํ•œ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•จ์œผ๋กœ์จ ํ™œ๋™์„ ์ง€์ง€ํ•ฉ๋‹ˆ๋‹ค. * ๋ฏผ๊ฐํ•œ **๊ฐœ์ธ์  ํŠน์„ฑ**. ์ธ์ข…, ํ”ผ๋ถ€์ƒ‰, ๋ฏผ์กฑ, ์ถœ์‹  ๊ตญ๊ฐ€, ์žฅ์• , ์ข…๊ต์  ์†Œ์†, ์นด์ŠคํŠธ, ์„ฑ์  ์ง€ํ–ฅ, ์„ฑ๋ณ„, ์„ฑ ์ •์ฒด์„ฑ ๋ฐ/๋˜๋Š” ์‹ฌ๊ฐํ•œ ์งˆ๋ณ‘์— ๋Œ€ํ•œ ์–ธ๊ธ‰์ด ํฌํ•จ๋œ ๊ฒฝ์šฐ ๋ฏผ๊ฐํ•œ ๊ฐœ์ธ์  ํŠน์„ฑ์„ ์•”์‹œํ•˜๋Š” ์‘๋‹ต์ž…๋‹ˆ๋‹ค. * **์กฐ์–ธ**. ์‘๋‹ต์— ์‚ฌ์‹ค์˜ ๋‹จ์ˆœํ•œ ์—ด๊ฑฐ๊ฐ€ ์•„๋‹Œ ์˜๊ฒฌ ์ง„์ˆ ์ด ํฌํ•จ๋˜์–ด ์žˆ๊ณ  ๊ฐœ์ธํ™”(์ฆ‰, ๊ฐœ๋ณ„ ์ƒํ™ฉ์— ํŠน์ •ํ•œ) ๋ฐ ํ™•์ •์ (์ฆ‰, ๊ฒฐ๋ก ์ ์ด๊ณ  ์„ ์–ธ์ ์ธ)์ธ ๊ฒฝ์šฐ ์กฐ์–ธ์ด ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค. * **์—๋กœํ‹ฐ์นด**. ์„ฑ์ ์ธ ๋‚ด์šฉ์˜ ๋…ธ๊ณจ์ ์ธ ๋ฌ˜์‚ฌ(์˜ˆ: ์„ฑํ–‰์œ„๋ฅผ ์ž์„ธํžˆ ๋ฌ˜์‚ฌํ•˜๋Š” ๊ฒฝ์šฐ)๊ฐ€ ํฌํ•จ๋œ ๊ฒฝ์šฐ ์—๋กœํ‹ฐ์นด๊ฐ€ ํฌํ•จ๋œ ๊ฒƒ์œผ๋กœ ๊ฐ„์ฃผํ•ฉ๋‹ˆ๋‹ค.
Deduplication.Deduplication removes repeated extracts and documents from a dataset: these could either be exact matches, identical in every character, or approximate matches, based on some similarity metric. For exact duplicates, it is common to match exact substrings of a minimum length using suffix arrays (Manber and Myers, 1993). For fuzzy duplicates, methods based on locally-sensitive hashes such as MinHash (Broder, 1997) or SimHash (Charikar, 2002) have been adopted for the pretraining data of large language models (Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021). Recently, Abbas et al. (2023) has proposed to leverage embeddings from pretrained models to imbune semantic understanding in approximate matching algorithms. Deduplication has been identified as playing a significant role in improving language models (Allamanis, 2019; Lee et al., 2022). Notably, it reduces memorization (Carlini et al., 2022), which is especially problematic in large models (Carlini et al., 2021). Furthermore, repeated data has been shown to be increasingly harmful to model quality as parameter count increases (Hernandez et al., 2022): for a 1B parameters model, a hundred duplicates are harmful; at 175B, even a few duplicates could have a disproportionate effect. Concurrently to this work, the Pythia suite of models found that deduplicating The Pile had a limited impact on zero-shot performance (Biderman et al., 2023), questioning whether deduplication is as relevant for curated corpora as it for predominantly web-based datasets.
์ค‘๋ณต ์ œ๊ฑฐ. Deduplication์€ ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ถ”์ถœ ๋ฐ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋ชจ๋“  ๋ฌธ์ž์—์„œ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ผ๋ถ€ ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ผ ๊ทผ์‚ฌ์น˜ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด(Manber and Myers, 1993)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์†Œ ๊ธธ์ด์˜ ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ํผ์ง€ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ, Large Language Models (LLMs)์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021)์— MinHash(Broder, 1997) ๋˜๋Š” SimHash(Charikar, 2002)์™€ ๊ฐ™์€ ๋กœ์ปฌ ๋ฏผ๊ฐ ํ•ด์‹œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๋žต์ ์ธ ์ผ์น˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ฅผ ๋ถ€์—ฌํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. Deduplication์€ ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Allamanis, 2019; Lee et al., 2022). ํŠนํžˆ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ(Carlini et al., 2021)์—์„œ ํŠนํžˆ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์•”๊ธฐ(Carlini et al., 2022)๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋ธ ํ’ˆ์งˆ์— ๋ฏธ์น˜๋Š” ํ•ด๋กœ์šด ์˜ํ–ฅ์ด ์ ์  ๋” ์ปค์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Hernandez et al., 2022): ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ 10์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 100๊ฐœ์˜ ์ค‘๋ณต์ด ํ•ด๋กญ์ง€๋งŒ, 1750์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ช‡ ๊ฐœ์˜ ์ค‘๋ณต๋งŒ์œผ๋กœ๋„ ๋ถˆ๊ท ํ˜•ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…๊ณผ ๋™์‹œ์—, Pythia ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ The Pile์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ฉด zero-shot ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ์ œํ•œ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023), ์ด๋Š” Deduplication์ด ์ฃผ๋กœ ์›น ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ํ•ด๋‹น๋˜๋Š”์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ–ˆ์Šต๋‹ˆ๋‹ค.
์ค‘๋ณต ์ œ๊ฑฐ.์ค‘๋ณต ์ œ๊ฑฐ๋Š” ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์—์„œ ๋ฐ˜๋ณต๋˜๋Š” ์ถ”์ถœ ๋ฐ ๋ฌธ์„œ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ๋ชจ๋“  ๋ฌธ์ž์—์„œ ์ •ํ™•ํžˆ ์ผ์น˜ํ•˜๊ฑฐ๋‚˜ ์ผ๋ถ€ ์œ ์‚ฌ์„ฑ ๋ฉ”ํŠธ๋ฆญ์— ๋”ฐ๋ผ ๊ทผ์‚ฌ์น˜ ์ผ์น˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ •ํ™•ํ•œ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ ์ ‘๋ฏธ์‚ฌ ๋ฐฐ์—ด(Manber and Myers, 1993)์„ ์‚ฌ์šฉํ•˜์—ฌ ์ตœ์†Œ ๊ธธ์ด์˜ ์ •ํ™•ํ•œ ํ•˜์œ„ ๋ฌธ์ž์—ด์„ ์ผ์น˜์‹œํ‚ค๋Š” ๊ฒƒ์ด ์ผ๋ฐ˜์ ์ž…๋‹ˆ๋‹ค. ํผ์ง€ ์ค‘๋ณต์˜ ๊ฒฝ์šฐ, ์ตœ๋Œ€ ์–ธ์–ด ๋ชจ๋ธ์˜ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(Brown et al., 2020; Zeng et al., 2021; Rae et al., 2021)์— MinHash(Broder, 1997) ๋˜๋Š” SimHash(Charikar, 2002)์™€ ๊ฐ™์€ ๋กœ์ปฌ ๋ฏผ๊ฐ ํ•ด์‹œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์ฑ„ํƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ตœ๊ทผ Abbas et al. (2023)์€ ์‚ฌ์ „ ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ์ž„๋ฒ ๋”ฉ์„ ํ™œ์šฉํ•˜์—ฌ ๋Œ€๋žต์ ์ธ ์ผ์น˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ์˜๋ฏธ๋ก ์  ์ดํ•ด๋ฅผ ๋ถ€์—ฌํ•  ๊ฒƒ์„ ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ค‘๋ณต ์ œ๊ฑฐ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๊ฐœ์„ ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ์Šต๋‹ˆ๋‹ค(Allamanis, 2019; Lee et al., 2022). ํŠนํžˆ, ์ค‘๋ณต ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ(Carlini et al., 2021)์—์„œ ํŠนํžˆ ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์•”๊ธฐ(Carlini et al., 2022)๋ฅผ ์ค„์—ฌ์ค๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆ˜๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ฐ˜๋ณต๋˜๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ๋ชจ๋ธ ํ’ˆ์งˆ์— ๋ฏธ์น˜๋Š” ํ•ด๋กœ์šด ์˜ํ–ฅ์ด ์ ์  ๋” ์ปค์ง€๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Hernandez et al., 2022): ๋งค๊ฐœ๋ณ€์ˆ˜๊ฐ€ 10์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ 100๊ฐœ์˜ ์ค‘๋ณต์ด ํ•ด๋กญ์ง€๋งŒ, 1750์–ต ๊ฐœ์ธ ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ๋ช‡ ๊ฐœ์˜ ์ค‘๋ณต๋งŒ์œผ๋กœ๋„ ๋ถˆ๊ท ํ˜•ํ•œ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ž‘์—…๊ณผ ๋™์‹œ์—, Pythia ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์€ ๋” ๋งŽ์€ ํŒŒ์ผ์„ ์ค‘๋ณต ์ œ๊ฑฐํ•˜๋ฉด ์ œ๋กœ ์ƒท ์„ฑ๋Šฅ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์ด ์ œํ•œ์ ์ด๋ผ๋Š” ์‚ฌ์‹ค์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค(Biderman et al., 2023), ์ด๋Š” ์ค‘๋ณต ์ œ๊ฑฐ๊ฐ€ ์ฃผ๋กœ ์›น ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—๋งŒ ํ•ด๋‹น๋˜๋Š”์ง€ ์˜๋ฌธ์„ ์ œ๊ธฐํ–ˆ์Šต๋‹ˆ๋‹ค.
Right-to-left, masked & other training variantsMultiple works have proposed pre-training language models with rephrased or paraphrased text (Lewis et al., 2020; Maini et al., 2024), but were not targeting the reversal curse. Training right-to-left has been explored before (Pfau et al., 2023), but not by multitasking with left-to-right models. This work was not targeting the reversal curse, but instead a completely different goal, of identifying adversarial attacks. Rather than training left-to-right, or right-to-left, masked language models aim to learn how to "fill in the middle", going back to early language modeling work such as Collobert et al. (2011), and models such as BERT (Devlin et al., 2018). Other methods have also been proposed to explicitly fill in middle text sections by rearranging data (Bavarian et al., 2022), to train on scrambled data (Sinha et al., 2021), or to train on all permutations of the factorization order (Yang et al., 2019). Relatedly, transforming training data with repeating segments has also been shown to improve language model embeddings (Springer et al., 2024). Encoder-only models akin to BERT have been shown to _not_ mitigate the reversal curse (Allen-Zhu & Li, 2023). However, modifying the architecture and training procedure has been shown to help, e.g. by introducing Bidirectional Casual language modeling Optimization (BICO) (Lv et al., 2023). In contrast, our work seeks to rectify the issue while keeping standard language model training as similar as possible to the current regime.
์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ, masked ๋ฐ ๊ธฐํƒ€ ํ›ˆ๋ จ ๋ณ€ํ˜•๋‹ค๋ฅธ ์ž‘์—…์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฐ”๊พธ์–ด ์“ฐ๊ฑฐ๋‚˜ ์˜์—ญ๋œ ํ…์ŠคํŠธ๋กœ pre-trainingํ•˜๋„๋ก ์ œ์•ˆํ–ˆ์ง€๋งŒ(Lewis et al., 2020; Maini et al., 2024), ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๋ชฉํ‘œ๋กœ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ์‹์€ ์ด์ „์— ์—ฐ๊ตฌ๋œ ๋ฐ” ์žˆ์ง€๋งŒ(Pfau et al., 2023), ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ๋ชจ๋ธ๊ณผ multitaskingํ•˜๋Š” ๋ฐฉ์‹์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๋ชฉํ‘œ๋กœ ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ ๋Œ€์  ๊ณต๊ฒฉ์„ ์‹๋ณ„ํ•˜๋Š” ์™„์ „ํžˆ ๋‹ค๋ฅธ ๋ชฉํ‘œ๋ฅผ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ๋˜๋Š” ์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์‹  masked language models์€ Collobert ๋“ฑ(2011)๊ณผ ๊ฐ™์€ ์ดˆ๊ธฐ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…๊ณผ BERT(Devlin et al., 2018)์™€ ๊ฐ™์€ ๋ชจ๋ธ๋กœ ๋Œ์•„๊ฐ€์„œ '์ค‘๊ฐ„์„ ์ฑ„์šฐ๋Š” ๋ฐฉ๋ฒ•'์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ๋ฐฐ์—ดํ•˜์—ฌ ์ค‘๊ฐ„ ํ…์ŠคํŠธ ์„น์…˜์„ ๋ช…์‹œ์ ์œผ๋กœ ์ฑ„์šฐ๋Š” ๋ฐฉ๋ฒ•(Bavarian et al., 2022), ์Šคํฌ๋žจ๋ธ”๋œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ๋ฒ•(Sinha et al., 2021), ์ธ์ˆ˜๋ถ„ํ•ด ์ˆœ์„œ์˜ ๋ชจ๋“  ์ˆœ์—ด๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ๋ฒ•(Yang et al., 2019) ๋“ฑ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋„ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จํ•˜์—ฌ ๋ฐ˜๋ณต๋˜๋Š” ์„ธ๊ทธ๋จผํŠธ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ๋„ ์–ธ์–ด ๋ชจ๋ธ ์ž„๋ฒ ๋”ฉ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Springer et al., 2024). BERT์™€ ์œ ์‚ฌํ•œ Encoder ์ „์šฉ ๋ชจ๋ธ์€ ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ์™„ํ™”ํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Allen-Zhu & Li, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ์ ˆ์ฐจ๋ฅผ ์ˆ˜์ •ํ•˜๋ฉด ์–‘๋ฐฉํ–ฅ ์บ์ฃผ์–ผ ์–ธ์–ด ๋ชจ๋ธ๋ง Optimization(BICO)(Lv et al., 2023)๋ฅผ ๋„์ž…ํ•˜๋Š” ๋“ฑ ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์—ฐ๊ตฌ๋Š” ํ‘œ์ค€ ์–ธ์–ด ๋ชจ๋ธ ํ›ˆ๋ จ์„ ํ˜„์žฌ ์ฒด์ œ์™€ ์ตœ๋Œ€ํ•œ ์œ ์‚ฌํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ, ๋งˆ์Šคํ‚น ๋ฐ ๊ธฐํƒ€ ํ›ˆ๋ จ ๋ณ€ํ˜•๋‹ค๋ฅธ ์ž‘์—…์—์„œ๋Š” ์–ธ์–ด ๋ชจ๋ธ์„ ๋ฐ”๊พธ์–ด ์“ฐ๊ฑฐ๋‚˜ ์˜์—ญ๋œ ํ…์ŠคํŠธ๋กœ ์‚ฌ์ „ ํ•™์Šตํ•˜๋„๋ก ์ œ์•ˆํ–ˆ์ง€๋งŒ(Lewis et al., 2020; Maini et al., 2024), ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๋ชฉํ‘œ๋กœ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ์‹์€ ์ด์ „์— ์—ฐ๊ตฌ๋œ ๋ฐ” ์žˆ์ง€๋งŒ(Pfau et al., 2023), ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ๋ชจ๋ธ๊ณผ ๋ฉ€ํ‹ฐํƒœ์Šคํ‚นํ•˜๋Š” ๋ฐฉ์‹์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ด ์ž‘์—…์€ ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ๋ชฉํ‘œ๋กœ ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ ๋Œ€์  ๊ณต๊ฒฉ์„ ์‹๋ณ„ํ•˜๋Š” ์™„์ „ํžˆ ๋‹ค๋ฅธ ๋ชฉํ‘œ๋ฅผ ๋ชฉํ‘œ๋กœ ํ–ˆ์Šต๋‹ˆ๋‹ค. ์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ ๋˜๋Š” ์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋Œ€์‹  ๋งˆ์Šคํ‚น๋œ ์–ธ์–ด ๋ชจ๋ธ์€ ์ฝœ๋กœ๋ฒ„ํŠธ ๋“ฑ(2011)๊ณผ ๊ฐ™์€ ์ดˆ๊ธฐ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ž‘์—…๊ณผ BERT(Devlin et al., 2018)์™€ ๊ฐ™์€ ๋ชจ๋ธ๋กœ ๋Œ์•„๊ฐ€์„œ '์ค‘๊ฐ„์„ ์ฑ„์šฐ๋Š” ๋ฐฉ๋ฒ•'์„ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ์žฌ๋ฐฐ์—ดํ•˜์—ฌ ์ค‘๊ฐ„ ํ…์ŠคํŠธ ์„น์…˜์„ ๋ช…์‹œ์ ์œผ๋กœ ์ฑ„์šฐ๋Š” ๋ฐฉ๋ฒ•(Bavarian et al., 2022), ์Šคํฌ๋žจ๋ธ”๋œ ๋ฐ์ดํ„ฐ๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ๋ฒ•(Sinha et al., 2021), ์ธ์ˆ˜๋ถ„ํ•ด ์ˆœ์„œ์˜ ๋ชจ๋“  ์ˆœ์—ด๋กœ ํ›ˆ๋ จํ•˜๋Š” ๋ฐฉ๋ฒ•(Yang et al., 2019) ๋“ฑ ๋‹ค๋ฅธ ๋ฐฉ๋ฒ•๋„ ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด์™€ ๊ด€๋ จํ•˜์—ฌ ๋ฐ˜๋ณต๋˜๋Š” ์„ธ๊ทธ๋จผํŠธ๋กœ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ๋„ ์–ธ์–ด ๋ชจ๋ธ ์ž„๋ฒ ๋”ฉ์„ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Springer et al., 2024). BERT์™€ ์œ ์‚ฌํ•œ ์ธ์ฝ”๋” ์ „์šฉ ๋ชจ๋ธ์€ ๋ฐ˜์ „ ์ €์ฃผ๋ฅผ ์™„ํ™”ํ•˜์ง€ ๋ชปํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค(Allen-Zhu & Li, 2023). ๊ทธ๋Ÿฌ๋‚˜ ์•„ํ‚คํ…์ฒ˜์™€ ํ›ˆ๋ จ ์ ˆ์ฐจ๋ฅผ ์ˆ˜์ •ํ•˜๋ฉด ์–‘๋ฐฉํ–ฅ ์บ์ฃผ์–ผ ์–ธ์–ด ๋ชจ๋ธ๋ง ์ตœ์ ํ™”(BICO)(Lv et al., 2023)๋ฅผ ๋„์ž…ํ•˜๋Š” ๋“ฑ ๋„์›€์ด ๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. ์ด์™€๋Š” ๋Œ€์กฐ์ ์œผ๋กœ, ์ €ํฌ ์—ฐ๊ตฌ๋Š” ํ‘œ์ค€ ์–ธ์–ด ๋ชจ๋ธ ํ›ˆ๋ จ์„ ํ˜„์žฌ ์ฒด์ œ์™€ ์ตœ๋Œ€ํ•œ ์œ ์‚ฌํ•˜๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.
1. We propose a two-stage instruction tuning method and a dataset curation recipe that can largely enhance LLM's capability of integrating user provided or retrieved context for zero-shot conversational QA tasks. We demonstrate that our method significantly outperforms regular instruction tuning or RLHF-based recipes (e.g., Llama-2-Chat). 2. For RAG in conversational QA, we show that fine-tuning state-of-the-art single-turn query retrievers on human-annotated multi-turn QA dataset works as well as utilizing the state-of-the-art LLM-based query rewriting model, i.e., GPT-3.5-turbo (OpenAI, 2022). 3. We build a family of ChatQA models based on Llama2-7B, Llama2-13B, Llama2-70B (Touvron et al., 2023) and a in-house 8B pretrained GPT. We conduct comprehensive study on 10 conversational QA datasets, including 5 datasets with long documents that need retrieval and 3 datasets with tables. In terms of average score, our ChatQA-70B model (54.14) can outperform GPT-3.5-turbo (50.37) and GPT-4 (53.90) without utilizing any synthetic data from ChatGPT models. 4. We study the "unanswerable" scenario, where the desired answer is not included in the provided or retrieved context, thus the LLM can easily hallucinate. We demonstrate that adding a small amount of "unanswerable" samples in instruction tuning can steer the model to generate "can not answer" output when it is necessary, thus largely reduce hallucination. Our ChatQA-70B outperforms GPT-3.5-turbo in this regard, while still has a slight gap compared to GPT-4 (around 3.5%).
1. ์ €ํฌ๋Š” zero-shot ๋Œ€ํ™”ํ˜• QA ์ž‘์—…์„ ์œ„ํ•ด ์‚ฌ์šฉ์ž๊ฐ€ ์ œ๊ณตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰ํ•œ ์ปจํ…์ŠคํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” LLM์˜ ๊ธฐ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” 2๋‹จ๊ณ„ Instruction Tuning ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํ๋ ˆ์ด์…˜ ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๋ฐฉ๋ฒ•์ด ์ผ๋ฐ˜์ ์ธ Instruction Tuning์ด๋‚˜ RLHF ๊ธฐ๋ฐ˜ ๋ ˆ์‹œํ”ผ(์˜ˆ: Llama-2-Chat)๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 2. ๋Œ€ํ™”ํ˜• QA์˜ RAG์˜ ๊ฒฝ์šฐ, ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์€ ๋‹ค์ค‘ ํ„ด QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ์‹  ์‹ฑ๊ธ€ ํ„ด ์ฟผ๋ฆฌ ๊ฒ€์ƒ‰๊ธฐ๋ฅผ Fine-tuningํ•˜๋Š” ๊ฒƒ์ด ์ตœ์‹  LLM ๊ธฐ๋ฐ˜ ์ฟผ๋ฆฌ ์žฌ์ž‘์„ฑ ๋ชจ๋ธ์ธ GPT-3.5-turbo(OpenAI, 2022)๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๋งŒํผ์ด๋‚˜ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 3. Llama2-7B, Llama2-13B, Llama2-70B(Touvron et al., 2023)์™€ ์‚ฌ๋‚ด 8B ์‚ฌ์ „ ํ•™์Šต๋œ GPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ChatQA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. Retrieval์ด ํ•„์š”ํ•œ ๊ธด ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ…Œ์ด๋ธ”์ด ํฌํ•จ๋œ 3๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํฌํ•จํ•˜์—ฌ 10๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  ์ ์ˆ˜ ๊ธฐ์ค€์œผ๋กœ, ์ €ํฌ์˜ ChatQA-70B ๋ชจ๋ธ(54.14)์€ ChatGPT ๋ชจ๋ธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ GPT-3.5-turbo(50.37) ๋ฐ GPT-4(53.90)๋ฅผ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 4. ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ์— ์›ํ•˜๋Š” ๋‹ต๋ณ€์ด ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์•„ LLM์ด ์‰ฝ๊ฒŒ Hallucination์„ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ๋Š” '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. Instruction Tuning์— ์†Œ๋Ÿ‰์˜ '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์ƒ˜ํ”Œ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ์ด ํ•„์š”ํ•  ๋•Œ '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋„๋ก ์œ ๋„ํ•˜์—ฌ Hallucination์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ์—์„œ ChatQA-70B๋Š” GPT-3.5-turbo๋ฅผ ๋Šฅ๊ฐ€ํ•˜์ง€๋งŒ, GPT-4์— ๋น„ํ•ด ์•ฝ๊ฐ„์˜ ๊ฒฉ์ฐจ(์•ฝ 3.5%)๊ฐ€ ์—ฌ์ „ํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
1. ์ €ํฌ๋Š” ์ œ๋กœ ์ƒท ๋Œ€ํ™”ํ˜• QA ์ž‘์—…์„ ์œ„ํ•ด ์‚ฌ์šฉ์ž๊ฐ€ ์ œ๊ณตํ•˜๊ฑฐ๋‚˜ ๊ฒ€์ƒ‰ํ•œ ์ปจํ…์ŠคํŠธ๋ฅผ ํ†ตํ•ฉํ•˜๋Š” LLM์˜ ๊ธฐ๋Šฅ์„ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š” 2๋‹จ๊ณ„ ๋ช…๋ น ํŠœ๋‹ ๋ฐฉ๋ฒ•๊ณผ ๋ฐ์ดํ„ฐ ์„ธํŠธ ํ๋ ˆ์ด์…˜ ๋ ˆ์‹œํ”ผ๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ์˜ ๋ฐฉ๋ฒ•์ด ์ผ๋ฐ˜์ ์ธ ์ง€์‹œ ํŠœ๋‹์ด๋‚˜ RLHF ๊ธฐ๋ฐ˜ ๋ ˆ์‹œํ”ผ(์˜ˆ: Llama-2-Chat)๋ณด๋‹ค ํ›จ์”ฌ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 2. ๋Œ€ํ™”ํ˜• QA์˜ RAG์˜ ๊ฒฝ์šฐ, ์ธ๊ฐ„์ด ์ฃผ์„์„ ๋‹ฌ์€ ๋‹ค์ค‘ ํ„ด QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ตœ์‹  ์‹ฑ๊ธ€ ํ„ด ์ฟผ๋ฆฌ ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ด ์ตœ์‹  LLM ๊ธฐ๋ฐ˜ ์ฟผ๋ฆฌ ์žฌ์ž‘์„ฑ ๋ชจ๋ธ์ธ GPT-3.5-turbo(OpenAI, 2022)๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ๋งŒํผ์ด๋‚˜ ํšจ๊ณผ์ ์ด๋ผ๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. 3. Llama2-7B, Llama2-13B, Llama2-70B(Touvron et al., 2023)์™€ ์‚ฌ๋‚ด 8B ์‚ฌ์ „ ํ•™์Šต๋œ GPT๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ChatQA ๋ชจ๋ธ ์ œํ’ˆ๊ตฐ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰์ด ํ•„์š”ํ•œ ๊ธด ๋ฌธ์„œ๊ฐ€ ํฌํ•จ๋œ 5๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ํ…Œ์ด๋ธ”์ด ํฌํ•จ๋œ 3๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํฌํ•จํ•˜์—ฌ 10๊ฐœ์˜ ๋Œ€ํ™”ํ˜• QA ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋Œ€ํ•œ ํฌ๊ด„์ ์ธ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ‰๊ท  ์ ์ˆ˜ ๊ธฐ์ค€์œผ๋กœ, ์ €ํฌ์˜ ChatQA-70B ๋ชจ๋ธ(54.14)์€ ChatGPT ๋ชจ๋ธ์˜ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ GPT-3.5-turbo(50.37) ๋ฐ GPT-4(53.90)๋ฅผ ๋Šฅ๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 4. ์ œ๊ณต๋œ ์ปจํ…์ŠคํŠธ๋‚˜ ๊ฒ€์ƒ‰๋œ ์ปจํ…์ŠคํŠธ์— ์›ํ•˜๋Š” ๋‹ต๋ณ€์ด ํฌํ•จ๋˜์–ด ์žˆ์ง€ ์•Š์•„ LLM์ด ์‰ฝ๊ฒŒ ํ™€๋ฃจ์‹ฑ์„ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ๋Š” '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ง€์‹œ ํŠœ๋‹์— ์†Œ๋Ÿ‰์˜ '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์ƒ˜ํ”Œ์„ ์ถ”๊ฐ€ํ•˜๋ฉด ๋ชจ๋ธ์ด ํ•„์š”ํ•  ๋•Œ '๋‹ต๋ณ€ ๋ถˆ๊ฐ€' ์ถœ๋ ฅ์„ ์ƒ์„ฑํ•˜๋„๋ก ์œ ๋„ํ•˜์—ฌ ํ™€๋ฃจ์‹ฑ์„ ํฌ๊ฒŒ ์ค„์ผ ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ์—์„œ ChatQA-70B๋Š” GPT-3.5-turbo๋ฅผ ๋Šฅ๊ฐ€ํ•˜์ง€๋งŒ, GPT-4์— ๋น„ํ•ด ์•ฝ๊ฐ„์˜ ๊ฒฉ์ฐจ(์•ฝ 3.5%)๊ฐ€ ์—ฌ์ „ํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_Fine-tuning with LLM Generated Datasets_. Building human annotated dataset needs to recruit people, which can be costly, especially when one needs to annotate a large amount of samples. Considering that LLMs can achieve human-like capabilities in a wide range of tasks, a natural idea is using LLMs to accomplish the annotation task. While the datasets produced from this method can be not as perfect as the human annotated ones, it is much cheaper, and can be leveraged to generate more samples. For example, in ToolBench [14], to enhance the tool-using capability of open-source LLMs, the authors collect 16,464 real-world APIs spanning 49 categories from the RapidAPI Hub. They used these APIs to prompt ChatGPT to generate diverse instructions, covering both single-tool and multi-tool scenarios. Based on the obtained dataset, the authors fine-tune LLaMA [9], and obtain significant performance improvement in terms of tool using. In [82], to empower the agent with social capability, the authors design a sandbox, and deploy multiple agents to interact with each other. Given a social question, the central agent first generates initial responses. Then, it shares the responses to its nearby agents for collecting their feedback. Based on the feedback as well as its detailed explanations, the central agent revise its initial responses to make them more consistent with social norms. In this process, the authors collect a large amount of agent social interaction data, which is then leveraged to fine-tune the LLMs.
\(\bullet\)_LLM ์ƒ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ์˜ Fine-tuning_. ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹ฌ ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ํŠนํžˆ ๋งŽ์€ ์–‘์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•  ๋•Œ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์—ญ๋Ÿ‰์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ž์—ฐ์Šค๋Ÿฌ์šด ์•„์ด๋””์–ด๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ์„ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒํผ ์™„๋ฒฝํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ํ›จ์”ฌ ์ €๋ ดํ•˜๊ณ  ๋” ๋งŽ์€ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ToolBench [14]์—์„œ๋Š” ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ๋„๊ตฌ ์‚ฌ์šฉ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ €์ž๋“ค์€ RapidAPI ํ—ˆ๋ธŒ์—์„œ 49๊ฐœ ์นดํ…Œ๊ณ ๋ฆฌ์— ๊ฑธ์ณ 16,464๊ฐœ์˜ ์‹ค์ œ API๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ChatGPT์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ๋‹จ์ผ ๋„๊ตฌ ๋ฐ ๋‹ค์ค‘ ๋„๊ตฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ๋‹ค์–‘ํ•œ Instructions์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ป์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ €์ž๋“ค์€ LLaMA [9]๋ฅผ Fine-tuningํ•˜์—ฌ ๋„๊ตฌ ์‚ฌ์šฉ ์ธก๋ฉด์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. [82]์—์„œ๋Š” ์—์ด์ „ํŠธ์—๊ฒŒ ์†Œ์…œ ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•˜๊ธฐ ์œ„ํ•ด ์ƒŒ๋“œ๋ฐ•์Šค๋ฅผ ์„ค๊ณ„ํ•˜๊ณ  ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ๋ฐฐํฌํ•˜์—ฌ ์„œ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ์…œ ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด ์ค‘์•™ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ด ์‘๋‹ต์„ ์ฃผ๋ณ€ ์—์ด์ „ํŠธ์™€ ๊ณต์œ ํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ์ค‘์•™ ์—์ด์ „ํŠธ๋Š” ํ”ผ๋“œ๋ฐฑ๊ณผ ์ž์„ธํ•œ ์„ค๋ช…์„ ๋ฐ”ํƒ•์œผ๋กœ ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ˆ˜์ •ํ•˜์—ฌ ์‚ฌํšŒ์  ๊ทœ๋ฒ”๊ณผ ๋” ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ์ €์ž๋“ค์€ ๋Œ€๋Ÿ‰์˜ ์—์ด์ „ํŠธ ์†Œ์…œ ์ƒํ˜ธ์ž‘์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ํ›„ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ Fine-tuningํ•ฉ๋‹ˆ๋‹ค.
\(\bullet\)_LLM ์ƒ์„ฑ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ์˜ ๋ฏธ์„ธ ์กฐ์ •_. ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹ฌ ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋ ค๋ฉด ํŠนํžˆ ๋งŽ์€ ์–‘์˜ ์ƒ˜ํ”Œ์— ๋Œ€ํ•œ ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•  ๋•Œ ๋น„์šฉ์ด ๋งŽ์ด ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. LLM์ด ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์—ญ๋Ÿ‰์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๊ณ ๋ คํ•  ๋•Œ, ์ž์—ฐ์Šค๋Ÿฌ์šด ์•„์ด๋””์–ด๋Š” LLM์„ ์‚ฌ์šฉํ•˜์—ฌ ์ฃผ์„ ์ž‘์—…์„ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์‚ฌ๋žŒ์ด ์ฃผ์„์„ ๋‹จ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋งŒํผ ์™„๋ฒฝํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ํ›จ์”ฌ ์ €๋ ดํ•˜๊ณ  ๋” ๋งŽ์€ ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ToolBench [14]์—์„œ๋Š” ์˜คํ”ˆ ์†Œ์Šค LLM์˜ ๋„๊ตฌ ์‚ฌ์šฉ ๊ธฐ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ์ €์ž๋“ค์€ RapidAPI ํ—ˆ๋ธŒ์—์„œ 49๊ฐœ ์นดํ…Œ๊ณ ๋ฆฌ์— ๊ฑธ์ณ 16,464๊ฐœ์˜ ์‹ค์ œ API๋ฅผ ์ˆ˜์ง‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ API๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ChatGPT์— ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด๋‚ด ๋‹จ์ผ ๋„๊ตฌ ๋ฐ ๋‹ค์ค‘ ๋„๊ตฌ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋ชจ๋‘ ํฌํ•จํ•˜๋Š” ๋‹ค์–‘ํ•œ ์ง€์นจ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์–ป์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ €์ž๋“ค์€ LLaMA [9]๋ฅผ ๋ฏธ์„ธ ์กฐ์ •ํ•˜์—ฌ ๋„๊ตฌ ์‚ฌ์šฉ ์ธก๋ฉด์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. [82]์—์„œ๋Š” ์—์ด์ „ํŠธ์—๊ฒŒ ์†Œ์…œ ๊ธฐ๋Šฅ์„ ๋ถ€์—ฌํ•˜๊ธฐ ์œ„ํ•ด ์ƒŒ๋“œ๋ฐ•์Šค๋ฅผ ์„ค๊ณ„ํ•˜๊ณ  ์—ฌ๋Ÿฌ ์—์ด์ „ํŠธ๋ฅผ ๋ฐฐํฌํ•˜์—ฌ ์„œ๋กœ ์ƒํ˜ธ ์ž‘์šฉํ•˜๋„๋ก ํ–ˆ์Šต๋‹ˆ๋‹ค. ์†Œ์…œ ์งˆ๋ฌธ์ด ์ฃผ์–ด์ง€๋ฉด ์ค‘์•™ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ ๋‹ค์Œ, ์ด ์‘๋‹ต์„ ์ฃผ๋ณ€ ์—์ด์ „ํŠธ์™€ ๊ณต์œ ํ•˜์—ฌ ํ”ผ๋“œ๋ฐฑ์„ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ์ค‘์•™ ์—์ด์ „ํŠธ๋Š” ํ”ผ๋“œ๋ฐฑ๊ณผ ์ž์„ธํ•œ ์„ค๋ช…์„ ๋ฐ”ํƒ•์œผ๋กœ ์ดˆ๊ธฐ ์‘๋‹ต์„ ์ˆ˜์ •ํ•˜์—ฌ ์‚ฌํšŒ์  ๊ทœ๋ฒ”๊ณผ ๋” ์ผ์น˜ํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ์ด ๊ณผ์ •์—์„œ ์ €์ž๋“ค์€ ๋Œ€๋Ÿ‰์˜ ์—์ด์ „ํŠธ ์†Œ์…œ ์ƒํ˜ธ์ž‘์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•œ ํ›„ ์ด๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ๋ฏธ์„ธ ์กฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
In this paper, we demonstrate how Large Language Models (LLMs) can effectively learn to use an off-the-shelf information retrieval (IR) system specifically when additional context is required to answer a given question. Given the performance of IR systems, the optimal strategy for question answering does not always entail external information retrieval; rather, it often involves leveraging the parametric memory of the LLM itself. Prior research has identified this phenomenon in the PopQA dataset, wherein the most popular questions are effectively addressed using the LLM's parametric memory, while less popular ones require IR system usage. Following this, we propose a tailored training approach for LLMs, leveraging existing open-domain question answering datasets. Here, LLMs are trained to generate a special token, \(\langle\text{RET}\rangle\), when they do not know the answer to a question. Our evaluation of the Adaptive Retrieval LLM (Adapt-LLM) on the PopQA dataset showcases improvements over the same LLM under three configurations: (i) retrieving information for all the questions, (ii) using always the parametric memory of the LLM, and (iii) using a popularity threshold to decide when to use a retriever. Through our analysis, we demonstrate that Adapt-LLM is able to generate the \(\langle\text{RET}\rangle\) token when it determines that it does not know how to answer a question, indicating the need for IR, while it achieves notably high accuracy levels when it chooses to rely only on its parametric memory.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํŠน์ • ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ Large Language Models (LLMs)์ด ์ฆ‰์‹œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด ๊ฒ€์ƒ‰(IR) ์‹œ์Šคํ…œ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. IR ์‹œ์Šคํ…œ์˜ ์„ฑ๋Šฅ์„ ๊ณ ๋ คํ•  ๋•Œ, ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์œ„ํ•œ ์ตœ์ ์˜ ์ „๋žต์ด ํ•ญ์ƒ ์™ธ๋ถ€ ์ •๋ณด ๊ฒ€์ƒ‰์„ ์ˆ˜๋ฐ˜ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์˜คํžˆ๋ ค LLM ์ž์ฒด์˜ parametric memory๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ๊ณผ์ ์ธ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์งˆ๋ฌธ์€ LLM์˜ parametric memory๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ธ๊ธฐ๊ฐ€ ๋‚ฎ์€ ์งˆ๋ฌธ์€ IR ์‹œ์Šคํ…œ ์‚ฌ์šฉ์ด ํ•„์š”ํ•œ PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๊ธฐ์กด์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ question answering ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ์œ„ํ•œ ๋งž์ถคํ˜• ํ›ˆ๋ จ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ LLM์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ๋ชจ๋ฅผ ๋•Œ ํŠน์ˆ˜ ํ† ํฐ์ธ \(\langle\text{RET}\rangle\)์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ Adaptive Retrieval LLM (Adapt-LLM)์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์„ธ ๊ฐ€์ง€ ๊ตฌ์„ฑ์—์„œ ๋™์ผํ•œ LLM๋ณด๋‹ค ๊ฐœ์„ ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค: (i) ๋ชจ๋“  ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰, (ii) ํ•ญ์ƒ LLM์˜ parametric memory๋ฅผ ์‚ฌ์šฉ, (iii) ์ธ๊ธฐ๋„ ์ž„๊ณ„๊ฐ’์„ ์‚ฌ์šฉํ•˜์—ฌ retriever๋ฅผ ์–ธ์ œ ์‚ฌ์šฉํ• ์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„์„ ํ†ตํ•ด Adapt-LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ๋ฐฉ๋ฒ•์„ ๋ชจ๋ฅธ๋‹ค๊ณ  ํŒ๋‹จํ•  ๋•Œ \(\langle\text{RET}\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•˜์—ฌ IR์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋™์‹œ์— parametric memory์—๋งŒ ์˜์กดํ•˜๊ธฐ๋กœ ์„ ํƒํ–ˆ์„ ๋•Œ ํ˜„์ €ํžˆ ๋†’์€ ์ •ํ™•๋„ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฐฑ์„œ์—์„œ๋Š” ํŠน์ • ์งˆ๋ฌธ์— ๋‹ตํ•˜๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ์ปจํ…์ŠคํŠธ๊ฐ€ ํ•„์š”ํ•œ ๊ฒฝ์šฐ ๋Œ€์šฉ๋Ÿ‰ ์–ธ์–ด ๋ชจ๋ธ(LLM)์ด ์ฆ‰์‹œ ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ์ •๋ณด ๊ฒ€์ƒ‰(IR) ์‹œ์Šคํ…œ์„ ํšจ๊ณผ์ ์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ ์ค๋‹ˆ๋‹ค. IR ์‹œ์Šคํ…œ์˜ ์„ฑ๋Šฅ์„ ๊ณ ๋ คํ•  ๋•Œ, ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์œ„ํ•œ ์ตœ์ ์˜ ์ „๋žต์ด ํ•ญ์ƒ ์™ธ๋ถ€ ์ •๋ณด ๊ฒ€์ƒ‰์„ ์ˆ˜๋ฐ˜ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹ˆ๋ฉฐ, ์˜คํžˆ๋ ค LLM ์ž์ฒด์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ํ™œ์šฉํ•˜๋Š” ๊ฒƒ์ด ๋” ํšจ๊ณผ์ ์ธ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. ์ด์ „ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” ์งˆ๋ฌธ์€ LLM์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ˜๋ฉด, ์ธ๊ธฐ๊ฐ€ ๋‚ฎ์€ ์งˆ๋ฌธ์€ IR ์‹œ์Šคํ…œ ์‚ฌ์šฉ์ด ํ•„์š”ํ•œ PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ด๋Ÿฌํ•œ ํ˜„์ƒ์„ ํ™•์ธํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ๊ธฐ์กด์˜ ์˜คํ”ˆ ๋„๋ฉ”์ธ ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ™œ์šฉํ•˜์—ฌ LLM์„ ์œ„ํ•œ ๋งž์ถคํ˜• ํ›ˆ๋ จ ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์„œ LLM์€ ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต์„ ๋ชจ๋ฅผ ๋•Œ ํŠน์ˆ˜ ํ† ํฐ์ธ \(\langle\text{RET}\rangle\)์„ ์ƒ์„ฑํ•˜๋„๋ก ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. PopQA ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ์ ์‘ํ˜• ๊ฒ€์ƒ‰ LLM(Adapt-LLM)์„ ํ‰๊ฐ€ํ•œ ๊ฒฐ๊ณผ, ์„ธ ๊ฐ€์ง€ ๊ตฌ์„ฑ์—์„œ ๋™์ผํ•œ LLM๋ณด๋‹ค ๊ฐœ์„ ๋œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค: (i) ๋ชจ๋“  ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰, (ii) ํ•ญ์ƒ LLM์˜ ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์‚ฌ์šฉ, (iii) ์ธ๊ธฐ๋„ ์ž„๊ณ„๊ฐ’์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฒ€์ƒ‰๊ธฐ๋ฅผ ์–ธ์ œ ์‚ฌ์šฉํ• ์ง€ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋ถ„์„์„ ํ†ตํ•ด Adapt-LLM์ด ์งˆ๋ฌธ์— ๋Œ€ํ•œ ๋‹ต๋ณ€ ๋ฐฉ๋ฒ•์„ ๋ชจ๋ฅธ๋‹ค๊ณ  ํŒ๋‹จํ•  ๋•Œ \(\langle\text{RET}\rangle\) ํ† ํฐ์„ ์ƒ์„ฑํ•˜์—ฌ IR์ด ํ•„์š”ํ•จ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋™์‹œ์— ํŒŒ๋ผ๋ฉ”ํŠธ๋ฆญ ๋ฉ”๋ชจ๋ฆฌ์—๋งŒ ์˜์กดํ•˜๊ธฐ๋กœ ์„ ํƒํ–ˆ์„ ๋•Œ ํ˜„์ €ํžˆ ๋†’์€ ์ •ํ™•๋„ ์ˆ˜์ค€์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค.
Recently, there have been significant advancements in large language models (LLMs), particularly focused on the English language. These advancements have enabled these LLMs to understand and execute complex instructions with unprecedented accuracy and fluency. However, despite these advancements, there remains a noticeable gap in the development of Chinese instruction tuning. The unique linguistic features and cultural depth of the Chinese language pose challenges for instruction tuning tasks. Existing datasets are either derived from English-centric LLMs or are ill-suited for aligning with the interaction patterns of real-world Chinese users. To bridge this gap, we introduce COIG-CQIA, a high-quality Chinese instruction tuning dataset. Our aim is to build a diverse, wide-ranging instruction-tuning dataset to better align model behavior with human interactions. To this end, we collect a high-quality human-written corpus from various sources on the Chinese Internet, including Q&A communities, Wikis, examinations, and existing NLP datasets. This corpus was rigorously filtered and carefully processed to form the COIG-CQIA dataset. Furthermore, we train models of various scales on different subsets of CQIA, following in-depth evaluation and analyses. The findings from our experiments offer valuable insights for selecting and developing Chinese instruction-tuning datasets. We also find that models trained on CQIA-Subset achieve competitive results in human assessment as well as knowledge and security benchmarks.
์ตœ๊ทผ Large Language Models (LLMs)์€ ํŠนํžˆ ์˜์–ด์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐœ์ „์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ LLM์€ ์ „๋ก€ ์—†๋Š” ์ •ํ™•์„ฑ๊ณผ ์œ ์ฐฝ์„ฑ์œผ๋กœ ๋ณต์žกํ•œ Instructions์„ ์ดํ•ดํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ค‘๊ตญ์–ด Instruction Tuning ๊ฐœ๋ฐœ์—๋Š” ๋ˆˆ์— ๋„๋Š” ๊ฒฉ์ฐจ๊ฐ€ ์—ฌ์ „ํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ณ ์œ ํ•œ ์–ธ์–ด์  ํŠน์ง•๊ณผ ๋ฌธํ™”์  ๊นŠ์ด๋กœ ์ธํ•ด Instruction Tuning ์ž‘์—…์— ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์˜์–ด ์ค‘์‹ฌ์˜ LLM์—์„œ ํŒŒ์ƒ๋˜์—ˆ๊ฑฐ๋‚˜ ์‹ค์ œ ์ค‘๊ตญ ์‚ฌ์šฉ์ž์˜ ์ƒํ˜ธ ์ž‘์šฉ ํŒจํ„ด๊ณผ ์ผ์น˜ํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ COIG-CQIA๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” ๋ชจ๋ธ ๋™์ž‘์„ ์ธ๊ฐ„๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ๋” ์ž˜ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•˜๊ณ  ๊ด‘๋ฒ”์œ„ํ•œ Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด Q&A ์ปค๋ฎค๋‹ˆํ‹ฐ, ์œ„ํ‚ค, ์‹œํ—˜, ๊ธฐ์กด NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋“ฑ ์ค‘๊ตญ ์ธํ„ฐ๋„ท์˜ ๋‹ค์–‘ํ•œ ์†Œ์Šค์—์„œ ์ˆ˜์ง‘ํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์‚ฌ๋žŒ์ด ์ง์ ‘ ์ž‘์„ฑํ•œ ์ฝ”ํผ์Šค๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์‹ ์ค‘ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜์—ฌ COIG-CQIA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹ฌ๋„ ์žˆ๋Š” ํ‰๊ฐ€์™€ ๋ถ„์„์„ ๊ฑฐ์ณ ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์„ CQIA์˜ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ์ค‘๊ตญ์–ด Instruction Tuning ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์„ ํƒํ•˜๊ณ  ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๊ท€์ค‘ํ•œ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ CQIA-Subset์—์„œ ํ•™์Šตํ•œ ๋ชจ๋ธ์ด ์ง€์‹ ๋ฐ ๋ณด์•ˆ ๋ฒค์น˜๋งˆํฌ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ๋„ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ์‚ฌ์‹ค๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ํŠนํžˆ ์˜์–ด์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ƒ๋‹นํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐœ์ „์œผ๋กœ ์ธํ•ด ์ด๋Ÿฌํ•œ LLM์€ ์ „๋ก€ ์—†๋Š” ์ •ํ™•์„ฑ๊ณผ ์œ ์ฐฝ์„ฑ์œผ๋กœ ๋ณต์žกํ•œ ์ง€์นจ์„ ์ดํ•ดํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Ÿฌํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ค‘๊ตญ์–ด ์ง€์นจ ํŠœ๋‹ ๊ฐœ๋ฐœ์—๋Š” ๋ˆˆ์— ๋„๋Š” ๊ฒฉ์ฐจ๊ฐ€ ์—ฌ์ „ํžˆ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค. ์ค‘๊ตญ์–ด์˜ ๊ณ ์œ ํ•œ ์–ธ์–ด์  ํŠน์ง•๊ณผ ๋ฌธํ™”์  ๊นŠ์ด๋กœ ์ธํ•ด ์ง€์นจ ํŠœ๋‹ ์ž‘์—…์— ์–ด๋ ค์›€์ด ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ์˜์–ด ์ค‘์‹ฌ์˜ LLM์—์„œ ํŒŒ์ƒ๋˜์—ˆ๊ฑฐ๋‚˜ ์‹ค์ œ ์ค‘๊ตญ ์‚ฌ์šฉ์ž์˜ ์ƒํ˜ธ ์ž‘์šฉ ํŒจํ„ด๊ณผ ์ผ์น˜ํ•˜๋Š” ๋ฐ ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒฉ์ฐจ๋ฅผ ํ•ด์†Œํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ํ’ˆ์งˆ ์ค‘๊ตญ์–ด ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ COIG-CQIA๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ชฉํ‘œ๋Š” ๋ชจ๋ธ ๋™์ž‘์„ ์ธ๊ฐ„๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ์— ๋” ์ž˜ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•˜๊ณ  ๊ด‘๋ฒ”์œ„ํ•œ ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด Q&A ์ปค๋ฎค๋‹ˆํ‹ฐ, ์œ„ํ‚ค, ์‹œํ—˜, ๊ธฐ์กด NLP ๋ฐ์ดํ„ฐ ์„ธํŠธ ๋“ฑ ์ค‘๊ตญ ์ธํ„ฐ๋„ท์˜ ๋‹ค์–‘ํ•œ ์†Œ์Šค์—์„œ ์ˆ˜์ง‘ํ•œ ๊ณ ํ’ˆ์งˆ์˜ ์‚ฌ๋žŒ์ด ์ง์ ‘ ์ž‘์„ฑํ•œ ์ฝ”ํผ์Šค๋ฅผ ์—„๊ฒฉํ•˜๊ฒŒ ํ•„ํ„ฐ๋งํ•˜๊ณ  ์‹ ์ค‘ํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜์—ฌ COIG-CQIA ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹ฌ๋„ ์žˆ๋Š” ํ‰๊ฐ€์™€ ๋ถ„์„์„ ๊ฑฐ์ณ ๋‹ค์–‘ํ•œ ๊ทœ๋ชจ์˜ ๋ชจ๋ธ์„ CQIA์˜ ๋‹ค์–‘ํ•œ ํ•˜์œ„ ์ง‘ํ•ฉ์— ๋Œ€ํ•ด ํ•™์Šต์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ์ค‘๊ตญ์–ด ์ง€์นจ ํŠœ๋‹ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์„ ํƒํ•˜๊ณ  ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๊ท€์ค‘ํ•œ ํ†ต์ฐฐ๋ ฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ CQIA-Subset์—์„œ ํ•™์Šตํ•œ ๋ชจ๋ธ์ด ์ง€์‹ ๋ฐ ๋ณด์•ˆ ๋ฒค์น˜๋งˆํฌ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ธ๊ฐ„ ํ‰๊ฐ€์—์„œ๋„ ๊ฒฝ์Ÿ๋ ฅ ์žˆ๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•œ๋‹ค๋Š” ์‚ฌ์‹ค๋„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
**Data Curriculum.** After preparing the data mixture, it is important to schedule the order that specific data is presented to LLMs for pre-training. It has been shown that, in some cases, to learn a certain skill, learning in a skill-set sequence (_e.g.,_ basic skills \(\rightarrow\) target skill) outperforms direct learning from a corpus focused solely on the target skill [234, 235]. Following the idea of curriculum learning [236], _data curriculum_ has been proposed and widely used in model pre-training [234, 235, 237, 238]. It aims to organize different parts of pre-training data for LLMs in a specific order, _e.g.,_ starting with easy/general examples and progressively introducing more challenging/specialized ones. More generally, it can broadly refer to the adaptive adjustment of data proportions for different sources during pre-training. Existing work about data curriculum mainly focuses on continual pre-training, such as specialized coding LLMs (_e.g.,_ CodeLLMa [235]) or long context LLMs (_e.g.,_ LongLLaMA [238]). However, it still lacks of more detailed report about data curriculum for general-purpose LLMs (_e.g.,_ LLaMA) in the literature. To determine data curriculum, a practical approach is to monitor the development of key abilities of LLMs based on specially constructed evaluation benchmarks, and then adaptively adjust the data mixture during pre-training. Next, we take three common abilities as examples to introduce how the concept of data curriculum23 applies in continual pre-training.
**Data Curriculum**. ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์„ ์ค€๋น„ํ•œ ํ›„์—๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ํŠน์ • ๋ฐ์ดํ„ฐ๋ฅผ LLM์— ์ œ์‹œํ•˜๋Š” ์ˆœ์„œ๋ฅผ ์˜ˆ์•ฝํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ๊ฒฝ์šฐ์—๋Š” ํŠน์ • ๊ธฐ์ˆ ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์ˆ  ์„ธํŠธ ์ˆœ์„œ(_์˜ˆ:_ ๊ธฐ๋ณธ ๊ธฐ์ˆ  \(\rightarrow\) ๋ชฉํ‘œ ๊ธฐ์ˆ )๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ ๊ธฐ์ˆ ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ˜ ์ฝ”ํผ์Šค์—์„œ ์ง์ ‘ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[234, 235]. ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต [236]์˜ ์•„์ด๋””์–ด์— ๋”ฐ๋ผ ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต [234, 235, 237, 238]์—์„œ _data curriculum_์ด ์ œ์•ˆ๋˜๊ณ  ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” LLM์„ ์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์—ฌ๋Ÿฌ ๋ถ€๋ถ„์„ ํŠน์ • ์ˆœ์„œ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ:_ ์‰ฌ์šด/์ผ๋ฐ˜์ ์ธ ์˜ˆ์ œ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ์ ์ฐจ ๋” ์–ด๋ ต๊ณ  ์ „๋ฌธ์ ์ธ ์˜ˆ์ œ๋ฅผ ์†Œ๊ฐœํ•˜๋Š” ๊ฒƒ). ๋” ์ผ๋ฐ˜์ ์œผ๋กœ, ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋‹ค์–‘ํ•œ ์†Œ์Šค์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ๋น„์œจ์„ ์ ์‘์ ์œผ๋กœ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ง€์นญํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. data curriculum์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์ „๋ฌธ ์ฝ”๋”ฉ LLM(_์˜ˆ:_ CodeLLMa [235]) ๋˜๋Š” ๊ธด ์ปจํ…์ŠคํŠธ LLM(_์˜ˆ:_ LongLLaMA [238])๊ณผ ๊ฐ™์€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜ ๋ชฉ์  LLM(_์˜ˆ:_ LLaMA)์— ๋Œ€ํ•œ data curriculum์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ž์„ธํ•œ ๋ณด๊ณ ์„œ๋Š” ์•„์ง ๋ฌธํ—Œ์— ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. data curriculum์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์‹ค์šฉ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน๋ณ„ํžˆ ๊ตฌ์„ฑ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์˜ ์ฃผ์š” ๋Šฅ๋ ฅ ๊ฐœ๋ฐœ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•œ ๋‹ค์Œ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์„ ์ ์‘์ ์œผ๋กœ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, data curriculum์˜ ๊ฐœ๋…์ด ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์–ด๋–ป๊ฒŒ ์ ์šฉ๋˜๋Š”์ง€ ์†Œ๊ฐœํ•˜๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ์„ ์˜ˆ๋กœ ๋“ค์–ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
**๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ**. ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ๋ฌผ์„ ์ค€๋น„ํ•œ ํ›„์—๋Š” ์‚ฌ์ „ ํ•™์Šต์„ ์œ„ํ•ด ํŠน์ • ๋ฐ์ดํ„ฐ๋ฅผ LLM์— ์ œ์‹œํ•˜๋Š” ์ˆœ์„œ๋ฅผ ์˜ˆ์•ฝํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์–ด๋–ค ๊ฒฝ์šฐ์—๋Š” ํŠน์ • ๊ธฐ์ˆ ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์ˆ  ์„ธํŠธ ์ˆœ์„œ(_์˜ˆ:_ ๊ธฐ๋ณธ ๊ธฐ์ˆ  \(\rightarrow\) ๋ชฉํ‘œ ๊ธฐ์ˆ )๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ ๊ธฐ์ˆ ์—๋งŒ ์ดˆ์ ์„ ๋งž์ถ˜ ์ฝ”ํผ์Šค์—์„œ ์ง์ ‘ ํ•™์Šตํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค[234, 235]. ์ปค๋ฆฌํ˜๋Ÿผ ํ•™์Šต [236]์˜ ์•„์ด๋””์–ด์— ๋”ฐ๋ผ ๋ชจ๋ธ ์‚ฌ์ „ ํ•™์Šต [234, 235, 237, 238]์—์„œ _๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ_์ด ์ œ์•ˆ๋˜๊ณ  ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” LLM์„ ์œ„ํ•œ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ์—ฌ๋Ÿฌ ๋ถ€๋ถ„์„ ํŠน์ • ์ˆœ์„œ๋กœ ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค(_์˜ˆ:_ ์‰ฌ์šด/์ผ๋ฐ˜์ ์ธ ์˜ˆ์ œ๋ถ€ํ„ฐ ์‹œ์ž‘ํ•˜์—ฌ ์ ์ฐจ ๋” ์–ด๋ ต๊ณ  ์ „๋ฌธ์ ์ธ ์˜ˆ์ œ๋ฅผ ์†Œ๊ฐœํ•˜๋Š” ๊ฒƒ). ๋” ์ผ๋ฐ˜์ ์œผ๋กœ, ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋‹ค์–‘ํ•œ ์†Œ์Šค์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ๋น„์œจ์„ ์ ์‘์ ์œผ๋กœ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์„ ๊ด‘๋ฒ”์œ„ํ•˜๊ฒŒ ์ง€์นญํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์— ๋Œ€ํ•œ ๊ธฐ์กด ์—ฐ๊ตฌ๋Š” ์ฃผ๋กœ ์ „๋ฌธ ์ฝ”๋”ฉ LLM(_์˜ˆ:_ CodeLLMa [235]) ๋˜๋Š” ๊ธด ์ปจํ…์ŠคํŠธ LLM(_์˜ˆ:_ LongLLaMA [238])๊ณผ ๊ฐ™์€ ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ผ๋ฐ˜ ๋ชฉ์  LLM(_์˜ˆ:_ LLaMA)์— ๋Œ€ํ•œ ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์— ๋Œ€ํ•œ ๋ณด๋‹ค ์ž์„ธํ•œ ๋ณด๊ณ ์„œ๋Š” ์•„์ง ๋ฌธํ—Œ์— ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ์„ ๊ฒฐ์ •ํ•˜๊ธฐ ์œ„ํ•œ ์‹ค์šฉ์ ์ธ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํŠน๋ณ„ํžˆ ๊ตฌ์„ฑ๋œ ํ‰๊ฐ€ ๋ฒค์น˜๋งˆํฌ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ LLM์˜ ์ฃผ์š” ๋Šฅ๋ ฅ ๊ฐœ๋ฐœ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•œ ๋‹ค์Œ ์‚ฌ์ „ ํ•™์Šต ์ค‘์— ๋ฐ์ดํ„ฐ ํ˜ผํ•ฉ์„ ์ ์‘์ ์œผ๋กœ ์กฐ์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค์Œ์œผ๋กœ, ๋ฐ์ดํ„ฐ ์ปค๋ฆฌํ˜๋Ÿผ23์˜ ๊ฐœ๋…์ด ์ง€์†์ ์ธ ์‚ฌ์ „ ํ•™์Šต์— ์–ด๋–ป๊ฒŒ ์ ์šฉ๋˜๋Š”์ง€ ์†Œ๊ฐœํ•˜๊ธฐ ์œ„ํ•ด ์„ธ ๊ฐ€์ง€ ์ผ๋ฐ˜์ ์ธ ๋Šฅ๋ ฅ์„ ์˜ˆ๋กœ ๋“ค์–ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
forms other sequence models by at least \(23\%\) in terms of mean squared error (MSE) across multiple real-world sensor datasets. LOCOST [168] introduced an encoder-decoder architecture based on state-space models for conditional text generation tasks with long-context inputs. This approach effectively reduced computational complexity and memory usage, significantly enhancing the speed of both training and inference stages. MambaStock [178] utilizes the structured state space (S4) architecture to capture the nonlinearity in stock data, enabling accurate predictions of future stock prices. Lu et al. [163] proposed an improvement to the simplified structured state-space sequence model (S5), enabling the reset of hidden states within trajectories during the model training phase. Specifically, to enable the model to handle variable-length sequences, this paper modifies the association operator and introduces a reset annotation that preserves association properties in S5. Additionally, to test the generalization ability of the model, a challenging Meta-RL setup is also introduced. [188] introduces a new method for developing a realistic digital dynamic range compressor model for digital audio production by analyzing its simulation prototype. The learned representations are often affected by the high order of the model, which makes them unsuitable for control design. [189] proposes a system theory based model order reduction technique specifically for linear dynamic blocks in SSM to address this challenge.
์—ฌ๋Ÿฌ ์‹ค์ œ ์„ผ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ‰๊ท  ์ œ๊ณฑ ์˜ค์ฐจ(MSE) ์ธก๋ฉด์—์„œ ์ตœ์†Œ \(23\%\) ์ด์ƒ์˜ ๋‹ค๋ฅธ ์‹œํ€€์Šค ๋ชจ๋ธ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. LOCOST [168]๋Š” ๊ธด ์ปจํ…์ŠคํŠธ ์ž…๋ ฅ์ด ์žˆ๋Š” ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด state-space ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Encoder-Decoder ์•„ํ‚คํ…์ฒ˜๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์—ฌ ํ›ˆ๋ จ๊ณผ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ์†๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. MambaStock [178]์€ structured state space (S4) ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ฃผ์‹ ๋ฐ์ดํ„ฐ์˜ ๋น„์„ ํ˜•์„ฑ์„ ํฌ์ฐฉํ•˜์—ฌ ํ–ฅํ›„ ์ฃผ๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Lu et al. [163]์€ ๋‹จ์ˆœํ™”๋œ structured state-space sequence ๋ชจ๋ธ(S5)์„ ๊ฐœ์„ ํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๊ถค์  ๋‚ด์—์„œ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ์žฌ์„ค์ •ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ๊ฐ€๋ณ€ ๊ธธ์ด์˜ ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์—ฐ๊ด€ ์—ฐ์‚ฐ์ž๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  S5์—์„œ ์—ฐ๊ด€ ์†์„ฑ์„ ๋ณด์กดํ•˜๋Š” ์žฌ์„ค์ • ์ฃผ์„์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋„์ „์ ์ธ Meta-RL ์„ค์ •๋„ ์†Œ๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. [188]์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ”„๋กœํ† ํƒ€์ž…์„ ๋ถ„์„ํ•˜์—ฌ ๋””์ง€ํ„ธ ์˜ค๋””์˜ค ํ”„๋กœ๋•์…˜์„ ์œ„ํ•œ ์‚ฌ์‹ค์ ์ธ ๋””์ง€ํ„ธ ๋™์  ๋ฒ”์œ„ ์ปดํ”„๋ ˆ์„œ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ ํ‘œํ˜„์€ ์ข…์ข… ๋ชจ๋ธ์˜ ๋†’์€ ์ฐจ์ˆ˜์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์ œ์–ด ์„ค๊ณ„์— ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. [189]๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด SSM์˜ ์„ ํ˜• ๋™์  ๋ธ”๋ก์— ํŠนํ™”๋œ ์‹œ์Šคํ…œ ์ด๋ก  ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์ˆœ์„œ ๊ฐ์†Œ ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
์—ฌ๋Ÿฌ ์‹ค์ œ ์„ผ์„œ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ํ‰๊ท  ์ œ๊ณฑ ์˜ค์ฐจ(MSE) ์ธก๋ฉด์—์„œ ์ตœ์†Œ \(23\%\) ์ด์ƒ์˜ ๋‹ค๋ฅธ ์‹œํ€€์Šค ๋ชจ๋ธ์„ ํ˜•์„ฑํ•ฉ๋‹ˆ๋‹ค. LOCOST [168]๋Š” ๊ธด ์ปจํ…์ŠคํŠธ ์ž…๋ ฅ์ด ์žˆ๋Š” ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์„ ์œ„ํ•ด ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ธ์ฝ”๋”-๋””์ฝ”๋” ์•„ํ‚คํ…์ฒ˜๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๊ณ„์‚ฐ ๋ณต์žก์„ฑ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ ํšจ๊ณผ์ ์œผ๋กœ ์ค„์—ฌ ํ›ˆ๋ จ๊ณผ ์ถ”๋ก  ๋‹จ๊ณ„์˜ ์†๋„๋ฅผ ํฌ๊ฒŒ ํ–ฅ์ƒ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ๋ง˜๋ฐ”์Šคํ†ก [178]์€ ๊ตฌ์กฐํ™”๋œ ์ƒํƒœ ๊ณต๊ฐ„(S4) ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ฃผ์‹ ๋ฐ์ดํ„ฐ์˜ ๋น„์„ ํ˜•์„ฑ์„ ํฌ์ฐฉํ•˜์—ฌ ํ–ฅํ›„ ์ฃผ๊ฐ€๋ฅผ ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Lu et al. [163]์€ ๋‹จ์ˆœํ™”๋œ ๊ตฌ์กฐํ™”๋œ ์ƒํƒœ ๊ณต๊ฐ„ ์‹œํ€€์Šค ๋ชจ๋ธ(S5)์„ ๊ฐœ์„ ํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต ๋‹จ๊ณ„์—์„œ ๊ถค์  ๋‚ด์—์„œ ์ˆจ๊ฒจ์ง„ ์ƒํƒœ๋ฅผ ์žฌ์„ค์ •ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ œ์•ˆํ–ˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ๋ชจ๋ธ์ด ๊ฐ€๋ณ€ ๊ธธ์ด์˜ ์‹œํ€€์Šค๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๊ธฐ ์œ„ํ•ด ์ด ๋…ผ๋ฌธ์—์„œ๋Š” ์—ฐ๊ด€ ์—ฐ์‚ฐ์ž๋ฅผ ์ˆ˜์ •ํ•˜๊ณ  S5์—์„œ ์—ฐ๊ด€ ์†์„ฑ์„ ๋ณด์กดํ•˜๋Š” ์žฌ์„ค์ • ์ฃผ์„์„ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ๋Šฅ๋ ฅ์„ ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•ด ๋„์ „์ ์ธ ๋ฉ”ํƒ€ RL ์„ค์ •๋„ ์†Œ๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. [188]์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํ”„๋กœํ† ํƒ€์ž…์„ ๋ถ„์„ํ•˜์—ฌ ๋””์ง€ํ„ธ ์˜ค๋””์˜ค ํ”„๋กœ๋•์…˜์„ ์œ„ํ•œ ์‚ฌ์‹ค์ ์ธ ๋””์ง€ํ„ธ ๋™์  ๋ฒ”์œ„ ์ปดํ”„๋ ˆ์„œ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ธฐ ์œ„ํ•œ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต๋œ ํ‘œํ˜„์€ ์ข…์ข… ๋ชจ๋ธ์˜ ๋†’์€ ์ฐจ์ˆ˜์— ์˜ํ•ด ์˜ํ–ฅ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ์ œ์–ด ์„ค๊ณ„์— ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. [189]๋Š” ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด SSM์˜ ์„ ํ˜• ๋™์  ๋ธ”๋ก์— ํŠนํ™”๋œ ์‹œ์Šคํ…œ ์ด๋ก  ๊ธฐ๋ฐ˜ ๋ชจ๋ธ ์ˆœ์„œ ๊ฐ์†Œ ๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.
Ablation studies.We conduct a set of ablations of our framework to identify which factors play key roles. We evaluate two model variants trained differently than our model: _No Retriever_ trains an LM using the standard instruction-following method given instruction-output pairs, without retrieved passages; _No Critic_ trains an LM trained with input-output pairs that are always augmented with the top one retrieved document without reflection tokens. This is similar to SAIL (Luo et al., 2023), and we use our instruction-output data instead of using the Alpaca dataset (Dubois et al., 2023), as in SAIL. We also conduct ablation on our inference-time algorithm, including _No retrieval_ disables retrieval during inference; _Hard constraints_ indicates the model performance that retrieves when \(\boxed{Retrieve}\)=Yes instead of using the adaptive threshold; _Retrieve top 1_ always retrieves and uses the top one document only, similar to standard RAG approaches; _Remove_ \(\boxed{Isstr}\) indicates the model performance that removes \(\boxed{Isstr}\) score only during critique-guided beam search in Eq. 4. In this ablation experiment, we use a training instance size of 50k for a more efficient exploration of training variations. Later in this section, we conduct an analysis of the effect of training data size. We conduct the ablation studies on three datasets, PopQA, PubHealth, and ASQA. On ASQA, we evaluate models on sampled 150 instances and exclude ablations involving adaptive or no retrieval processes.
์ ˆ์ œ ์—ฐ๊ตฌ. ์–ด๋–ค ์š”์†Œ๊ฐ€ ํ•ต์‹ฌ ์—ญํ• ์„ ํ•˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋ ˆ์ž„์›Œํฌ์— ๋Œ€ํ•œ ์ผ๋ จ์˜ ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์ด ์—†๋Š” ํ‘œ์ค€ Instruction-following ๋ฐฉ์‹์œผ๋กœ Instruction-output ์Œ์ด ์ฃผ์–ด์ง€๋ฉด ๋ชจ๋ธ๊ณผ ๋‹ค๋ฅด๊ฒŒ ํ›ˆ๋ จ๋œ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ๋ณ€ํ˜•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค; _No Critic_์€ ๋ฐ˜์˜ ํ† ํฐ ์—†์ด ํ•ญ์ƒ ๊ฒ€์ƒ‰๋œ ์ƒ์œ„ ํ•œ ๋ฌธ์„œ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์œผ๋กœ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” SAIL(Luo et al., 2023)๊ณผ ์œ ์‚ฌํ•˜๋ฉฐ, SAIL์—์„œ์™€ ๊ฐ™์ด ์•ŒํŒŒ์นด ๋ฐ์ดํ„ฐ ์„ธํŠธ(Dubois et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  Instruction-output ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”๋ก  ์‹œ๊ฐ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ์ ˆ์ œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก  ์ค‘ ๊ฒ€์ƒ‰์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” _No retrieval_, ์ ์‘ํ˜• ์ž„๊ณ„๊ฐ’์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  \(\boxed{Retrieve}\)=Yes์ผ ๋•Œ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” _Hard constraints_, ํ‘œ์ค€ Retrieval-augmented generation (RAG) ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ํ•ญ์ƒ ์ƒ์œ„ ๋ฌธ์„œ ํ•˜๋‚˜๋งŒ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” _Retrieve top 1_, _Remove_ \(\boxed{Isstr}\)๋Š” ์‹ 4์˜ ๋น„ํ‰ ๊ฐ€์ด๋“œ ๋น” ๊ฒ€์ƒ‰ ์ค‘์—๋งŒ \(\boxed{Isstr}\) ์ ์ˆ˜๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์ ˆ์ œ ์‹คํ—˜์—์„œ๋Š” ํ›ˆ๋ จ ๋ณ€ํ˜•์„ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํƒ์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด 50k์˜ ํ›ˆ๋ จ ์ธ์Šคํ„ด์Šค ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜ ํ›„๋ฐ˜๋ถ€์—์„œ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ PopQA, PubHealth, ASQA์— ๋Œ€ํ•ด ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ASQA์—์„œ๋Š” 150๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ ์‘ํ˜• ๋˜๋Š” ๋น„๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค์™€ ๊ด€๋ จ๋œ ์ ˆ์ œ๋ฅผ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค.
์ ˆ์ œ ์—ฐ๊ตฌ. ์–ด๋–ค ์š”์†Œ๊ฐ€ ํ•ต์‹ฌ ์—ญํ• ์„ ํ•˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋ ˆ์ž„์›Œํฌ์— ๋Œ€ํ•œ ์ผ๋ จ์˜ ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ƒ‰๋œ ๊ตฌ์ ˆ์ด ์—†๋Š” ํ‘œ์ค€ ๋ช…๋ น์–ด-์ถ”์ข… ๋ฐฉ์‹์œผ๋กœ ๋ช…๋ น์–ด-์ถœ๋ ฅ ์Œ์ด ์ฃผ์–ด์ง€๋ฉด ๋ชจ๋ธ๊ณผ ๋‹ค๋ฅด๊ฒŒ ํ›ˆ๋ จ๋œ ๋‘ ๊ฐ€์ง€ ๋ชจ๋ธ ๋ณ€ํ˜•์„ ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค; _No Critic_์€ ๋ฐ˜์˜ ํ† ํฐ ์—†์ด ํ•ญ์ƒ ๊ฒ€์ƒ‰๋œ ์ƒ์œ„ ํ•œ ๋ฌธ์„œ๋กœ ์ฆ๊ฐ•๋œ ์ž…๋ ฅ-์ถœ๋ ฅ ์Œ์œผ๋กœ LM์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” SAIL(Luo et al., 2023)๊ณผ ์œ ์‚ฌํ•˜๋ฉฐ, SAIL์—์„œ์™€ ๊ฐ™์ด ์•ŒํŒŒ์นด ๋ฐ์ดํ„ฐ ์„ธํŠธ(Dubois et al., 2023)๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  ๋ช…๋ น์–ด-์ถœ๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ์ถ”๋ก  ์‹œ๊ฐ„ ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋Œ€ํ•œ ์ ˆ์ œ ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์—ฌ ์ถ”๋ก  ์ค‘ ๊ฒ€์ƒ‰์„ ๋น„ํ™œ์„ฑํ™”ํ•˜๋Š” _No ๊ฒ€์ƒ‰_, ์ ์‘ํ˜• ์ž„๊ณ„๊ฐ’์„ ์‚ฌ์šฉํ•˜๋Š” ๋Œ€์‹  \(\boxed{Retrieve}\)=Yes์ผ ๋•Œ ๊ฒ€์ƒ‰ํ•˜๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” _Hard constraints_, ํ‘œ์ค€ RAG ์ ‘๊ทผ ๋ฐฉ์‹๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ํ•ญ์ƒ ์ƒ์œ„ ๋ฌธ์„œ ํ•˜๋‚˜๋งŒ ๊ฒ€์ƒ‰ํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” _Retrieve top 1_, _Remove_ \(\boxed{Isstr}\)๋Š” ์‹ 4์˜ ๋น„ํ‰ ๊ฐ€์ด๋“œ ๋น” ๊ฒ€์ƒ‰ ์ค‘์—๋งŒ \(\boxed{Isstr}\) ์ ์ˆ˜๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์ด ์ ˆ์ œ ์‹คํ—˜์—์„œ๋Š” ํ›ˆ๋ จ ๋ณ€ํ˜•์„ ๋ณด๋‹ค ํšจ์œจ์ ์œผ๋กœ ํƒ์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด 50k์˜ ํ›ˆ๋ จ ์ธ์Šคํ„ด์Šค ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ์„น์…˜ ํ›„๋ฐ˜๋ถ€์—์„œ๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ ํฌ๊ธฐ์˜ ์˜ํ–ฅ์— ๋Œ€ํ•œ ๋ถ„์„์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์„ธ ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์ธ PopQA, PubHealth, ASQA์— ๋Œ€ํ•ด ์ ˆ์ œ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ASQA์—์„œ๋Š” 150๊ฐœ์˜ ์ธ์Šคํ„ด์Šค๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ณ  ์ ์‘ํ˜• ๋˜๋Š” ๋น„๊ฒ€์ƒ‰ ํ”„๋กœ์„ธ์Šค์™€ ๊ด€๋ จ๋œ ์ ˆ์ œ๋ฅผ ์ œ์™ธํ•ฉ๋‹ˆ๋‹ค.
* **GSM8K**[190] is designed to evaluate the model's ability for multi-step mathematical reasoning. GSM8K includes 8.5K linguistically diverse grade school math word problems written by humans. The dataset is split into two sets: a training set with \(7.5K\) problems, and a test set with \(1\)K problems. These problems need \(2\) to \(8\) steps to be solved. Solutions mainly are a series of elementary calculations using basic arithmetic operations. * **MATH**[191] enables to assess how well models can solve math problems. MATH dataset last \(12\), \(500\) problems from high school math competitions. Each problem in the dataset has a step-by-step solution and a final answer enclosed in a box. The problems cover a wide range of topics and have different levels of complexity. There are seven subjects in total. Furthermore, the difficulty of each problem is rated based on the AoPS standards on a scale from \({}^{\prime}1^{\prime}\) to \({}^{\prime}5^{\prime}\). A \({}^{\prime}1^{\prime}\) shows the easiest problems in a subject, while \({}^{\prime}5^{\prime}\) represents the most difficult. In terms of formatting, all problems and solutions are presented using LATEX and the Asymptote vector graphics language. * **HellaSwag**[192] is designed to assess commonsense reasoning in LLMs. This benchmark includes \(70,000\) multiple-choice questions. Each question is derived from one of two domains: ActivityNet or WikiHow, and presents four answer choices regarding what might happen in the following situation.
* **GSM8K**[190]์€ ๋ชจ๋ธ์˜ ๋‹ค๋‹จ๊ณ„ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. GSM8K์—๋Š” ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์–ธ์–ด์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ์ดˆ๋“ฑํ•™์ƒ ์ˆ˜ํ•™ ์›Œ๋“œ ๋ฌธ์ œ 8.5K๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋‘ ๊ฐœ์˜ ์„ธํŠธ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค: 7.5K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ training set์™€ 1K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ test set์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋“ค์€ ํ’€๊ธฐ ์œ„ํ•ด 2~8๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์†”๋ฃจ์…˜์€ ์ฃผ๋กœ ๊ธฐ๋ณธ ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์‚ฌ์šฉํ•˜๋Š” ์ผ๋ จ์˜ ์ดˆ๋“ฑ ๊ณ„์‚ฐ์ž…๋‹ˆ๋‹ค. * **MATH**[191]๋Š” ๋ชจ๋ธ์ด ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ํ’€ ์ˆ˜ ์žˆ๋Š”์ง€ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณ ๋“ฑํ•™๊ต ์ˆ˜ํ•™ ๊ฒฝ์‹œ๋Œ€ํšŒ์—์„œ ๋งˆ์ง€๋ง‰์œผ๋กœ 12,500 ๋ฌธ์ œ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ๋ฌธ์ œ์—๋Š” ๋‹จ๊ณ„๋ณ„ ํ’€์ด์™€ ์ตœ์ข… ๋‹ต์ด ๋ฐ•์Šค ์•ˆ์— ๋“ค์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ์ œ๋Š” ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ๋ฉฐ ๋ณต์žก๋„๊ฐ€ ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด 7๊ฐœ์˜ ๊ณผ๋ชฉ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ ๋ฌธ์ œ์˜ ๋‚œ์ด๋„๋Š” AoPS ํ‘œ์ค€์— ๋”ฐ๋ผ '1'์—์„œ '5'๊นŒ์ง€์˜ ์ฒ™๋„๋กœ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. '1'์€ ํ•œ ๊ณผ๋ชฉ์—์„œ ๊ฐ€์žฅ ์‰ฌ์šด ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, '5'๋Š” ๊ฐ€์žฅ ์–ด๋ ค์šด ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์„œ์‹ ๋ฉด์—์„œ ๋ชจ๋“  ๋ฌธ์ œ์™€ ํ•ด๊ฒฐ์ฑ…์€ LATEX์™€ Asymptote ๋ฒกํ„ฐ ๊ทธ๋ž˜ํ”ฝ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. * **HellaSwag**[192]๋Š” LLM์˜ ์ƒ์‹์  ์ถ”๋ก ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด benchmark์—๋Š” 70,000๊ฐœ์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์ œ๋Š” ActivityNet ๋˜๋Š” WikiHow ์ค‘ ํ•˜๋‚˜์˜ ๋„๋ฉ”์ธ์—์„œ ํŒŒ์ƒ๋˜๋ฉฐ, ๋‹ค์Œ ์ƒํ™ฉ์—์„œ ์ผ์–ด๋‚  ์ˆ˜ ์žˆ๋Š” ์ผ์— ๋Œ€ํ•ด ๋„ค ๊ฐ€์ง€ ๋‹ต์•ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
* **GSM8K**[190]์€ ๋ชจ๋ธ์˜ ๋‹ค๋‹จ๊ณ„ ์ˆ˜ํ•™์  ์ถ”๋ก  ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. GSM8K์—๋Š” ์ธ๊ฐ„์ด ์ž‘์„ฑํ•œ ์–ธ์–ด์ ์œผ๋กœ ๋‹ค์–‘ํ•œ ์ดˆ๋“ฑํ•™์ƒ ์ˆ˜ํ•™ ์›Œ๋“œ ๋ฌธ์ œ 8.5K๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋‘ ๊ฐœ์˜ ์„ธํŠธ๋กœ ๋‚˜๋‰ฉ๋‹ˆ๋‹ค: 7.5K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ ํ•™์Šต ์„ธํŠธ์™€ 1K๊ฐœ์˜ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋œ ํ…Œ์ŠคํŠธ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. ์ด ๋ฌธ์ œ๋“ค์€ ํ’€๊ธฐ ์œ„ํ•ด 2~8๋‹จ๊ณ„๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์†”๋ฃจ์…˜์€ ์ฃผ๋กœ ๊ธฐ๋ณธ ์‚ฐ์ˆ  ์—ฐ์‚ฐ์„ ์‚ฌ์šฉํ•˜๋Š” ์ผ๋ จ์˜ ์ดˆ๋“ฑ ๊ณ„์‚ฐ์ž…๋‹ˆ๋‹ค. * **MATH**[191]๋Š” ๋ชจ๋ธ์ด ์ˆ˜ํ•™ ๋ฌธ์ œ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ํ’€ ์ˆ˜ ์žˆ๋Š”์ง€ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. MATH ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๊ณ ๋“ฑํ•™๊ต ์ˆ˜ํ•™ ๊ฒฝ์‹œ๋Œ€ํšŒ์—์„œ ๋งˆ์ง€๋ง‰์œผ๋กœ \(12\), \(500\) ๋ฌธ์ œ๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๊ฐ ๋ฌธ์ œ์—๋Š” ๋‹จ๊ณ„๋ณ„ ํ’€์ด์™€ ์ตœ์ข… ๋‹ต์ด ๋ฐ•์Šค ์•ˆ์— ๋“ค์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฌธ์ œ๋Š” ๋‹ค์–‘ํ•œ ์ฃผ์ œ๋ฅผ ๋‹ค๋ฃจ๋ฉฐ ๋ณต์žก๋„๊ฐ€ ๋‹ค์–‘ํ•ฉ๋‹ˆ๋‹ค. ์ด 7๊ฐœ์˜ ๊ณผ๋ชฉ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ฐ ๋ฌธ์ œ์˜ ๋‚œ์ด๋„๋Š” AoPS ํ‘œ์ค€์— ๋”ฐ๋ผ \({}^{\prime}1^{\prime}\)์—์„œ \({}^{\prime}5^{\prime}\)๊นŒ์ง€์˜ ์ฒ™๋„๋กœ ํ‰๊ฐ€๋ฉ๋‹ˆ๋‹ค. 1^{\ํ”„๋ผ์ž„}\์€ ํ•œ ๊ณผ๋ชฉ์—์„œ ๊ฐ€์žฅ ์‰ฌ์šด ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ, \({}^{\ํ”„๋ผ์ž„}5^{\ํ”„๋ผ์ž„}\์€ ๊ฐ€์žฅ ์–ด๋ ค์šด ๋ฌธ์ œ๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ์„œ์‹ ๋ฉด์—์„œ ๋ชจ๋“  ๋ฌธ์ œ์™€ ํ•ด๊ฒฐ์ฑ…์€ LATEX์™€ Asymptote ๋ฒกํ„ฐ ๊ทธ๋ž˜ํ”ฝ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. * ํ—ฌ๋ผ์Šค์›จ๊ทธ**[192]๋Š” LLM์˜ ์ƒ์‹์  ์ถ”๋ก ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ณ ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฒค์น˜๋งˆํฌ์—๋Š” \(70,000\)๊ฐœ์˜ ๊ฐ๊ด€์‹ ๋ฌธ์ œ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๋ฌธ์ œ๋Š” ์•กํ‹ฐ๋น„ํ‹ฐ๋„ท ๋˜๋Š” ์œ„ํ‚คํ•˜์šฐ ์ค‘ ํ•˜๋‚˜์˜ ๋„๋ฉ”์ธ์—์„œ ํŒŒ์ƒ๋˜๋ฉฐ, ๋‹ค์Œ ์ƒํ™ฉ์—์„œ ์ผ์–ด๋‚  ์ˆ˜ ์žˆ๋Š” ์ผ์— ๋Œ€ํ•ด ๋„ค ๊ฐ€์ง€ ๋‹ต์•ˆ์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
The burgeoning interest in developing Large Language Models (LLMs) with up to trillion parameters has been met with concerns regarding resource efficiency and practical expense, particularly given the immense cost of experimentation. This scenario underscores the importance of exploring the potential of Small Language Models (SLMs) as a resource-efficient alternative. In this context, we introduce MiniCPM, specifically the 1.2B and 2.4B non-embedding parameter variants, not only excel in their respective categories but also demonstrate capabilities on par with 7B-13B LLMs. While focusing on SLMs, our approach exhibits scalability in both model and data dimensions for future LLM research. Regarding model scaling, we employ extensive model wind tunnel experiments for stable and optimal scaling. For data scaling, we introduce a Warmup-Stable-Decay (WSD) learning rate scheduler (LRS), conducive to continuous training and domain adaptation. We present an in-depth analysis of the intriguing training dynamics that occurred in the WSD LRS. With WSD LRS, we are now able to efficiently study data-model scaling law without extensive retraining experiments on both axes of model and data, from which we derive the much higher compute optimal data-model ratio than Chinchilla Optimal. Additionally, we introduce MiniCPM family, including MiniCPM-DPO, MiniCPM-MoE and MiniCPM-128K, whose excellent performance further cementing MiniCPM's foundation in diverse SLM applications. MiniCPM models are available publicly 1.
์ตœ๋Œ€ ์ˆ˜์กฐ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ Large Language Models (LLMs)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๋ฉด์„œ ๋ฆฌ์†Œ์Šค ํšจ์œจ์„ฑ๊ณผ ์‹ค์งˆ์ ์ธ ๋น„์šฉ, ํŠนํžˆ ์‹คํ—˜์— ๋“œ๋Š” ๋ง‰๋Œ€ํ•œ ๋น„์šฉ์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋Š” ๋ฆฌ์†Œ์Šค ํšจ์œจ์ ์ธ ๋Œ€์•ˆ์œผ๋กœ์„œ Small Language Models (SLMs)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ์šฐ๋ฆฌ๋Š” MiniCPM, ํŠนํžˆ 1.2B ๋ฐ 2.4B ๋น„์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ณ€ํ˜•์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ 7B-13B LLM๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. SLM์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ–ฅํ›„ LLM ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์ฐจ์› ๋ชจ๋‘์—์„œ ํ™•์žฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋ธ scaling๊ณผ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ์•ˆ์ •์ ์ด๊ณ  ์ตœ์ ์˜ scaling์„ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ scaling์˜ ๊ฒฝ์šฐ, ์ง€์†์ ์ธ ํ›ˆ๋ จ๊ณผ ๋„๋ฉ”์ธ ์ ์‘์— ๋„์›€์ด ๋˜๋Š” Warmup-Stable-Decay (WSD) Learning Rate Scheduler (LRS)๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. WSD LRS์—์„œ ๋ฐœ์ƒํ•œ ํฅ๋ฏธ๋กœ์šด ํ›ˆ๋ จ ๋™ํ•™์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. WSD LRS๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ์˜ ๋‘ ์ถ•์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์žฌ๊ต์œก ์‹คํ—˜ ์—†์ด๋„ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ scaling ๋ฒ•์น™์„ ํšจ์œจ์ ์œผ๋กœ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด Chinchilla Optimal๋ณด๋‹ค ํ›จ์”ฌ ๋†’์€ ๊ณ„์‚ฐ ์ตœ์  ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ๋น„์œจ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ MiniCPM-DPO, MiniCPM-MoE, MiniCPM-128K ๋“ฑ MiniCPM ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ์ด๋“ค์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์€ ๋‹ค์–‘ํ•œ SLM ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ MiniCPM์˜ ๊ธฐ๋ฐ˜์„ ๋”์šฑ ๊ณต๊ณ ํžˆ ํ•ฉ๋‹ˆ๋‹ค. MiniCPM ๋ชจ๋ธ์€ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ตœ๋Œ€ ์ˆ˜์กฐ ๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง„ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’์•„์ง€๋ฉด์„œ ๋ฆฌ์†Œ์Šค ํšจ์œจ์„ฑ๊ณผ ์‹ค์งˆ์ ์ธ ๋น„์šฉ, ํŠนํžˆ ์‹คํ—˜์— ๋“œ๋Š” ๋ง‰๋Œ€ํ•œ ๋น„์šฉ์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ œ๊ธฐ๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ๋‚˜๋ฆฌ์˜ค๋Š” ๋ฆฌ์†Œ์Šค ํšจ์œจ์ ์ธ ๋Œ€์•ˆ์œผ๋กœ์„œ ์†Œ๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(SLM)์˜ ์ž ์žฌ๋ ฅ์„ ํƒ๊ตฌํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋งฅ๋ฝ์—์„œ ์šฐ๋ฆฌ๋Š” ๋ฏธ๋‹ˆCPM, ํŠนํžˆ 1.2B ๋ฐ 2.4B ๋น„์ž„๋ฒ ๋”ฉ ํŒŒ๋ผ๋ฏธํ„ฐ ๋ณ€ํ˜•์„ ์†Œ๊ฐœํ•˜๋Š”๋ฐ, ์ด๋Š” ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ 7B-13B LLM๊ณผ ๋™๋“ฑํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. SLM์— ์ดˆ์ ์„ ๋งž์ถ˜ ์ ‘๊ทผ ๋ฐฉ์‹์€ ํ–ฅํ›„ LLM ์—ฐ๊ตฌ๋ฅผ ์œ„ํ•ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ ์ฐจ์› ๋ชจ๋‘์—์„œ ํ™•์žฅ์„ฑ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ๋ชจ๋ธ ์Šค์ผ€์ผ๋ง๊ณผ ๊ด€๋ จํ•˜์—ฌ, ์šฐ๋ฆฌ๋Š” ์•ˆ์ •์ ์ด๊ณ  ์ตœ์ ์˜ ์Šค์ผ€์ผ๋ง์„ ์œ„ํ•ด ๊ด‘๋ฒ”์œ„ํ•œ ๋ชจ๋ธ ์œˆ๋“œ ํ„ฐ๋„ ์‹คํ—˜์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์Šค์ผ€์ผ๋ง์˜ ๊ฒฝ์šฐ, ์ง€์†์ ์ธ ํ›ˆ๋ จ๊ณผ ๋„๋ฉ”์ธ ์ ์‘์— ๋„์›€์ด ๋˜๋Š” ์›Œ๋จธํ”„-์Šคํ…Œ์ด๋ธ”-๋””์ผ€์ด(WSD) ํ•™์Šต ์†๋„ ์Šค์ผ€์ค„๋Ÿฌ(LRS)๋ฅผ ๋„์ž…ํ–ˆ์Šต๋‹ˆ๋‹ค. WSD LRS์—์„œ ๋ฐœ์ƒํ•œ ํฅ๋ฏธ๋กœ์šด ํ›ˆ๋ จ ๋™ํ•™์— ๋Œ€ํ•œ ์‹ฌ์ธต ๋ถ„์„์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. WSD LRS๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ๊ณผ ๋ฐ์ดํ„ฐ์˜ ๋‘ ์ถ•์— ๋Œ€ํ•œ ๊ด‘๋ฒ”์œ„ํ•œ ์žฌ๊ต์œก ์‹คํ—˜ ์—†์ด๋„ ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ์Šค์ผ€์ผ๋ง ๋ฒ•์น™์„ ํšจ์œจ์ ์œผ๋กœ ์—ฐ๊ตฌํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์œผ๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ์นญ์น ๋ผ ์ตœ์ ๋ณด๋‹ค ํ›จ์”ฌ ๋†’์€ ๊ณ„์‚ฐ ์ตœ์  ๋ฐ์ดํ„ฐ-๋ชจ๋ธ ๋น„์œจ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฏธ๋‹ˆCPM-DPO, ๋ฏธ๋‹ˆCPM-MoE, ๋ฏธ๋‹ˆCPM-128K ๋“ฑ ๋ฏธ๋‹ˆCPM ์ œํ’ˆ๊ตฐ์„ ์†Œ๊ฐœํ•˜๋ฉฐ, ์ด๋“ค์˜ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์€ ๋‹ค์–‘ํ•œ SLM ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ๋ฏธ๋‹ˆCPM์˜ ๊ธฐ๋ฐ˜์„ ๋”์šฑ ๊ณต๊ณ ํžˆ ํ•ฉ๋‹ˆ๋‹ค. ๋ฏธ๋‹ˆCPM ๋ชจ๋ธ์€ ๊ณต๊ฐœ์ ์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
strategy that converts a point cloud to a 1-D sequence of points and ensures that adjacent points in the sequence are also adjacent in space. Consistent traverse serialization strategy produces six variants by arranging the order of x, y, and z coordinates, and the author introduces order prompt to inform Mampa of the arrangement rules of the sequence. In addition, a positional embedding method based on spatial coordinate mapping is proposed to add point cloud position information. Point Mamba [155] designs an octree-based ordering mechanism for irregular points,ensuring the preservation of their spatial proximity and causal dependence. The points undergo a sequence of Point Mamba blocks and downsampling layers to extract layered point characteristics. Zhou et al. propose 3DMambaIPF [156], which integrates Mamba into the point cloud filtering task and introduces a fast differentiable rendering loss. This approach has demonstrated strong performance in handling large-scale point clouds. Li et al. propose 3DMambaComplete [157], which incorporates the HyperPoint Generation module, HyperPoint Spread module, and deformation method for point cloud reconstruction. The HyperPoint Generation module introduces Mamba's selection mechanism to encode point cloud features. Zubi'c et al. [158] introduce a state-space models (SSMs) with learnable time-scale parameters to the event-based vision, enabling adaptation to different frequency time inputs without having to retrain the network at different frequencies.
ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ๋ฅผ 1-D์˜ ์  ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ์‹œํ€€์Šค์˜ ์ธ์ ‘ํ•œ ์ ์ด ๊ณต๊ฐ„์ ์œผ๋กœ๋„ ์ธ์ ‘ํ•˜๋„๋ก ํ•˜๋Š” ์ „๋žต์ž…๋‹ˆ๋‹ค. ์ผ๊ด€๋œ traverse serialization ์ „๋žต์€ x, y, z ์ขŒํ‘œ์˜ ์ˆœ์„œ๋ฅผ ์ •๋ ฌํ•˜์—ฌ 6๊ฐ€์ง€ ๋ณ€ํ˜•์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ž‘์„ฑ์ž๋Š” order prompt๋ฅผ ๋„์ž…ํ•˜์—ฌ Mampa์— ์‹œํ€€์Šค์˜ ๋ฐฐ์—ด ๊ทœ์น™์„ ์•Œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณต๊ฐ„ ์ขŒํ‘œ ๋งคํ•‘ ๊ธฐ๋ฐ˜์˜ positional embedding ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์—ฌ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ์œ„์น˜ ์ •๋ณด๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Point Mamba [155]๋Š” ๋ถˆ๊ทœ์น™ํ•œ ํฌ์ธํŠธ์— ๋Œ€ํ•ด octree ๊ธฐ๋ฐ˜์˜ ordering ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ•˜์—ฌ ๊ณต๊ฐ„์  ๊ทผ์ ‘์„ฑ๊ณผ ์ธ๊ณผ์  ์ข…์†์„ฑ์„ ๋ณด์กดํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ํฌ์ธํŠธ๋“ค์€ ์ผ๋ จ์˜ Point Mamba ๋ธ”๋ก๊ณผ downsampling ๋ ˆ์ด์–ด๋ฅผ ๊ฑฐ์ณ ๊ณ„์ธต์ ์ธ ํฌ์ธํŠธ ํŠน์„ฑ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. Zhou ๋“ฑ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ํ•„ํ„ฐ๋ง ์ž‘์—…์— Mamba๋ฅผ ํ†ตํ•ฉํ•˜๊ณ  ๋น ๋ฅด๊ฒŒ ์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ rendering loss๋ฅผ ๋„์ž…ํ•œ 3DMambaIPF [156]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๊ทœ๋ชจ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. Li ๋“ฑ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•œ HyperPoint Generation ๋ชจ๋“ˆ, HyperPoint Spread ๋ชจ๋“ˆ ๋ฐ ๋ณ€ํ˜• ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•œ 3DMambaComplete [157]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. HyperPoint Generation ๋ชจ๋“ˆ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ํŠน์ง•์„ ์ธ์ฝ”๋”ฉํ•˜๊ธฐ ์œ„ํ•ด Mamba์˜ ์„ ํƒ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Zubi'c ๋“ฑ [158]์€ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์‹œ๊ฐ„ ์Šค์ผ€์ผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–์ถ˜ state-space models (SSMs)์„ ์ด๋ฒคํŠธ ๊ธฐ๋ฐ˜ ๋น„์ „์— ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ฃผํŒŒ์ˆ˜์—์„œ ๋„คํŠธ์›Œํฌ๋ฅผ ๋‹ค์‹œ ํ›ˆ๋ จํ•  ํ•„์š” ์—†์ด ๋‹ค์–‘ํ•œ ์ฃผํŒŒ์ˆ˜ ์‹œ๊ฐ„ ์ž…๋ ฅ์— ์ ์‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ๋ฅผ 1-D์˜ ์  ์‹œํ€€์Šค๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ์‹œํ€€์Šค์˜ ์ธ์ ‘ํ•œ ์ ์ด ๊ณต๊ฐ„์ ์œผ๋กœ๋„ ์ธ์ ‘ํ•˜๋„๋ก ํ•˜๋Š” ์ „๋žต์ž…๋‹ˆ๋‹ค. ์ผ๊ด€๋œ ํŠธ๋ž˜๋ฒ„์Šค ์ง๋ ฌํ™” ์ „๋žต์€ x, y, z ์ขŒํ‘œ์˜ ์ˆœ์„œ๋ฅผ ์ •๋ ฌํ•˜์—ฌ 6๊ฐ€์ง€ ๋ณ€ํ˜•์„ ์ƒ์„ฑํ•˜๋ฉฐ, ์ž‘์„ฑ์ž๋Š” ์ˆœ์„œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋„์ž…ํ•˜์—ฌ Mampa์— ์‹œํ€€์Šค์˜ ๋ฐฐ์—ด ๊ทœ์น™์„ ์•Œ๋ฆฝ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ณต๊ฐ„ ์ขŒํ‘œ ๋งคํ•‘ ๊ธฐ๋ฐ˜์˜ ์œ„์น˜ ์ž„๋ฒ ๋”ฉ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์—ฌ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ์œ„์น˜ ์ •๋ณด๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. Point Mamba [155]๋Š” ๋ถˆ๊ทœ์น™ํ•œ ํฌ์ธํŠธ์— ๋Œ€ํ•ด ์˜ฅํŠธ๋ฆฌ ๊ธฐ๋ฐ˜์˜ ์ˆœ์„œ ์ง€์ • ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์„ค๊ณ„ํ•˜์—ฌ ๊ณต๊ฐ„์  ๊ทผ์ ‘์„ฑ๊ณผ ์ธ๊ณผ์  ์ข…์†์„ฑ์„ ๋ณด์กดํ•˜๋„๋ก ํ•ฉ๋‹ˆ๋‹ค. ํฌ์ธํŠธ๋“ค์€ ์ผ๋ จ์˜ ํฌ์ธํŠธ ๋ง˜๋ฐ” ๋ธ”๋ก๊ณผ ๋‹ค์šด์ƒ˜ํ”Œ๋ง ๋ ˆ์ด์–ด๋ฅผ ๊ฑฐ์ณ ๊ณ„์ธต์ ์ธ ํฌ์ธํŠธ ํŠน์„ฑ์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค. Zhou ๋“ฑ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ํ•„ํ„ฐ๋ง ์ž‘์—…์— Mamba๋ฅผ ํ†ตํ•ฉํ•˜๊ณ  ๋น ๋ฅด๊ฒŒ ์ฐจ๋ณ„ํ™” ๊ฐ€๋Šฅํ•œ ๋ Œ๋”๋ง ์†์‹ค์„ ๋„์ž…ํ•œ 3DMambaIPF [156]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด ์ ‘๊ทผ ๋ฐฉ์‹์€ ๋Œ€๊ทœ๋ชจ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” ๋ฐ ๊ฐ•๋ ฅํ•œ ์„ฑ๋Šฅ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค. Li ๋“ฑ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ์žฌ๊ตฌ์„ฑ์„ ์œ„ํ•œ ํ•˜์ดํผํฌ์ธํŠธ ์ƒ์„ฑ ๋ชจ๋“ˆ, ํ•˜์ดํผํฌ์ธํŠธ ์Šคํ”„๋ ˆ๋“œ ๋ชจ๋“ˆ ๋ฐ ๋ณ€ํ˜• ๋ฐฉ๋ฒ•์„ ํ†ตํ•ฉํ•œ 3DMambaComplete [157]๋ฅผ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ดํผํฌ์ธํŠธ ์ƒ์„ฑ ๋ชจ๋“ˆ์€ ํฌ์ธํŠธ ํด๋ผ์šฐ๋“œ ํŠน์ง•์„ ์ธ์ฝ”๋”ฉํ•˜๊ธฐ ์œ„ํ•ด Mamba์˜ ์„ ํƒ ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๋„์ž…ํ•ฉ๋‹ˆ๋‹ค. Zubi'c ๋“ฑ [158]์€ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ์‹œ๊ฐ„ ์Šค์ผ€์ผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ–์ถ˜ ์ƒํƒœ ๊ณต๊ฐ„ ๋ชจ๋ธ(SSM)์„ ์ด๋ฒคํŠธ ๊ธฐ๋ฐ˜ ๋น„์ „์— ๋„์ž…ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ฃผํŒŒ์ˆ˜์—์„œ ๋„คํŠธ์›Œํฌ๋ฅผ ๋‹ค์‹œ ํ›ˆ๋ จํ•  ํ•„์š” ์—†์ด ๋‹ค์–‘ํ•œ ์ฃผํŒŒ์ˆ˜ ์‹œ๊ฐ„ ์ž…๋ ฅ์— ์ ์‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
**Conditional Text Generation.** As an important topic in language generation, conditional text generation [48] focuses on generating texts satisfying specific task demands based on the given conditions, typically including machine translation [624], text summarization [548], and question answering [557]. To measure the quality of the generated text, automatic metrics (_e.g._, Accuracy, BLEU [625] and ROUGE [626]) and human ratings have been typically used for evaluating the performance. Due to the powerful language generation capabilities, LLMs have achieved remarkable performance on existing datasets and benchmarks. For instance, GPT-4 exhibits comparable performance as commercial translation products, even for the translation task of languages that are with significant linguistic distance [627]. On news summarization tasks (_i.e._, CNN/DM and XSUM), LLMs also demonstrate comparable performance with human freelance writers [628]. Despite the rapid progress on model capacity, there are increasing concerns on the feasibility of existing automatic metrics to faithfully assess the performance of LLMs in conditional text generation tasks [628, 629, 630]. As the alternatives to automatic metrics, recent studies also propose to incorporate LLMs as generation evaluators to examine the quality of the generated content [631, 632, 138]. Moreover, researchers also explore more challenging language generation tasks for LLMs, such as structured data generation [458] and long text generation [633, 46, 46].
**Conditional Text Generation**. ์–ธ์–ด ์ƒ์„ฑ์˜ ์ค‘์š”ํ•œ ์ฃผ์ œ๋กœ์„œ conditional text generation[48]์€ ์ฃผ๋กœ machine translation[624], text summarization[548], question answering[557] ๋“ฑ ์ฃผ์–ด์ง„ ์กฐ๊ฑด์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํŠน์ • ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์˜ ํ’ˆ์งˆ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ ๋ฉ”ํŠธ๋ฆญ(์˜ˆ: Accuracy, BLEU[625] ๋ฐ ROUGE[626])๊ณผ ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์œผ๋กœ ์ธํ•ด LLMs์€ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4๋Š” ์–ธ์–ด์  ๊ฑฐ๋ฆฌ๊ฐ€ ํฐ ์–ธ์–ด์˜ ๋ฒˆ์—ญ ์ž‘์—…์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์šฉ ๋ฒˆ์—ญ ์ œํ’ˆ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[627]. ๋‰ด์Šค ์š”์•ฝ ์ž‘์—…(์˜ˆ: CNN/DM ๋ฐ XSUM)์—์„œ๋„ LLMs์€ ํ”„๋ฆฌ๋žœ์„œ ์ž‘๊ฐ€์™€ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[628]. ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๊ธ‰๊ฒฉํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  conditional text generation ์ž‘์—…์—์„œ LLMs์˜ ์„ฑ๋Šฅ์„ ์ถฉ์‹คํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์กด ์ž๋™ ๋ฉ”ํŠธ๋ฆญ์˜ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[628, 629, 630]. ์ž๋™ ๋ฉ”ํŠธ๋ฆญ์˜ ๋Œ€์•ˆ์œผ๋กœ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด LLMs์„ ์ƒ์„ฑ ํ‰๊ฐ€์ž๋กœ ํ†ตํ•ฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์ œ์•ˆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[631, 632, 138]. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ structured data generation[458] ๋ฐ long text generation[633, 46, 46]๊ณผ ๊ฐ™์€ ๋”์šฑ ๋„์ „์ ์ธ ์–ธ์–ด ์ƒ์„ฑ ์ž‘์—…์„ LLMs์— ๋Œ€ํ•ด ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
**์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ**. ์–ธ์–ด ์ƒ์„ฑ์˜ ์ค‘์š”ํ•œ ์ฃผ์ œ๋กœ์„œ ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ[48]์€ ์ฃผ๋กœ ๊ธฐ๊ณ„ ๋ฒˆ์—ญ[624], ํ…์ŠคํŠธ ์š”์•ฝ[548], ์งˆ๋ฌธ ๋‹ต๋ณ€[557] ๋“ฑ ์ฃผ์–ด์ง„ ์กฐ๊ฑด์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํŠน์ • ์ž‘์—… ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ถฉ์กฑํ•˜๋Š” ํ…์ŠคํŠธ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘ก๋‹ˆ๋‹ค. ์ƒ์„ฑ๋œ ํ…์ŠคํŠธ์˜ ํ’ˆ์งˆ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•ด ์ž๋™ ๋ฉ”ํŠธ๋ฆญ(์˜ˆ: ์ •ํ™•๋„, BLEU [625] ๋ฐ ROUGE [626])๊ณผ ์‚ฌ๋žŒ์˜ ํ‰๊ฐ€๊ฐ€ ์ผ๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ์ƒ์„ฑ ๊ธฐ๋Šฅ์œผ๋กœ ์ธํ•ด LLM์€ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ์™€ ๋ฒค์น˜๋งˆํฌ์—์„œ ๋†€๋ผ์šด ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ–ˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT-4๋Š” ์–ธ์–ด์  ๊ฑฐ๋ฆฌ๊ฐ€ ํฐ ์–ธ์–ด์˜ ๋ฒˆ์—ญ ์ž‘์—…์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์ƒ์šฉ ๋ฒˆ์—ญ ์ œํ’ˆ๊ณผ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[627]. ๋‰ด์Šค ์š”์•ฝ ์ž‘์—…(์˜ˆ: CNN/DM ๋ฐ XSUM)์—์„œ๋„ LLM์€ ํ”„๋ฆฌ๋žœ์„œ ์ž‘๊ฐ€์™€ ๋น„์Šทํ•œ ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค[628]. ๋ชจ๋ธ ์šฉ๋Ÿ‰์˜ ๊ธ‰๊ฒฉํ•œ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์กฐ๊ฑด๋ถ€ ํ…์ŠคํŠธ ์ƒ์„ฑ ์ž‘์—…์—์„œ LLM์˜ ์„ฑ๋Šฅ์„ ์ถฉ์‹คํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๊ธฐ์กด ์ž๋™ ๋ฉ”ํŠธ๋ฆญ์˜ ์‹คํ˜„ ๊ฐ€๋Šฅ์„ฑ์— ๋Œ€ํ•œ ์šฐ๋ ค๊ฐ€ ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[628, 629, 630]. ์ž๋™ ๋ฉ”ํŠธ๋ฆญ์˜ ๋Œ€์•ˆ์œผ๋กœ ์ตœ๊ทผ ์—ฐ๊ตฌ์—์„œ๋Š” ์ƒ์„ฑ๋œ ์ฝ˜ํ…์ธ ์˜ ํ’ˆ์งˆ์„ ๊ฒ€ํ† ํ•˜๊ธฐ ์œ„ํ•ด LLM์„ ์ƒ์„ฑ ํ‰๊ฐ€์ž๋กœ ํ†ตํ•ฉํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์ œ์•ˆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค[631, 632, 138]. ๋˜ํ•œ ์—ฐ๊ตฌ์ž๋“ค์€ ๊ตฌ์กฐํ™”๋œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ[458] ๋ฐ ๊ธด ํ…์ŠคํŠธ ์ƒ์„ฑ[633, 46, 46]๊ณผ ๊ฐ™์€ ๋”์šฑ ๋„์ „์ ์ธ ์–ธ์–ด ์ƒ์„ฑ ์ž‘์—…์„ LLM์— ๋Œ€ํ•ด ํƒ๊ตฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Components._ Typically, there are three main components in an LLM-based agent: _memory, planning50_, and _execution_. Specifically, the _memory_ component aims to store the information perceived from the environment and can be utilized to support decision-making. In particular, LLM-based agents usually maintain information in both short-term memory and long-term memory with the operations of reading and writing. Short-term memory usually refers to the internal context window of LLMs (_i.e._, input), where LLMs can read and write through actions like reasoning [880]. While long-term memory can be mapped to the external storage like vector databases [537], where LLMs can read through retrieval and write with reflection [686]. Specially, profiles are usually implemented with long-term memory, which is an important feature for an agent that specifies its role and function [818]. The _planning_ component is responsible for generating the action plan based on the information from the memory component. In data format, the plan usually takes the form of text-based instructions [441] or code-based programs [443]. To generate it, LLM-based agents will first propose several candidates and then select a more suitable one among them [436]. The initial plan can be further refined with execution feedback from the environment [528]. The _execution_ component is in charge of carrying out the plan from the planning component, which can be fulfilled by the internal LLM [441] or external tools [880].
\(\bullet\)_Components._ ์ผ๋ฐ˜์ ์œผ๋กœ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์—๋Š” ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค: _memory, planning50_, _execution_. ํŠนํžˆ, _memory_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ํ™˜๊ฒฝ์—์„œ ์ธ์‹ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ ์˜์‚ฌ ๊ฒฐ์ •์„ ์ง€์›ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ฝ๊ธฐ ๋ฐ ์“ฐ๊ธฐ ์ž‘์—…์„ ํ†ตํ•ด ๋‹จ๊ธฐ ๊ธฐ์–ต๊ณผ ์žฅ๊ธฐ ๊ธฐ์–ต ๋ชจ๋‘์— ์ •๋ณด๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋‹จ๊ธฐ ๊ธฐ์–ต์€ ์ผ๋ฐ˜์ ์œผ๋กœ LLM์˜ ๋‚ด๋ถ€ ์ปจํ…์ŠคํŠธ ์ฐฝ(_์ฆ‰, ์ž…๋ ฅ)์„ ์˜๋ฏธํ•˜๋ฉฐ, LLM์€ ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๋™์ž‘์„ ํ†ตํ•ด ์ฝ๊ณ  ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[880]. ์žฅ๊ธฐ ๊ธฐ์–ต์€ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์™€ ๊ฐ™์€ ์™ธ๋ถ€ ์ €์žฅ์†Œ์— ๋งคํ•‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[537], LLM์€ retrieval์„ ํ†ตํ•ด ์ฝ๊ณ  reflection์„ ํ†ตํ•ด ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[686]. ํŠนํžˆ ํ”„๋กœํ•„์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์žฅ๊ธฐ ๊ธฐ์–ต์œผ๋กœ ๊ตฌํ˜„๋˜๋ฉฐ, ์ด๋Š” ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ์ง€์ •ํ•˜๋Š” ์—์ด์ „ํŠธ์˜ ์ค‘์š”ํ•œ ๊ธฐ๋Šฅ์ž…๋‹ˆ๋‹ค[818]. _planning_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” memory ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹คํ–‰ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ˜•์‹์—์„œ ๊ณ„ํš์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ instruction[441] ๋˜๋Š” ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ํ”„๋กœ๊ทธ๋žจ[443]์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์—ฌ๋Ÿฌ ํ›„๋ณด๋ฅผ ์ œ์•ˆํ•œ ๋‹ค์Œ ๊ทธ ์ค‘์—์„œ ๋” ์ ํ•ฉํ•œ ํ›„๋ณด๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค[436]. ์ดˆ๊ธฐ ๊ณ„ํš์€ ํ™˜๊ฒฝ์˜ ์‹คํ–‰ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ๋”์šฑ ๊ฐœ์„ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[528]. _execution_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” planning ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ณ„ํš์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์ž‘์—…์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ๋‚ด๋ถ€ LLM[441] ๋˜๋Š” ์™ธ๋ถ€ ๋„๊ตฌ[880]๋ฅผ ํ†ตํ•ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
\(\bullet\)_Components._ ์ผ๋ฐ˜์ ์œผ๋กœ LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ์—๋Š” ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค: _๋ฉ”๋ชจ๋ฆฌ, ๊ณ„ํš50_, _์‹คํ–‰_. ํŠนํžˆ, _๋ฉ”๋ชจ๋ฆฌ_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ํ™˜๊ฒฝ์—์„œ ์ธ์‹ํ•œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ ์˜์‚ฌ ๊ฒฐ์ •์„ ์ง€์›ํ•˜๋Š” ๋ฐ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ฝ๊ธฐ ๋ฐ ์“ฐ๊ธฐ ์ž‘์—…์„ ํ†ตํ•ด ๋‹จ๊ธฐ ๊ธฐ์–ต๊ณผ ์žฅ๊ธฐ ๊ธฐ์–ต ๋ชจ๋‘์— ์ •๋ณด๋ฅผ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ๋‹จ๊ธฐ ๊ธฐ์–ต์€ ์ผ๋ฐ˜์ ์œผ๋กœ LLM์˜ ๋‚ด๋ถ€ ์ปจํ…์ŠคํŠธ ์ฐฝ(_์ฆ‰, ์ž…๋ ฅ)์„ ์˜๋ฏธํ•˜๋ฉฐ, LLM์€ ์ถ”๋ก ๊ณผ ๊ฐ™์€ ๋™์ž‘์„ ํ†ตํ•ด ์ฝ๊ณ  ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[880]. ์žฅ๊ธฐ ๊ธฐ์–ต์€ ๋ฒกํ„ฐ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์™€ ๊ฐ™์€ ์™ธ๋ถ€ ์ €์žฅ์†Œ์— ๋งคํ•‘ํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ[537], LLM์€ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ์ฝ๊ณ  ๋ฐ˜์˜์„ ํ†ตํ•ด ์“ธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[686]. ํŠนํžˆ ํ”„๋กœํ•„์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์žฅ๊ธฐ ๊ธฐ์–ต์œผ๋กœ ๊ตฌํ˜„๋˜๋ฉฐ, ์ด๋Š” ์—ญํ• ๊ณผ ๊ธฐ๋Šฅ์„ ์ง€์ •ํ•˜๋Š” ์—์ด์ „ํŠธ์˜ ์ค‘์š”ํ•œ ๊ธฐ๋Šฅ์ž…๋‹ˆ๋‹ค[818]. ๊ณ„ํš_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ๊ตฌ์„ฑ ์š”์†Œ์˜ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹คํ–‰ ๊ณ„ํš์„ ์ƒ์„ฑํ•˜๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ํ˜•์‹์—์„œ ๊ณ„ํš์€ ์ผ๋ฐ˜์ ์œผ๋กœ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ์ง€์นจ[441] ๋˜๋Š” ์ฝ”๋“œ ๊ธฐ๋ฐ˜ ํ”„๋กœ๊ทธ๋žจ[443]์˜ ํ˜•ํƒœ๋ฅผ ์ทจํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด LLM ๊ธฐ๋ฐ˜ ์—์ด์ „ํŠธ๋Š” ๋จผ์ € ์—ฌ๋Ÿฌ ํ›„๋ณด๋ฅผ ์ œ์•ˆํ•œ ๋‹ค์Œ ๊ทธ ์ค‘์—์„œ ๋” ์ ํ•ฉํ•œ ํ›„๋ณด๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค[436]. ์ดˆ๊ธฐ ๊ณ„ํš์€ ํ™˜๊ฒฝ์˜ ์‹คํ–‰ ํ”ผ๋“œ๋ฐฑ์„ ํ†ตํ•ด ๋”์šฑ ๊ฐœ์„ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค[528]. ์‹คํ–‰_ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๊ณ„ํš ๊ตฌ์„ฑ ์š”์†Œ์˜ ๊ณ„ํš์„ ์ˆ˜ํ–‰ํ•˜๋Š” ์ž‘์—…์„ ๋‹ด๋‹นํ•˜๋ฉฐ, ๋‚ด๋ถ€ LLM[441] ๋˜๋Š” ์™ธ๋ถ€ ๋„๊ตฌ[880]๋ฅผ ํ†ตํ•ด ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
Math.Recent advancements in mathematical reasoning for language models (LMs) have led to the development of various approaches to improve performance on math-related tasks. One approach is to train on math-targeted pre-training data, such as Minerva (Lewkowycz et al., 2022), Llemina (Azerbayev et al., 2023), and DeepSeeKMath (Shao et al., 2024). Another mainstream method is to generate synthetic questions and answers to imitate the training or validation set of target benchmarks. For instance, WizardMath (Luo et al., 2023) leverages a series of operations to increase the complexity of questions and answers using GPT-3.5, while MetaMath (Yu et al., 2023) bootstraps the questions in MATH and GSM8K by rewriting them in different ways, such as semantic rephrasing, self-verification, and backward reasoning. GAIR-Abel (Chern et al., 2023) found that the format of the augmented answers is crucial to final performance, with answers that begin with a paraphrasing of the question followed by a step-by-step solution showing better performance than those in vanilla format. Xwin-Math (Li et al., 2024) further scaled up synthetic SFT data to one million examples and found that the LLaMA-2 7B model (Touvron et al., 2023) can still benefit from data scaling. MMIQC (Liu and Yao, 2024) composed a bundle of datasets that infuse SFT style data (via question-answer rephrasing or directly taken from MetaMath) with a subset of high-quality mathematical pre-training data, such as OpenWebMath (Paster et al., 2023).
์ตœ๊ทผ ์–ธ์–ด ๋ชจ๋ธ(LM)์˜ ์ˆ˜ํ•™์  ์ถ”๋ก ์— ๋Œ€ํ•œ ๋ฐœ์ „์œผ๋กœ ์ˆ˜ํ•™ ๊ด€๋ จ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์— ๋งž๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(์˜ˆ: Minerva(Lewkowycz et al., 2022), Llemina(Azerbayev et al., 2023), DeepSeeKMath(Shao et al., 2024))๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ฃผ๋ฅ˜ ๋ฐฉ๋ฒ•์€ ๋Œ€์ƒ ๋ฒค์น˜๋งˆํฌ์˜ ํ•™์Šต ๋˜๋Š” ๊ฒ€์ฆ ์„ธํŠธ๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WizardMath(Luo et al., 2023)๋Š” ์ผ๋ จ์˜ ์—ฐ์‚ฐ์„ ํ™œ์šฉํ•˜์—ฌ GPT-3.5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ณ , MetaMath(Yu et al., 2023)๋Š” ์˜๋ฏธ๋ก ์  ์žฌ๊ตฌ๋ฌธํ™”, ์ž๊ธฐ ๊ฒ€์ฆ, ์—ญ ์ถ”๋ก  ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ MATH ๋ฐ GSM8K์˜ ์งˆ๋ฌธ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•ฉ๋‹ˆ๋‹ค. GAIR-Abel(Chern et al., 2023)์€ ์ฆ๊ฐ•๋œ ๋‹ต๋ณ€์˜ ํ˜•์‹์ด ์ตœ์ข… ์„ฑ๋Šฅ์— ๊ฒฐ์ •์ ์ด๋ฉฐ, ์งˆ๋ฌธ์„ ์˜์—ญํ•œ ๋‹ค์Œ ๋‹จ๊ณ„๋ณ„ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ ์‹œ์ž‘ํ•˜๋Š” ๋‹ต๋ณ€์ด ์ผ๋ฐ˜ ํ˜•์‹์˜ ๋‹ต๋ณ€๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Xwin-Math(Li et al., 2024)๋Š” ํ•ฉ์„ฑ Supervised Fine-tuning (SFT) ๋ฐ์ดํ„ฐ๋ฅผ 100๋งŒ ๊ฐœ์˜ ์˜ˆ์ œ๋กœ ํ™•์žฅํ•˜์—ฌ LLaMA-2 7B ๋ชจ๋ธ(Touvron et al., 2023)์ด ๋ฐ์ดํ„ฐ ํ™•์žฅ์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. MMIQC(Liu and Yao, 2024)๋Š” ์งˆ๋ฌธ-๋‹ต๋ณ€์˜ ๋ฌธ๊ตฌ๋ฅผ ๋ฐ”๊พธ๊ฑฐ๋‚˜ MetaMath์—์„œ ์ง์ ‘ ๊ฐ€์ ธ์˜จ SFT ์Šคํƒ€์ผ์˜ ๋ฐ์ดํ„ฐ๋ฅผ OpenWebMath(Paster et al., 2023)์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ์ˆ˜ํ•™ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ํ˜ผํ•ฉํ•œ ๋ฐ์ดํ„ฐ ๋ฌถ์Œ์„ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
์ตœ๊ทผ ์–ธ์–ด ๋ชจ๋ธ(LM)์˜ ์ˆ˜ํ•™์  ์ถ”๋ก ์— ๋Œ€ํ•œ ๋ฐœ์ „์œผ๋กœ ์ˆ˜ํ•™ ๊ด€๋ จ ์ž‘์—…์˜ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์ด ๊ฐœ๋ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ์ ‘๊ทผ ๋ฐฉ์‹์€ ์ˆ˜ํ•™์— ๋งž๋Š” ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ(์˜ˆ: Minerva(Lewkowycz et al., 2022), Llemina(Azerbayev et al., 2023), DeepSeeKMath(Shao et al., 2024))๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋˜ ๋‹ค๋ฅธ ์ฃผ๋ฅ˜ ๋ฐฉ๋ฒ•์€ ๋Œ€์ƒ ๋ฒค์น˜๋งˆํฌ์˜ ํ•™์Šต ๋˜๋Š” ๊ฒ€์ฆ ์„ธํŠธ๋ฅผ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด ํ•ฉ์„ฑ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์„ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, WizardMath(Luo et al., 2023)๋Š” ์ผ๋ จ์˜ ์—ฐ์‚ฐ์„ ํ™œ์šฉํ•˜์—ฌ GPT-3.5๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์งˆ๋ฌธ๊ณผ ๋‹ต๋ณ€์˜ ๋ณต์žก์„ฑ์„ ๋†’์ด๊ณ , MetaMath(Yu et al., 2023)๋Š” ์˜๋ฏธ๋ก ์  ์žฌ๊ตฌ๋ฌธํ™”, ์ž๊ธฐ ๊ฒ€์ฆ, ์—ญ ์ถ”๋ก  ๋“ฑ ๋‹ค์–‘ํ•œ ๋ฐฉ์‹์œผ๋กœ MATH ๋ฐ GSM8K์˜ ์งˆ๋ฌธ์„ ๋ถ€ํŠธ์ŠคํŠธ๋žฉํ•ฉ๋‹ˆ๋‹ค. GAIR-Abel(Chern et al., 2023)์€ ์ฆ๊ฐ•๋œ ๋‹ต๋ณ€์˜ ํ˜•์‹์ด ์ตœ์ข… ์„ฑ๋Šฅ์— ๊ฒฐ์ •์ ์ด๋ฉฐ, ์งˆ๋ฌธ์„ ์˜์—ญํ•œ ๋‹ค์Œ ๋‹จ๊ณ„๋ณ„ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ ์‹œ์ž‘ํ•˜๋Š” ๋‹ต๋ณ€์ด ์ผ๋ฐ˜ ํ˜•์‹์˜ ๋‹ต๋ณ€๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค. Xwin-Math(Li et al., 2024)๋Š” ํ•ฉ์„ฑ SFT ๋ฐ์ดํ„ฐ๋ฅผ 100๋งŒ ๊ฐœ์˜ ์˜ˆ์ œ๋กœ ํ™•์žฅํ•˜์—ฌ LLaMA-2 7B ๋ชจ๋ธ(Touvron et al., 2023)์ด ๋ฐ์ดํ„ฐ ํ™•์žฅ์˜ ์ด์ ์„ ๋ˆ„๋ฆด ์ˆ˜ ์žˆ์Œ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. MMIQC(Liu and Yao, 2024)๋Š” ์งˆ๋ฌธ-๋‹ต๋ณ€์˜ ๋ฌธ๊ตฌ๋ฅผ ๋ฐ”๊พธ๊ฑฐ๋‚˜ MetaMath์—์„œ ์ง์ ‘ ๊ฐ€์ ธ์˜จ SFT ์Šคํƒ€์ผ์˜ ๋ฐ์ดํ„ฐ๋ฅผ OpenWebMath(Paster et al., 2023)์™€ ๊ฐ™์€ ๊ณ ํ’ˆ์งˆ ์ˆ˜ํ•™ ์‚ฌ์ „ ํ•™์Šต ๋ฐ์ดํ„ฐ์˜ ํ•˜์œ„ ์ง‘ํ•ฉ์œผ๋กœ ํ˜ผํ•ฉํ•œ ๋ฐ์ดํ„ฐ ๋ฌถ์Œ์„ ๊ตฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.
Synthetic data might cause ambiguity in AI alignment.The increasing use of synthetic data in aligning AI models (e.g., Constitutional AI (Bai et al., 2022)) can introduce significant ambiguity and uncertainty. The goal of AI alignment is to ensure that AI systems behave in ways that are aligned with human values and intentions. However, synthetic data, which is artificially generated rather than collected from real-world sources, may not accurately represent the nuances and complexities of human values and preferences (Zhou et al., 2024). This discrepancy can lead to AI models learning from data that is biased (Feng et al., 2023; Liu et al., 2021), ungrounded (Liu et al., 2022; Patel and Pavlick, 2022), or misrepresentative of real-world scenarios (Ji et al., 2023; Weidinger et al., 2021). As a result, AI systems trained on synthetic data may exhibit behaviors that are misaligned with human expectations, potentially leading to unintended consequences or even harmful actions (Anderljung et al., 2023; Zou et al., 2023). Moreover, the ambiguity introduced by synthetic data can make it challenging to interpret and understand the decision-making processes of AI models (Lightman et al., 2023), further complicating the task of ensuring alignment. To mitigate these risks, it is crucial for researchers to carefully consider the limitations and potential drawbacks of using synthetic data in alignment research and to develop robust methods for validating and testing AI models trained on such data.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” AI alignment์— ๋ชจํ˜ธํ•จ์„ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI ๋ชจ๋ธ์„ alignmentํ•˜๋Š” ๋ฐ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ(์˜ˆ: Constitutional AI(Bai et al., 2022)) ์ƒ๋‹นํ•œ ๋ชจํ˜ธํ•จ๊ณผ ๋ถˆํ™•์‹ค์„ฑ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI alignment์˜ ๋ชฉํ‘œ๋Š” AI ์‹œ์Šคํ…œ์ด ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์˜๋„์— ๋ถ€ํ•ฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์†Œ์Šค์—์„œ ์ˆ˜์ง‘ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ธ๊ณต์ ์œผ๋กœ ์ƒ์„ฑ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„์˜ ๋‰˜์•™์Šค์™€ ๋ณต์žก์„ฑ์„ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Zhou et al., 2024). ์ด๋Ÿฌํ•œ ๋ถˆ์ผ์น˜๋Š” AI ๋ชจ๋ธ์ด ํŽธํ–ฅ๋œ(Feng et al., 2023; Liu et al., 2021), ๊ทผ๊ฑฐ๊ฐ€ ์—†๋Š”(Liu et al., 2022; Patel and Pavlick, 2022) ๋˜๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ž˜๋ชป ๋Œ€ํ‘œํ•˜๋Š” ๋ฐ์ดํ„ฐ(Ji et al., 2023; Weidinger et al., 2021)๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒฐ๊ณผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ์‹œ์Šคํ…œ์€ ์ธ๊ฐ„์˜ ๊ธฐ๋Œ€์— ๋งž์ง€ ์•Š๋Š” ํ–‰๋™์„ ๋ณด์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋‚˜ ์‹ฌ์ง€์–ด ์œ ํ•ดํ•œ ํ–‰๋™์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Anderljung et al., 2023; Zou et al., 2023). ๋˜ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋„์ž…๋œ ๋ชจํ˜ธ์„ฑ์œผ๋กœ ์ธํ•ด AI ๋ชจ๋ธ์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์„ ํ•ด์„ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์›Œ์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ(Lightman et al., 2023), alignment๋ฅผ ๋ณด์žฅํ•˜๋Š” ์ž‘์—…์ด ๋”์šฑ ๋ณต์žกํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ„ํ—˜์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ alignment ์—ฐ๊ตฌ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ์˜ ํ•œ๊ณ„์™€ ์ž ์žฌ์  ๋‹จ์ ์„ ์‹ ์ค‘ํ•˜๊ฒŒ ๊ณ ๋ คํ•˜๊ณ , ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ๋ชจ๋ธ์„ ๊ฒ€์ฆํ•˜๊ณ  ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” AI ์ •๋ ฌ์— ๋ชจํ˜ธํ•จ์„ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI ๋ชจ๋ธ์„ ์ •๋ ฌํ•˜๋Š” ๋ฐ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์˜ ์‚ฌ์šฉ์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ(์˜ˆ: ํ—Œ๋ฒ• AI(Bai et al., 2022)) ์ƒ๋‹นํ•œ ๋ชจํ˜ธํ•จ๊ณผ ๋ถˆํ™•์‹ค์„ฑ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. AI ์ •๋ ฌ์˜ ๋ชฉํ‘œ๋Š” AI ์‹œ์Šคํ…œ์ด ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์˜๋„์— ๋ถ€ํ•ฉํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ž‘๋™ํ•˜๋„๋ก ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์†Œ์Šค์—์„œ ์ˆ˜์ง‘ํ•œ ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ธ๊ณต์ ์œผ๋กœ ์ƒ์„ฑ๋œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ์ธ๊ฐ„์˜ ๊ฐ€์น˜์™€ ์„ ํ˜ธ๋„์˜ ๋‰˜์•™์Šค์™€ ๋ณต์žก์„ฑ์„ ์ •ํ™•ํ•˜๊ฒŒ ํ‘œํ˜„ํ•˜์ง€ ๋ชปํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Zhou et al., 2024). ์ด๋Ÿฌํ•œ ๋ถˆ์ผ์น˜๋Š” AI ๋ชจ๋ธ์ด ํŽธํ–ฅ๋œ(Feng et al., 2023; Liu et al., 2021), ๊ทผ๊ฑฐ๊ฐ€ ์—†๋Š”(Liu et al., 2022; Patel and Pavlick, 2022) ๋˜๋Š” ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ์ž˜๋ชป ๋Œ€ํ‘œํ•˜๋Š” ๋ฐ์ดํ„ฐ(Ji et al., 2023; Weidinger et al., 2021)๋กœ ํ•™์Šตํ•˜๋Š” ๊ฒฐ๊ณผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ์ ์œผ๋กœ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ์‹œ์Šคํ…œ์€ ์ธ๊ฐ„์˜ ๊ธฐ๋Œ€์— ๋งž์ง€ ์•Š๋Š” ํ–‰๋™์„ ๋ณด์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜๋„ํ•˜์ง€ ์•Š์€ ๊ฒฐ๊ณผ๋‚˜ ์‹ฌ์ง€์–ด ์œ ํ•ดํ•œ ํ–‰๋™์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(Anderljung et al., 2023; Zou et al., 2023). ๋˜ํ•œ ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋กœ ์ธํ•ด ๋„์ž…๋œ ๋ชจํ˜ธ์„ฑ์œผ๋กœ ์ธํ•ด AI ๋ชจ๋ธ์˜ ์˜์‚ฌ ๊ฒฐ์ • ๊ณผ์ •์„ ํ•ด์„ํ•˜๊ณ  ์ดํ•ดํ•˜๊ธฐ ์–ด๋ ค์›Œ์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ(Lightman et al., 2023), ์ •๋ ฌ์„ ๋ณด์žฅํ•˜๋Š” ์ž‘์—…์ด ๋”์šฑ ๋ณต์žกํ•ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์œ„ํ—˜์„ ์™„ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌ์ž๋“ค์€ ์ •๋ ฌ ์—ฐ๊ตฌ์— ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ์˜ ํ•œ๊ณ„์™€ ์ž ์žฌ์  ๋‹จ์ ์„ ์‹ ์ค‘ํ•˜๊ฒŒ ๊ณ ๋ คํ•˜๊ณ , ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ AI ๋ชจ๋ธ์„ ๊ฒ€์ฆํ•˜๊ณ  ํ…Œ์ŠคํŠธํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ•๋ ฅํ•œ ๋ฐฉ๋ฒ•์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.
Multilingual language models have demonstrated a high level of zero-shot or few-shot cross-lingual transferability across a wide range of tasks Wu and Dredze (2019); Pires et al. (2019); Winata et al. (2021). This means that they can acquire the language capability from supervised data in one language and apply it to another without or with few additional training data. The mechanism behind the strong cross-lingual performance has been investigated by the researchers. It has been shown that multilingual language models have inferred universal rules applicable to any language Artetxe et al. (2020); Chi et al. (2020); Conneau et al. (2020). Contrary to the common hypothesis that multilingual multilingual language models such as mBERT Devlin et al. (2019) rely on a shared subword vocabulary and joint pretraining across multiple languages Pires et al. (2019); Cao et al. (2020); Wu and Dredze (2019), researchers have developed new understandings on the models, emphasizing the models' ability to learn universal semantic abstractions Artetxe et al. (2020); Chi et al. (2020). In terms of the factors that influence cross-lingual performance, researchers have associated transferability with parameter sharing Conneau et al. (2020); Dufter and Schutze (2020); Wu et al. (2022) and language distance Conneau et al. (2020); Eronen et al. (2023). We here further investigate the cross-lingual transferability of language models with new LLaMA-based experiments, presenting outcomes from a different aspect.
๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋†’์€ ์ˆ˜์ค€์˜ zero-shot ๋˜๋Š” few-shot ๊ต์ฐจ ์–ธ์–ด ์ „์†ก์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค Wu์™€ Dredze(2019); Pires ๋“ฑ(2019); Winata ๋“ฑ(2021). ์ฆ‰, ํ•œ ์–ธ์–ด์˜ supervised data์—์„œ ์–ธ์–ด ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•˜์—ฌ ์ถ”๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๊ฑฐ๋‚˜ ๊ฑฐ์˜ ์—†๋Š” ์ƒํƒœ์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ๊ฐ„ ์„ฑ๋Šฅ์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌ์ž๋“ค์ด ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ๋ชจ๋“  ์–ธ์–ด์— ์ ์šฉ๋˜๋Š” ๋ณดํŽธ์ ์ธ ๊ทœ์น™์„ ์ถ”๋ก ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค Artetxe et al. (2020); Chi et al. (2020); Conneau et al. (2020). mBERT์™€ ๊ฐ™์€ ๋‹ค๊ตญ์–ด ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ณต์œ  ์„œ๋ธŒ์›Œ๋“œ ์–ดํœ˜์™€ ์—ฌ๋Ÿฌ ์–ธ์–ด์— ๋Œ€ํ•œ joint pretraining์— ์˜์กดํ•œ๋‹ค๋Š” ์ผ๋ฐ˜์ ์ธ ๊ฐ€์„ค๊ณผ๋Š” ๋‹ฌ๋ฆฌ Pires et al. (2019); Cao et al. (2020); Wu์™€ Dredze(2019), ์—ฐ๊ตฌ์ž๋“ค์€ ๋ชจ๋ธ์˜ ๋ณดํŽธ์ ์ธ ์˜๋ฏธ ์ถ”์ƒํ™” ํ•™์Šต ๋Šฅ๋ ฅ์„ ๊ฐ•์กฐํ•˜๋ฉด์„œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ์ดํ•ด๋ฅผ ๋ฐœ์ „์‹œ์ผฐ์Šต๋‹ˆ๋‹ค Artetxe et al. (2020); Chi et al. (2020). ์–ธ์–ด ๊ฐ„ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์š”์ธ ์ธก๋ฉด์—์„œ ์—ฐ๊ตฌ์ž๋“ค์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ณต์œ  Conneau et al. (2020); Dufter and Schutze (2020); Wu et al. (2022) ๋ฐ ์–ธ์–ด ๊ฑฐ๋ฆฌ Conneau et al. (2020); Eronen et al. (2023). ์—ฌ๊ธฐ์„œ๋Š” ์ƒˆ๋กœ์šด LLaMA ๊ธฐ๋ฐ˜ ์‹คํ—˜์„ ํ†ตํ•ด ์–ธ์–ด ๋ชจ๋ธ์˜ ์–ธ์–ด ๊ฐ„ ์ „์ด์„ฑ์„ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ•˜์—ฌ ๋‹ค๋ฅธ ์ธก๋ฉด์—์„œ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์€ ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ ๋†’์€ ์ˆ˜์ค€์˜ ์ œ๋กœ ์ƒท ๋˜๋Š” ๋ช‡ ๊ฐœ์˜ ์ƒท ๊ต์ฐจ ์–ธ์–ด ์ „์†ก์„ฑ์„ ์ž…์ฆํ–ˆ์Šต๋‹ˆ๋‹ค Wu์™€ Dredze(2019); Pires ๋“ฑ(2019); Winata ๋“ฑ(2021). ์ฆ‰, ํ•œ ์–ธ์–ด์˜ ์ง€๋„ ๋ฐ์ดํ„ฐ์—์„œ ์–ธ์–ด ๊ธฐ๋Šฅ์„ ์Šต๋“ํ•˜์—ฌ ์ถ”๊ฐ€ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ์—†๊ฑฐ๋‚˜ ๊ฑฐ์˜ ์—†๋Š” ์ƒํƒœ์—์„œ ๋‹ค๋ฅธ ์–ธ์–ด์— ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ•๋ ฅํ•œ ์–ธ์–ด ๊ฐ„ ์„ฑ๋Šฅ์˜ ๊ธฐ๋ฐ˜์ด ๋˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์—ฐ๊ตฌ์ž๋“ค์ด ์กฐ์‚ฌํ–ˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ๋ชจ๋“  ์–ธ์–ด์— ์ ์šฉ๋˜๋Š” ๋ณดํŽธ์ ์ธ ๊ทœ์น™์„ ์ถ”๋ก ํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์Šต๋‹ˆ๋‹ค Artetxe et al. (2020); Chi et al. (2020); Conneau et al. (2020). mBERT์™€ ๊ฐ™์€ ๋‹ค๊ตญ์–ด ๋‹ค๊ตญ์–ด ์–ธ์–ด ๋ชจ๋ธ์ด ๊ณต์œ  ์„œ๋ธŒ์›Œ๋“œ ์–ดํœ˜์™€ ์—ฌ๋Ÿฌ ์–ธ์–ด์— ๋Œ€ํ•œ ๊ณต๋™ ์‚ฌ์ „ ํ•™์Šต์— ์˜์กดํ•œ๋‹ค๋Š” ์ผ๋ฐ˜์ ์ธ ๊ฐ€์„ค๊ณผ๋Š” ๋‹ฌ๋ฆฌ Pires et al. (2019); Cao et al. (2020); Wu์™€ Dredze(2019), ์—ฐ๊ตฌ์ž๋“ค์€ ๋ชจ๋ธ์˜ ๋ณดํŽธ์ ์ธ ์˜๋ฏธ ์ถ”์ƒํ™” ํ•™์Šต ๋Šฅ๋ ฅ์„ ๊ฐ•์กฐํ•˜๋ฉด์„œ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ƒˆ๋กœ์šด ์ดํ•ด๋ฅผ ๋ฐœ์ „์‹œ์ผฐ์Šต๋‹ˆ๋‹ค Artetxe et al. (2020); Chi et al. (2020). ์–ธ์–ด ๊ฐ„ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์š”์ธ ์ธก๋ฉด์—์„œ ์—ฐ๊ตฌ์ž๋“ค์€ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ณต์œ  Conneau et al. (2020); Dufter and Schutze (2020); Wu et al. (2022) ๋ฐ ์–ธ์–ด ๊ฑฐ๋ฆฌ Conneau et al. (2020); Eronen et al. (2023). ์—ฌ๊ธฐ์„œ๋Š” ์ƒˆ๋กœ์šด LLaMA ๊ธฐ๋ฐ˜ ์‹คํ—˜์„ ํ†ตํ•ด ์–ธ์–ด ๋ชจ๋ธ์˜ ์–ธ์–ด ๊ฐ„ ์ „์ด์„ฑ์„ ์ถ”๊ฐ€๋กœ ์กฐ์‚ฌํ•˜์—ฌ ๋‹ค๋ฅธ ์ธก๋ฉด์—์„œ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.
**Output**[[**Retriee**]=No[The manager realizes that there is an extensive list of tasks that need to be completed in order to make the event a success.[[**Retriee**]=Yes]<p>Delegation-Delegation is the assignment of any responsibility or authority to another person (normally from a manager to a subordinate) to carry out specific activities. It is one of the core concepts of management leadership. However, the person who delegated the work remains accountable for the outcome of the delegated work. Delegation empowers a subordinate to make decisions, i.e. it is a shifting of decision-making authority from one organizational level to a lower one. Delegation, if properly done, is not fabrication. The opposite of effective delegation is micromanagement, where a manager provides too much input,<p>[[**Isstr**]=RelevantThis includes tasks such as creating a guest list, designing and sending invitations, booking a venue, coordinating with vendors, planning the event program, and monitoring the budget.[[[**Isstr**]=Partially Supported[**Retriee**]=No[Since it is not possible for the manager to handle everything alone, they must delegate tasks to their team members.[[**Retriee**]=No[The manager would then assign each team member specific responsibilities based on their individual skillsets and strengths.[[**Retriee**]=Yes]<p>Joint application design - Subject Matter Experts: These are the business users, the IS professionals, and the outside experts that will be needed for a successful workshop.
**์ถœ๋ ฅ**[[**Retriee**]=No[๊ด€๋ฆฌ์ž๋Š” ์ด๋ฒคํŠธ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ์ง„ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์™„๋ฃŒํ•ด์•ผ ํ•  ์ž‘์—… ๋ชฉ๋ก์ด ๊ด‘๋ฒ”์œ„ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.[[**Retriee**]=Yes]<p>Delegation-Delegation์€ ํŠน์ • ํ™œ๋™์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค๋ฅธ ์‚ฌ๋žŒ(์ผ๋ฐ˜์ ์œผ๋กœ ๊ด€๋ฆฌ์ž๊ฐ€ ๋ถ€ํ•˜ ์ง์›์—๊ฒŒ)์—๊ฒŒ ์ฑ…์ž„์ด๋‚˜ ๊ถŒํ•œ์„ ํ• ๋‹นํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๊ด€๋ฆฌ ๋ฆฌ๋”์‹ญ์˜ ํ•ต์‹ฌ ๊ฐœ๋… ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—…๋ฌด๋ฅผ ์œ„์ž„ํ•œ ์‚ฌ๋žŒ์€ ์œ„์ž„๋œ ์—…๋ฌด์˜ ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด ์ฑ…์ž„์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. Delegation์€ ๋ถ€ํ•˜ ์ง์›์ด ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋„๋ก ๊ถŒํ•œ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์˜์‚ฌ ๊ฒฐ์ • ๊ถŒํ•œ์„ ํ•œ ์กฐ์ง ์ˆ˜์ค€์—์„œ ํ•˜์œ„ ์ˆ˜์ค€์œผ๋กœ ์ด๋™ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Delegation์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€๋ฉด ์œ„์กฐ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ Delegation์˜ ๋ฐ˜๋Œ€๋Š” ๊ด€๋ฆฌ์ž๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์€ ์˜๊ฒฌ์„ ์ œ์‹œํ•˜๋Š” micromanagement๋กœ, Delegation์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์œผ๋ฉด ์œ„์ž„์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์Šต๋‹ˆ๋‹ค.<p>[[**Isstr**]=RelevantThis includes tasks such as creating a guest list, designing and sending invitations, booking a venue, coordinating with vendors, planning the event program, and monitoring the budget.[[[**Isstr**]=Partially Supported[**Retriee**]=No[๊ด€๋ฆฌ์ž๊ฐ€ ํ˜ผ์ž์„œ ๋ชจ๋“  ๊ฒƒ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ํŒ€์›์—๊ฒŒ ์—…๋ฌด๋ฅผ ์œ„์ž„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[[**Retriee**]=No[๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ด€๋ฆฌ์ž๋Š” ๊ฐ ํŒ€์›์—๊ฒŒ ๊ฐœ์ธ์˜ ๊ธฐ์ˆ ๊ณผ ๊ฐ•์ ์— ๋”ฐ๋ผ ํŠน์ • ์ฑ…์ž„์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค.[[**Retriee**]=Yes]<p>Joint application design - Subject Matter Experts: These are the business users, the IS professionals, and the outside experts that will be needed for a successful workshop.
**์ถœ๋ ฅ**[[**Retriee**]=No[๊ด€๋ฆฌ์ž๋Š” ์ด๋ฒคํŠธ๋ฅผ ์„ฑ๊ณต์ ์œผ๋กœ ์ง„ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์™„๋ฃŒํ•ด์•ผ ํ•  ์ž‘์—… ๋ชฉ๋ก์ด ๊ด‘๋ฒ”์œ„ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.[[**Retriee**]=Yes]<p>์œ„์ž„-์œ„์ž„์€ ํŠน์ • ํ™œ๋™์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค๋ฅธ ์‚ฌ๋žŒ(์ผ๋ฐ˜์ ์œผ๋กœ ๊ด€๋ฆฌ์ž๊ฐ€ ๋ถ€ํ•˜ ์ง์›์—๊ฒŒ)์—๊ฒŒ ์ฑ…์ž„์ด๋‚˜ ๊ถŒํ•œ์„ ํ• ๋‹นํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ๊ด€๋ฆฌ ๋ฆฌ๋”์‹ญ์˜ ํ•ต์‹ฌ ๊ฐœ๋… ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์—…๋ฌด๋ฅผ ์œ„์ž„ํ•œ ์‚ฌ๋žŒ์€ ์œ„์ž„๋œ ์—…๋ฌด์˜ ๊ฒฐ๊ณผ์— ๋Œ€ํ•ด ์ฑ…์ž„์„ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ์œ„์ž„์€ ๋ถ€ํ•˜ ์ง์›์ด ์˜์‚ฌ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ๋„๋ก ๊ถŒํ•œ์„ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์œผ๋กœ, ์˜์‚ฌ ๊ฒฐ์ • ๊ถŒํ•œ์„ ํ•œ ์กฐ์ง ์ˆ˜์ค€์—์„œ ํ•˜์œ„ ์ˆ˜์ค€์œผ๋กœ ์ด๋™ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์œ„์ž„์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€๋ฉด ์œ„์กฐ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. ํšจ๊ณผ์ ์ธ ์œ„์ž„์˜ ๋ฐ˜๋Œ€๋Š” ๊ด€๋ฆฌ์ž๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์€ ์˜๊ฒฌ์„ ์ œ์‹œํ•˜๋Š” ๋งˆ์ดํฌ๋กœ ๊ด€๋ฆฌ๋กœ, ์œ„์ž„์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์œผ๋ฉด ์œ„์ž„์ด ์ œ๋Œ€๋กœ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š์Šต๋‹ˆ๋‹ค.<p>[[**Isstr**]=๊ด€๋ จ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ฒŒ์ŠคํŠธ ๋ชฉ๋ก ์ž‘์„ฑ, ์ดˆ๋Œ€์žฅ ๋””์ž์ธ ๋ฐ ๋ฐœ์†ก, ์žฅ์†Œ ์˜ˆ์•ฝ, ๊ณต๊ธ‰์—…์ฒด์™€์˜ ์กฐ์œจ, ์ด๋ฒคํŠธ ํ”„๋กœ๊ทธ๋žจ ๊ณ„ํš, ์˜ˆ์‚ฐ ๋ชจ๋‹ˆํ„ฐ๋ง ๋“ฑ์˜ ์ž‘์—…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.[[[**Isstr**]=๋ถ€๋ถ„์ ์œผ๋กœ ์ง€์›๋จ[**Retriee**]=No[๊ด€๋ฆฌ์ž๊ฐ€ ํ˜ผ์ž์„œ ๋ชจ๋“  ๊ฒƒ์„ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ํŒ€์›์—๊ฒŒ ์—…๋ฌด๋ฅผ ์œ„์ž„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.[[**Retriee**]=No[๊ทธ๋Ÿฐ ๋‹ค์Œ ๊ด€๋ฆฌ์ž๋Š” ๊ฐ ํŒ€์›์—๊ฒŒ ๊ฐœ์ธ์˜ ๊ธฐ์ˆ ๊ณผ ๊ฐ•์ ์— ๋”ฐ๋ผ ํŠน์ • ์ฑ…์ž„์„ ํ• ๋‹นํ•ฉ๋‹ˆ๋‹ค.[[**Retriee**]=Yes]<p>๊ณต๋™ ์ ์šฉ ์„ค๊ณ„ - ์ฃผ์ œ๋ณ„ ์ „๋ฌธ๊ฐ€: ์„ฑ๊ณต์ ์ธ ์›Œํฌ์ˆ์„ ์œ„ํ•ด ํ•„์š”ํ•œ ๋น„์ฆˆ๋‹ˆ์Šค ์‚ฌ์šฉ์ž, IS ์ „๋ฌธ๊ฐ€ ๋ฐ ์™ธ๋ถ€ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค.
To construct the dataset, we initialized it with the 52K instruction dataset of Alpaca. After iteratively performing \(M\) evolutions, where \(M=4\), we obtained a 250K instructions. For each instruction in each round of evolution, we randomly select one evolving prompt from total six prompts (i.e., five from in-depth evolving and one from in-breadth evolving) with equal probability. We execute above process using Azure OpenAI ChatGPT API7. Finally, we also leveraged the ChatGPT to generate responses. We use a temperature of 1 to generate response and set the maximum number of tokens for generation to \(2048\). Additionally, we set the frequency penalty to zero and top-p to 0.9. Totally, we request the API 52 \(\times\) 4 \(\times\) 3 = 624K times to construct the full dataset. We use pre-trained LLaMA 7B [4] to initialize our model. We adopt Adam optimizer as an initial learning rate of 2 \(\times 10^{-5}\), a maximum number of tokens \(2048\), and the batch size is 8 for each GPU. We train our model on 8 V100 GPUs with Deepspeed Zero-3 for 70 hours on 3 epochs. For fair comparison, we replace Alpaca's original Davici-003 response with ChatGPT's response, and also sample 70K instructions subset to train _WizardLM_. For inference, we set the temperature to 1 and the top-p to 0.9 in the same setting for _WizardLM_ and baseline models to reduce output randomness and ensure more focused and deterministic outputs. We also use a beam size of 1, and set the maximum generation length to 2048.
๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Alpaca์˜ 52K instruction ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. \(M=4\)์ธ \(M\) ์ง„ํ™”๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•œ ํ›„ 250K instructions๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ง„ํ™” ๋ผ์šด๋“œ์˜ ๊ฐ instruction์— ๋Œ€ํ•ด ์ด 6๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ(์ฆ‰, ์‹ฌ์ธต ์ง„ํ™” 5๊ฐœ, ํญ ์ง„ํ™” 1๊ฐœ)์—์„œ ๋™์ผํ•œ ํ™•๋ฅ ๋กœ ํ•˜๋‚˜์˜ ์ง„ํ™” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์œ„์˜ ํ”„๋กœ์„ธ์Šค๋Š” Azure OpenAI ChatGPT API7์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ChatGPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์‘๋‹ต์„ ์ƒ์„ฑํ•  ๋•Œ temperature๋ฅผ 1๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์ƒ์„ฑ์„ ์œ„ํ•œ ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ \(2048\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ frequency penalty๋ฅผ 0์œผ๋กœ ์„ค์ •ํ•˜๊ณ  top-p๋ฅผ 0.9๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด 52 \(\times\) 4 \(\times\) 3 = 624K๋ฒˆ์˜ API๋ฅผ ์š”์ฒญํ•˜์—ฌ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ LLaMA 7B [4]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ดˆ๊ธฐ learning rate๋Š” 2 \(\times 10^{-5}\), ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋Š” \(2048\), batch size๋Š” ๊ฐ GPU๋‹น 8์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. 8๊ฐœ์˜ V100 GPU์—์„œ 3ํšŒ์— ๊ฑธ์ณ 70์‹œ๊ฐ„ ๋™์•ˆ Deepspeed Zero-3์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด Alpaca์˜ ์›๋ž˜ Davici-003 ์‘๋‹ต์„ ChatGPT์˜ ์‘๋‹ต์œผ๋กœ ๋Œ€์ฒดํ•˜๊ณ , 70K instructions ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ _WizardLM_์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด _WizardLM_๊ณผ baseline ๋ชจ๋ธ์— ๋™์ผํ•œ ์„ค์ •์œผ๋กœ temperature๋ฅผ 1๋กœ, top-p๋ฅผ 0.9๋กœ ์„ค์ •ํ•˜์—ฌ ์ถœ๋ ฅ ๋ฌด์ž‘์œ„์„ฑ์„ ์ค„์ด๊ณ  ๋ณด๋‹ค ์ง‘์ค‘๋˜๊ณ  ๊ฒฐ์ •๋ก ์ ์ธ ์ถœ๋ ฅ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ beam size๋ฅผ 1๋กœ ์„ค์ •ํ•˜๊ณ  ์ตœ๋Œ€ ์ƒ์„ฑ ๊ธธ์ด๋ฅผ 2048๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด Alpaca์˜ 52K ๋ช…๋ น์–ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋กœ ์ดˆ๊ธฐํ™”ํ–ˆ์Šต๋‹ˆ๋‹ค. M=4์ธ \(M\) ์ง„ํ™”๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ˆ˜ํ–‰ํ•œ ํ›„ 250K ๋ช…๋ น์–ด๋ฅผ ์–ป์—ˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ์ง„ํ™” ๋ผ์šด๋“œ์˜ ๊ฐ ๋ช…๋ น์— ๋Œ€ํ•ด ์ด 6๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ(์ฆ‰, ์‹ฌ์ธต ์ง„ํ™” 5๊ฐœ, ํญ ์ง„ํ™” 1๊ฐœ)์—์„œ ๋™์ผํ•œ ํ™•๋ฅ ๋กœ ํ•˜๋‚˜์˜ ์ง„ํ™” ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค. ์œ„์˜ ํ”„๋กœ์„ธ์Šค๋Š” Azure OpenAI ChatGPT API7์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ChatGPT๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‘๋‹ต์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์‘๋‹ต์„ ์ƒ์„ฑํ•  ๋•Œ ์˜จ๋„๋ฅผ 1๋กœ ์‚ฌ์šฉํ•˜๊ณ  ์ƒ์„ฑ์„ ์œ„ํ•œ ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋ฅผ \(2048\)๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋นˆ๋„ ํŽ˜๋„ํ‹ฐ๋ฅผ 0์œผ๋กœ ์„ค์ •ํ•˜๊ณ  ์ƒ์œ„ p๋ฅผ 0.9๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ์ด 52 \(\๋ฐฐ์ˆ˜\) 4 \(\๋ฐฐ์ˆ˜\) 3 = 624K๋ฒˆ์˜ API๋ฅผ ์š”์ฒญํ•˜์—ฌ ์ „์ฒด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ์ „ ํ•™์Šต๋œ LLaMA 7B [4]๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•ฉ๋‹ˆ๋‹ค. ์ดˆ๊ธฐ ํ•™์Šต๋ฅ ์€ 2 \(\times 10^{-5}\), ์ตœ๋Œ€ ํ† ํฐ ์ˆ˜๋Š” \(2048\), ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” ๊ฐ GPU๋‹น 8์„ ์ฑ„ํƒํ•ฉ๋‹ˆ๋‹ค. 8๊ฐœ์˜ V100 GPU์—์„œ 3ํšŒ์— ๊ฑธ์ณ 70์‹œ๊ฐ„ ๋™์•ˆ Deepspeed Zero-3์œผ๋กœ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ๊ณต์ •ํ•œ ๋น„๊ต๋ฅผ ์œ„ํ•ด ์•ŒํŒŒ์นด์˜ ์›๋ž˜ Davici-003 ์‘๋‹ต์„ ChatGPT์˜ ์‘๋‹ต์œผ๋กœ ๋Œ€์ฒดํ•˜๊ณ , 7๋งŒ ๊ฐœ์˜ ๋ช…๋ น์–ด ํ•˜์œ„ ์ง‘ํ•ฉ์„ ์ƒ˜ํ”Œ๋งํ•˜์—ฌ _WizardLM_์„ ํ›ˆ๋ จํ•ฉ๋‹ˆ๋‹ค. ์ถ”๋ก ์„ ์œ„ํ•ด _WizardLM_๊ณผ ๊ธฐ์ค€ ๋ชจ๋ธ์— ๋™์ผํ•œ ์„ค์ •์œผ๋กœ ์˜จ๋„๋ฅผ 1๋กœ, ์ƒ์œ„ p๋ฅผ 0.9๋กœ ์„ค์ •ํ•˜์—ฌ ์ถœ๋ ฅ ๋ฌด์ž‘์œ„์„ฑ์„ ์ค„์ด๊ณ  ๋ณด๋‹ค ์ง‘์ค‘๋˜๊ณ  ๊ฒฐ์ •๋ก ์ ์ธ ์ถœ๋ ฅ์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๋น” ํฌ๊ธฐ๋ฅผ 1๋กœ ์„ค์ •ํ•˜๊ณ  ์ตœ๋Œ€ ์ƒ์„ฑ ๊ธธ์ด๋ฅผ 2048๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.