daman1209arora commited on
Commit
d070d11
1 Parent(s): 8ce189e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +30 -1
README.md CHANGED
@@ -11,4 +11,33 @@ tags:
11
  pretty_name: jeebench
12
  size_categories:
13
  - n<1K
14
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
  pretty_name: jeebench
12
  size_categories:
13
  - n<1K
14
+ ---
15
+
16
+ # JEEBench(EMNLP 2023)
17
+ Repository for the code and dataset for the paper: "Have LLMs Advanced Enough? A Harder Problem Solving Benchmark For Large Language Models" accepted in EMNLP 2023 as a Main conference paper.
18
+ https://aclanthology.org/2023.emnlp-main.468/
19
+
20
+
21
+ ## Citation
22
+
23
+ If you use our dataset in your research, please cite it using the following
24
+ ```latex
25
+ @inproceedings{arora-etal-2023-llms,
26
+ title = "Have {LLM}s Advanced Enough? A Challenging Problem Solving Benchmark For Large Language Models",
27
+ author = "Arora, Daman and
28
+ Singh, Himanshu and
29
+ {Mausam}",
30
+ editor = "Bouamor, Houda and
31
+ Pino, Juan and
32
+ Bali, Kalika",
33
+ booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
34
+ month = dec,
35
+ year = "2023",
36
+ address = "Singapore",
37
+ publisher = "Association for Computational Linguistics",
38
+ url = "https://aclanthology.org/2023.emnlp-main.468",
39
+ doi = "10.18653/v1/2023.emnlp-main.468",
40
+ pages = "7527--7543",
41
+ abstract = "The performance of large language models (LLMs) on existing reasoning benchmarks has significantly improved over the past years. In response, we present JEEBench, a considerably more challenging benchmark dataset for evaluating the problem solving abilities of LLMs. We curate 515 challenging pre-engineering mathematics, physics and chemistry problems from the highly competitive IIT JEE-Advanced exam. Long-horizon reasoning on top of deep in-domain knowledge is essential for solving problems in this benchmark. Our evaluation on various open-source and proprietary models reveals that the highest performance, even after using techniques like self-consistency, self-refinement and chain-of-thought prompting, is less than 40{\%}. The typical failure modes of GPT-4, the best model, are errors in algebraic manipulation, difficulty in grounding abstract concepts into mathematical equations accurately and failure in retrieving relevant domain-specific concepts. We also observe that by mere prompting, GPT-4 is unable to assess risk introduced by negative marking for incorrect answers. For this, we develop a post-hoc confidence-thresholding method over self-consistency, which enables effective response selection. We hope that our challenging benchmark will guide future re-search in problem-solving using LLMs.",
42
+ }
43
+ ```