Yes Graphs are a Structured input now days !
So We can also think of solutions as graphs .
In this respect the plan generated by a model is the solution pathway. ie: the solution to the un completed task . we could say this output graph is a blueprint for similar tasks which can be structured and solved in the same way by plugging in the correct inputs: The graph is more diverse than code or a single algorithm as it also presents a series of Tasks to be performed as well as sub tasks which grab inputs where required which may be detailed and dynamic :
So : we find that the graph as a plan is a great generalizing methodology for training models as we know that agents are created to perform single tasks . most models can be prompted with a role, and a collection of tools and perform the task given a plan to follow . So a weak model may perform tasks which a stronger model has composed. in which the weak model in its formulation of a plan may falter : SO the grandfather model or Orchestrator , can hand jobs off to a collection of agents to perform and then examine the state outputs and make the main decisions for all actions within the task :
So Training for planning is the key to a models ability to perform as an orchestration model, enabling for specialized training of slim specific agents which perform dedicated tasks : Hence the specialist model also will in general outperform larger models on this specialist teaks but indeed fail at others :
I have found that in training a multi-model and a multi modal model , for general tasks this planning training is the most important step :
For a maths model it should be adept at understanding the steps required to solve tasks as well as display correctly algorithms as well as having the grasp in all methodology and formulae etc : IT should not perform the calculation , but prepare the lessons and step by step guides , display the information, interpret the information , gather requirements ... Correctly label inputs and expected outputs: Then this model can be trained to plan Jobs !
It should be also trained on various chain of thoughts as ell as self analysis of results and scoring of agents work etc ! so it can make reviews and critiques: IT should be able to generate graphs ad correct prompts for agents to perform the tasks as well as select the correct tools set for the roles required : then it should be trained in an agentic system to fine tune its role : Here we did not train for the task of calculation , that is for the sub agents , this model can be trained in these things also as it can use itself as an agent ! then e will find the model solving tasks and producing valid graphs which can be initiated and solved with an agentic team and set of tools !
