Different LLM models for different tasks

[fraricci]

Reading stuff around, I think a better approach to deal with a query from the user
and the different tool is to use dedicated LLM agents, possibly with different models. In particular, one agent at higher level that digest the initial user query and decide what to do. And another (or multiple) agent which are dedicated to generate json input for the different tools and gather the json output.

I think this is what LLaMP is also doing (@chiang-yuan could comment on that).
Also, it seems a general approach suggested by Groq when presenting their new
model, specifically tuned for tool calling: https://wow.groq.com/introducing-llama-3-groq-tool-use-models/.

[jan-janssen]

I was wondering why separate agents are better than one LLM even though the agents use the same LLM in the background. And I found the following paper https://doi.org/10.1016/S0004-3702(99)00022-3 which nicely explains it with a soccer analogy. Basically, the general goal is to win the game, you win the game by scoring goals, so naively you could say every player should aim for scoring goals, then the whole team is focused on scoring goals and that should lead to a success. But the more effective strategy is for the different players to have dedicated roles, a goal keeper, defence players, offensive players and so on. The paper shows that this implementation of roles was successful for their 3rd place in the robocup a simulated soccer competition for robots. I think in the context of agents for materials science this means it makes sense to have distinguished agents with separate levels of expertise. If they can form a coherent picture with their different expertises, then we most likely generated a new insight. If they disagree, for example the experimental agent based on literature comes to a different conclusion than the simulation agent, then the current information is not yet conclusive and they need to study more, do more simulation, read more paper and so on.