Recently @karpathy built an “LLM council” to provide advice to users. This got me thinking: what are the best governance rules for the council?
So I built a little experiment testing four different procedures:
(1) The LLMs do a simple majority vote without communicating to each other
(2) LLMs vote, then deliberate, then can update their vote.
(3) LLMs vote, deliberate, update their vote, and then a Chairman makes the decision.
(4) LLMs vote, evaluate each other’s answers, and then the Chairman makes the decision (like Karpathy’s procedure).
I evaluated how well these procedures do compared to individual models on two very basic eval sets with known answers (GSM8K and TruthfulQA). I focused on very cheap models to spare my wallet.
For the very particular setup I created---and I don’t claim this generalizes---it looks like all four approaches do a little bit better than the best model on its own, but option 2, voting with deliberation, does the best. (I didn’t run the study long enough to get “statistical significance” but if someone wants to they probably could).
It turns out, unbeknownst to me, there’s a cool literature in cs already exploring some of these procedural questions, so I don’t think I’ve discovered anything new here (one main example: a Du et al 2023 paper showing that councils outperform individual models).
That being said, this is a cool opportunity for people in the social sciences who study collective decision-making mechanisms and governance.
In the future there will be tons of situations where multiple agents have to interact and make decisions together, and we should develop a science of how to make optimal collective decisions with AI agents.
