I don’t think this [sc. that AlphaProof uses an LLM to generate candidate next steps] is true, actually.
Hmm, maybe you’re right. I thought I’d seen something that said it did that, but perhaps I hallucinated it. (What they’ve written isn’t specific enough to make it clear that it doesn’t do that either, at least to me. They say “AlphaProof generates solution candidates”, but nothing about how it generates them. I get the impression that it’s something at least kinda LLM-like, but could be wrong.)
The AlphaZero algorithm doesn’t obviously not involve an LLM. It has a “policy network” to propose moves, and I don’t know what that looks like in the case of AlphaProof. If I had to guess blindly I would guess it’s an LLM, but maybe they’ve got something else instead.
Hmm, maybe you’re right. I thought I’d seen something that said it did that, but perhaps I hallucinated it. (What they’ve written isn’t specific enough to make it clear that it doesn’t do that either, at least to me. They say “AlphaProof generates solution candidates”, but nothing about how it generates them. I get the impression that it’s something at least kinda LLM-like, but could be wrong.)
The diagram actually says it uses the AlphaZero algorithm. Which obviously doesn’t involve an LLM.
The AlphaZero algorithm doesn’t obviously not involve an LLM. It has a “policy network” to propose moves, and I don’t know what that looks like in the case of AlphaProof. If I had to guess blindly I would guess it’s an LLM, but maybe they’ve got something else instead.