This is my personal opinion
LLM is essentially a probability model that selects the word (token) with the highest probability of coming next based on context. However, if you start selecting word combinations (branches) with low winning probabilities or unverified outcomes for the 'exploration' of MCTS, it may look plausible the first one or two times, but over time, it generates nonsense (hallucinations) that are completely out of context. In coding terms, this is a phenomenon where you arbitrarily imagine and write a library that does not exist.
If you run an MCTS coin toss relying solely on the token generation probabilities of a simple LLM without any verification mechanisms (compiler, test code, reward model), the result is code that looks plausible on the surface but is completely broken on the inside—a 'lump of hallucination.'
If verification mechanisms are used, enormous resources are consumed.
This is my personal opinion LLM is essentially a probability model that selects the word (token) with the highest probability of coming next based on context. However, if you start selecting word combinations (branches) with low winning probabilities or unverified outcomes for the 'exploration' of MCTS, it may look plausible the first one or two times, but over time, it generates nonsense (hallucinations) that are completely out of context. In coding terms, this is a phenomenon where you arbitrarily imagine and write a library that does not exist.
hallucinations are often good ideas for things that should exist, missing functions or libraries
That's right! Extreme hallucination in a well-aligned state is extremely helpful.
If you run an MCTS coin toss relying solely on the token generation probabilities of a simple LLM without any verification mechanisms (compiler, test code, reward model), the result is code that looks plausible on the surface but is completely broken on the inside—a 'lump of hallucination.'
If verification mechanisms are used, enormous resources are consumed.
What is your question here? Has anyone considered monte carlo search for code gen?
Yes. Alphazero etc. Surely the labs extensively research ideas like this. Do you have a specific task you are inquiring about?