I think it has to do with the fact that LLM's are stochastic, autoregressive models. There are "trajectories" that it can take that are wrong (hallucinations) a

I think it has to do with the fact that LLM's are stochastic, autoregressive models. There are "trajectories" that it can take that are wrong (hallucinations) and there are trajectories that lead to the right answer. In essence multi-agent is doing something like self consistency, self-reflection.