Reactions from ranked influencers
3 posts
Sampling many solutions and taking the majority answer is the most reliable label-free way to improve LLM reasoning. Training methods reduce the consensus to a filter, a preference, or a scalar reward. What if the model could read them instead?
@iatitov Hard "E-M"?
Consensus is usually compressed into a vote or scalar reward. This work instead uses a consensus solution as context for a frozen self-teacher, turning agreement into dense token-level supervision. Recovers much of the benefit of gold solutions, without labels
Sampling many solutions and taking the majority answer is the most reliable label-free way to improve LLM reasoning. Training methods reduce the consensus to a filter, a preference, or a scalar reward. What if the model could read them instead?