PAC-CF introduces Probably Approximately Correct Conformal Filtering to reduce bias in LLM-guided tree pruning. It uses a calibrated margin derived from the score deficit of verifier-valid continuations to ensure that valid solutions are not prematurely eliminated during search.
HOW THIS AFFECTS YOU
●
builderYou can improve the utility of LLM-guided search by reducing the risk of pruning correct solutions.
●
researcherThis provides a theoretical PAC-guarantee for the decision problem of tree pruning in LLM search.