RLPR: Grounding RLVR in Domains without Verifiers with Probability
The core idea of RLPR is that, the probability the model assigns to the answer, conditioned on its reasoning, is already a good built-in estimator of how good the answer is.
The core idea of RLPR is that, the probability the model assigns to the answer, conditioned on its reasoning, is already a good built-in estimator of how good the answer is.
Date: 2026-10-22 Thu