Intra-trajectory consistency regularization, weighted by next-token generation probabilities, improves outcome reward models on RewardBench and in downstream DPO and best-of-N evaluations.
Rewarding progress: Scaling automated process verifiers for LLM reasoning
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Intra-Trajectory Consistency for Reward Modeling
Intra-trajectory consistency regularization, weighted by next-token generation probabilities, improves outcome reward models on RewardBench and in downstream DPO and best-of-N evaluations.