REFORM uses reward-guided controlled decoding to generate preference-class-consistent responses that the reward model mis-scores, then retrains the reward model on these failure modes to improve robustness.
Interpreting language reward models via contrastive explanations
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling
REFORM uses reward-guided controlled decoding to generate preference-class-consistent responses that the reward model mis-scores, then retrains the reward model on these failure modes to improve robustness.