Selecting the five safety rules with largest response discrepancy maximizes mutual information under stated assumptions, and a reward model trained with this adaptive labeling achieves top RewardBench safety.
{severe level} harm question:
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
-
Data-adaptive Safety Rules for Training Reward Models
Selecting the five safety rules with largest response discrepancy maximizes mutual information under stated assumptions, and a reward model trained with this adaptive labeling achieves top RewardBench safety.