RewardBench

'RewardBench' is a dataset used to evaluate the performance and reliability of large language model (LLM) evaluation panels by measuring the informational value and independence of their aggregated votes.

Papers using RewardBench (17)

PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling2025 · 4 citesNine Judges, Two Effective Votes: Correlated Errors Undermine LLM Evaluation Panels2026CDRRM: Contrast-Driven Rubric Generation for Reliable and Interpretable Reward Modeling2026SCOPE: Selective Conformal Optimized Pairwise LLM Judging2026SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment2025Explicit Reasoning Makes Better Judges: A Systematic Study on Accuracy, Efficiency, and Robustness2025Intra-Trajectory Consistency for Reward Modeling2025Time To Impeach LLM-as-a-Judge: Programs are the Future of Evaluation2025Robust Reward Modeling via Causal Rubrics2025ENCORE: Entropy-guided Reward Composition for Multi-head Safety Reward Models2025Sentence-level Reward Model can Generalize Better for Aligning LLM from Human Preference2025IPO: Your Language Model is Secretly a Preference Classifier2025Data-adaptive Safety Rules for Training Reward Models2025RRM: Robust Reward Model Training Mitigates Reward Hacking2024Uncertainty-aware Reward Model: Teaching Reward Models to Know What is Unknown2024Margin Matching Preference Optimization: Enhanced Model Alignment with Granular Feedback2024M-RewardBench: Evaluating Reward Models in Multilingual Settings2024