RewardBench

RewardBench is a benchmark dataset used to evaluate the performance of Generative Reward Models (GRMs) in reward modeling by providing a set of preference data for training and assessing pointwise reward predictions.

Papers using RewardBench (15)

PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling2025 · 4 citesreward-lens: A Mechanistic Interpretability Library for Reward Models2026IRPM: Intergroup Relative Preference Modeling for Pointwise Generative Reward Models2026Multi-Agent Collaborative Reward Design for Enhancing Reasoning in Reinforcement Learning2025Tiny Reward Models2025Efficient Online RFT with Plug-and-Play LLM Judges: Unlocking State-of-the-Art Performance2025Intra-Trajectory Consistency for Reward Modeling2025Act-Adaptive Margin: Dynamically Calibrating Reward Models for Subjective Ambiguity2025Sentence-level Reward Model can Generalize Better for Aligning LLM from Human Preference2025Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts2024 · 10 citesRewardBench: Evaluating Reward Models for Language Modeling2024 · 5 citesHelpSteer2: Open-source dataset for training top-performing reward models2024 · 2 citesPost-hoc Reward Calibration: A Case Study on Length Bias2024 · 1 citesQuantile Regression for Distributional Reward Models in RLHF2024Evaluating Robustness of Reward Models for Mathematical Reasoning2024