RewardBench 2

Properties

axis: methodology
verification: det-gold
whatItTests: Multi-skill reward-model evaluation benchmark (instruction following, reasoning, safety and other domains) using newly-sourced human prompts rather than recycled downstream-eval prompts, scored for accuracy against preference data and checked for correlation with downstream RLHF/best-of-N performance.
discriminates: [judge-calibration]
saturation: open
patternLabFit: Relevant wherever a reward-model-style verifier or preference-scorer stands in for a judge; its finding that scores drop ~20 points versus the original RewardBench flags overfitting risk in judge/verifier calibration.
source: RewardBench 2: Advancing Reward Model Evaluation (arXiv:2506.01937)
notes: Scope note: evaluates reward models (verifiers used in RLHF/best-of-N), not LLM judges of agent output directly, but is the closest reward/verifier-calibration analog to JudgeBench and is distinct from it (JudgeBench = LLM judges on objectively verifiable tasks; RewardBench 2 = reward models scored for accuracy on preference data with downstream-correlation checks).

axis: methodology
verification: det-gold
whatItTests: Multi-skill reward-model evaluation benchmark (instruction following, reasoning, safety and other domains) using newly-sourced human prompts rather than recycled downstream-eval prompts, scored for accuracy against preference data and checked for correlation with downstream RLHF/best-of-N performance.
discriminates: [judge-calibration]
saturation: open
patternLabFit: Relevant wherever a reward-model-style verifier or preference-scorer stands in for a judge; its finding that scores drop ~20 points versus the original RewardBench flags overfitting risk in judge/verifier calibration.
source: RewardBench 2: Advancing Reward Model Evaluation (arXiv:2506.01937)
notes: Scope note: evaluates reward models (verifiers used in RLHF/best-of-N), not LLM judges of agent output directly, but is the closest reward/verifier-calibration analog to JudgeBench and is distinct from it (JudgeBench = LLM judges on objectively verifiable tasks; RewardBench 2 = reward models scored for accuracy on preference data with downstream-correlation checks).