RewardBench

Benchmarks

TaskDatasetSOTA resultTrendResults

Listwise Judging

RewardBench listwise 2

58.1IF Score

Reward Modeling

RewardBench mathprm

72.5Score

Reward Modeling

RewardBench 2 (test)

76.3RWBench2 Score

LLM-as-a-Judge

RewardBench (test)

2.72Std Dev (Reward)

Pairwise Reference Alignment

RewardBench

0.771A^M

Reward Model Evaluation

RewardBench (test)

1.65Kuiper

LLM-as-a-Judge Calibration

RewardBench (test)

0.022Test Risk (MSE)

Reward Model Evaluation

RewardBench

0.001P-value

Reward Modeling

RewardBench External Evaluation

93.3Chat Score

Reward Modeling

RewardBench and RM-Bench Aggregate 2

90.3Average Score

Reward Modeling

RewardBench RLHFlow source Chat Chat-Hard Safety Reasoning Total 1.0 (train)

80.62Chat Score

Reward Modeling

RewardBench UltraFeedback training source Chat Chat-Hard Safety Reasoning Total 1.0

80.98Chat Score

Multimodal Reward Modeling

RewardBench 2

96.7Safety Score

Reward Modeling

RewardBench v2

57Score

Reward Modeling

RewardBench Filtered (test)

98.3Chat

Reward Modeling

RewardBench Out-of-distribution (OOD) evaluation

99.4Chat

Reward Modeling Evaluation

RewardBench

88.1R-Bench Score

Best-of-N evaluation

RewardBench v2

58.69Accuracy

Identifying preferred LLM outputs

RewardBench Chat Chat-Hard Safety

80.1Overall Accuracy