Reward Modeling

RewardBench 2

RewardBench 2 tests whether reward models rank the intended best response above the alternatives, on prompts covering areas such as factuality, math and safety.

1,824items
188subjects
ODC-BYlicense
reward_modelingdomain
textmodality
item-level responses released
Saturation status: No

Response matrix

Loading session strips…

lowhighUnobserved

Scale: {0, 0.25, 0.33, 0.5, 1} (fractions = ties, ~2% of rows)