OpenRubrics' ReVeR Lifts 4B Rubric RM Over 8B Pairwise Baselines Using Tool-Augmented Per-Criterion Judges

ReVeR, a pointwise rubric reward model distilled from DeepSeek-V4-Flash, derives rubrics from the task alone, then judges each criterion with a sandboxed Python interpreter plus LLM reasoning. A 4B ReVeR beats 8B pairwise rubric RMs on preference accuracy, hitting 83.3% on IFEval, and the paper documents that rubric fidelity, preference accuracy, and downstream policy utility are distinct, often diverging properties.