Rubrics from the Task, Verdicts from Evidence: Pointwise Rubric Rewards with Tool-Augmented Judges
ReVeR builds rubrics from the instruction itself, then judges each answer independently with tool help.
The paper says existing rubric reward models can inherit bias when their criteria are derived from candidate responses. ReVeR separates hard rules from graded principles, uses label-blind distillation to train rubric reasoning and judging traces, and reports stronger average preference accuracy than pairwise rubric RMs. The gains are clearest on instruction following, but the authors also find that preference accuracy, rubric fidelity, and downstream policy value do not move together cleanly. ArXiv · AI/CL/LG's note
The paper says existing rubric reward models can inherit bias when their criteria are derived from candidate responses. ReVeR separates hard rules from graded principles, uses label-blind distillation to train rubric reasoning and judging traces, and reports stronger average preference accuracy than pairwise rubric RMs. The gains are clearest on instruction following, but the authors also find that preference accuracy, rubric fidelity, and downstream policy value do not move together cleanly. ArXiv · AI/CL/LG's note
score 5