ReVeR: pointwise rubric rewards with tool-augmented judges beat pairwise rubric RMs
OpenRubrics · hf · 2026-10-14
ReVeR is a pointwise rubric reward model that derives hard rules and graded principles from the instruction alone and judges responses independently via tool-augmented evaluation, trained through label-blind distillation into the ReVeR-Traces dataset. It outperforms pairwise rubric RMs in preference accuracy (especially instruction following), test-time scaling and downstream policy optimization. Notably, generator RL raises preference accuracy while lowering rubric fidelity, and higher preference accuracy doesn't guarantee better downstream policies — three distinct properties.
More from Research
- RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding — Guowen Zhang · 2026-10-14
- JSIR dataset drops: 267 jazz recordings with beat, chord and structure annotations — umpedronosapato · 2026-10-14
- There is no such thing as an AI drug, argues researcher Patrick Kidger — PatrickKidger · 2026-10-14
- More thinking, worse scores: refusals tank closed-model evals at max reasoning effort — _lewtun · 2026-10-14
- Integer multiplication repo becomes an agent harness benchmark — and collaboration loses — iskander · 2026-10-14
- Recurrent self-conditioning: framing recent diffusion methods as depth decoupled from time — alec_helbling · 2026-10-14