DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
DecoEvo separates solver improvement from rubric evolution so the evaluator does not get rewarded for becoming easier.
The paper frames text-space optimization as editing inspectable natural-language skills while keeping the model itself as a black box. Its method co-evolves a solver skill and a rubric-generator skill, but updates the rubric generator through coverage and discrimination audits rather than the solver’s aggregate score. That is meant to surface weaknesses the solver has not already learned to satisfy. Under official benchmark evaluations, the authors report DecoEvo beating compared methods across five benchmarks and three LLM backbones, with 2.8-5.0% relative gains over SkillOpt on the five-benchmark average. HF Daily Papers' note
The paper frames text-space optimization as editing inspectable natural-language skills while keeping the model itself as a black box. Its method co-evolves a solver skill and a rubric-generator skill, but updates the rubric generator through coverage and discrimination audits rather than the solver’s aggregate score. That is meant to surface weaknesses the solver has not already learned to satisfy. Under official benchmark evaluations, the authors report DecoEvo beating compared methods across five benchmarks and three LLM backbones, with 2.8-5.0% relative gains over SkillOpt on the five-benchmark average. HF Daily Papers' note
score 4