DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space

Researchers introduced DecoEvo, a decoupled co-evolution method for text-space optimization in LLMs.

It updates solver skills using criterion-level feedback and improves rubric generation through coverage and discrimination audits, without gold rubrics during optimization.

Across five benchmarks and three LLM backbones, DecoEvo outperformed prior methods and delivered 2.8–5.0% average relative gains over SkillOpt.

The approach aims to boost open-ended task performance without changing model weights, while keeping rubrics from becoming too easy or missing new weaknesses.