DecoEvo introduces a text-space optimization method that co-evolves a solver skill and a rubric-generator skill without using gold rubrics during optimization. The solver is updated with criterion-level feedback, while the rubric generator is revised through separate audits of requirement coverage and response discrimination, rather than aggregate solver scores. This is designed to prevent rubric changes that merely make the task easier. According to the abstract, DecoEvo beats all compared methods under official evaluation across five benchmarks and three LLM backbones, with a 2.8–5.0% relative average gain over SkillOpt.
No heat snapshots are available in the last 24 hours.