DecoEvo Co-Evolves the Solver and Its Own Rubric in Text Space, Auditing the Rubric Independently So It Cannot Just Get Easier

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

This 13-author July 28 paper (led by Jiangwang Chen, 51 upvotes on Hugging Face) improves LLM performance without touching weights by evolving two text artifacts together: a solver that improves from criterion-level feedback, and a rubric generator that evolves via 'complementary audits of requirement coverage and response discrimination that are independent of aggregate solver score.' That independence is the whole design — coupling the rubric to the solver's score is what makes self-improving evaluators drift toward being easier to satisfy rather than more discriminating. Across five benchmarks and three LLM backbones it reports 2.8–5.0% relative gains over SkillOpt. Anyone running a self-improving harness with an LLM-authored quality gate should steal the decoupled-audit idea.