Paper 2607.25675
DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
- Published
- Jul 2026
- Research lab
- Independent
- Citations
- 2
- GitHub
- Not linked
01 In brief
Summary
DecoEvo is a text-space optimization framework that co-evolves a solver skill and a rubric-generator skill under decoupled objectives, avoiding the pitfalls of score-coupled co-adaptation.
The solver skill is updated using criterion-level feedback from generated rubrics, while the generator skill is revised through two audits: a task-conditioned structural audit that identifies omitted requirements, and a near-tie contrastive audit that exposes missed distinctions among score-similar responses.
These audits are independent of aggregate solver score, preventing the rubric from drifting toward the solver's current strengths.
The generator updates are accepted via a Pareto-style rule on held-out audit objectives.
Across five benchmarks (HealthBench, WritingBench, ResearchQA, and transfers to LLMEval-Med and EQ-Bench Creative Writing) and three backbones (GPT-4o, Qwen3-4B, Qwen3-8B), DecoEvo achieves the highest mean score in all 15 backbone-benchmark combinations, with relative gains of 2.8–5.0% over SkillOpt.
Ablations show both audits, rubric blindness, near-tie sampling, verification, and distillation are important.
Controls rule out task priors, extra compute, and direct audit feedback as sole explanations.
Rubric alignment analysis shows DecoEvo improves criterion-level F1 by ~12 points over SkillOpt, while SC-CoEvo degrades it.
02 From the paper
Abstract
Text-space optimization adapts large language models (LLMs) by editing external natural-language artifacts rather than model weights, so the optimized artifacts remain inspectable and the model can be treated as a black box. However, most existing text-space methods keep evaluation fixed. On open-ended tasks, this can become a bottleneck: once the solver improves on the criteria a rubric measures, omitted dimensions remain invisible to the optimization signal. Simply evolving the rubric is also unreliable when updates are selected by the current solver's score, because apparent progress can come from making the rubric easier to satisfy. We introduce DecoEvo (Decoupled Co-Evolution), which co-evolves a solver skill and a rubric-generator skill under decoupled objectives without using gold rubrics during optimization. The solver skill is updated using criterion-level feedback, while the rubric-generator skill is revised through complementary audits of requirement coverage and response discrimination that are independent of aggregate solver score. This separation focuses generator updates on newly exposed solver weaknesses, reducing repeated emphasis on criteria the solver already satisfies. Under each benchmark's official evaluation, DecoEvo outperforms all compared methods across five benchmarks and three LLM backbones, yielding 2.8--5.0\% relative gains over SkillOpt in the five-benchmark average.