Abstract
Text-space optimization adapts large language models (LLMs) by editing external natural-language artifacts rather than model weights, allowing optimized artifacts to remain inspectable while treating the model as a black box. However, most existing text-space methods keep evaluation fixed. This can become a bottleneck in open-ended tasks, where 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, as updates selected by the current solver's score may lead to apparent progress 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.
Blogger's Review: DecoEvo effectively addresses the bottlenecks of fixed evaluation in open-ended tasks by decoupling the updates of solver and rubric-generator skills. This method not only enhances the adaptability of the models but also provides new insights into LLM optimization, making it a worthy area for further exploration.