Source-linked AI summary
DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
Jiangwang Chen, Zixin Song, Junlin Liu, Shuaiyu Zhou, Haiyan Wu, Haihan Shi, Chenxi Zhou, Hanqing Li, Xiao Yang, Da Zhu, Guanjun Jiang, Hai Wan, Xibin Zhao
TL;DR
Fixed evaluation can leave omitted quality dimensions invisible and make rubric evolution appear successful merely because the rubric becomes easier to satisfy. DecoEvo co-evolves solver and rubric-generator skills with separate objectives, and achieves the highest mean across all 15 backbone–benchmark settings, with 2.8–5.0% average gains over SkillOpt.
Problem
Fixed rubrics can limit feedback on multidimensional open-ended tasks, while aggregate scores cannot separate solver improvement from easier rubric changes.
Method
DecoEvo co-evolves persistent solver and rubric-generator skills using criterion-level feedback plus structural and rubric-blind audits decoupled from aggregate solver scores.
Results
DecoEvo achieves the highest mean in all 15 backbone–benchmark combinations, with 5.0% average improvement over SkillOpt for GPT-4o.
Takeaways & Limitations
The results support score decoupling as a practical design principle for evolving evaluation artifacts alongside black-box LLM skills.
Takeaways & Limitations
The evidence is limited to settings where optimization roles share a backbone, with validation sets used for model selection rather than unbiased evaluation.
Abstract
from arXiv · showhide
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.
1. Introduction
DecoEvo addresses score-coupled co-adaptation in text-space optimization by separately evolving solver and rubric-generator skills under decoupled objectives. It uses complementary audits to revise rubric generation and outperforms compared methods across 15 backbone–benchmark combinations.
- Text-space optimization edits external natural-language artifacts, enabling black-box adaptation while keeping the optimized state inspectable.
- Fixed evaluation objectives can bottleneck open-ended optimization when initial rubrics omit dimensions of multidimensional task quality.Multiple responses may be valid, making omitted criteria invisible to a fixed rubric.
- Aggregate solver scores cannot distinguish solver improvement from a rubric that is easier to satisfy, creating an identification problem during joint adaptation.
- DecoEvo separately updates persistent solver and rubric-generator skills, using criterion-level feedback plus structural and rubric-blind near-tie audits.Structural audits identify omitted requirements, while near-tie comparisons expose missed quality distinctions for reusable generator-skill revisions.
- 2.8–5.0% average relative improvements over SkillOpt accompany the highest mean score in all 15 backbone–benchmark combinations.
2. Related Work
Prior work spans text-space optimization of external language artifacts, rubric-based and adaptive LLM judging, and co-evolution under imperfect proxy objectives. DecoEvo is positioned at their intersection, addressing risks from fixed evaluation and solver-score-driven rubric updates.
- Text-Space Optimization: Text-space optimization adapts frozen LLMs by editing external natural-language artifacts, including prompts, instructions, demonstrations, textual feedback, and reusable skill documents.Examples include OPRO, DSPy, TextGrad, and SkillOpt.
- LLM Judges and Adaptive Rubrics: Rubric-based LLM judges provide criterion-level feedback for open-ended evaluation, while contrastive methods derive discriminative criteria from differences between candidate responses.The related work also notes persistent weaknesses in specialized mathematical and general reasoning benchmarks and introduces OpenRS as pairwise adaptive rubric work.
- Co-Evolution and Proxy Optimization: Imperfect proxy optimization can raise measured reward without improving externally evaluated quality, and co-evolution can become unstable when interacting components adapt to each other.These concerns motivate avoiding generator updates driven solely by the solver’s aggregate score, which might favor easier-to-satisfy rubrics over true task quality.
3. Method
DecoEvo co-evolves editable solver and generator skills with fixed model parameters, using criterion-level feedback for solver updates and independent structural and contrastive audits for generator updates. Gold rubrics remain hidden during optimization, while generator acceptance relies on audit objectives rather than aggregate solver score.
- Two-skill framework: DecoEvo maintains two editable natural-language skills: a solver skill for reusable task strategies and a generator skill for constructing question-specific rubrics.Model parameters remain fixed, and each skill is updated under a distinct objective.
- Optimization assumptions: Gold rubrics are never exposed during optimization; all roles use the same frozen backbone under role-specific prompts and update objectives.Role separation changes prompts and objectives rather than model parameters.
- Score decoupling: Generator updates are score-decoupled because neither proposal nor acceptance uses aggregate solver score as the generator objective.Aggregate score can rise because the solver improves or because the generator produces easier rubrics.
- Solver update: The inner loop updates the solver skill from criterion-level rubric feedback, consolidating selected failures into reusable diagnoses and accepting candidates on paired validation comparisons.Paired questions and cached rubrics isolate solver edits from rubric-generation noise.
- Generator update: The outer loop follows audit–distill–verify: structural audits identify rubric omissions, while near-tie contrastive audits identify distinctions missed among score-similar responses.The contrastive auditor first judges responses without seeing the rubric or scores, then diagnoses rubric blind spots using criterion-level breakdowns.
4. Experiments
Across five benchmarks and three backbones, DecoEvo consistently outperforms the baselines, including on untouched cross-benchmark transfer, while controls and alignment analyses attribute the gains to persistent rubric evolution rather than score coupling, extra compute, or richer priors.
- Overall effectiveness: DecoEvo leads all baselines on all 15 backbone–benchmark pairs, exceeding SkillOpt by 5.0%, 2.9%, and 2.8% on GPT-4o, Qwen3-4B, and Qwen3-8B.These correspond to gains of 3.10, 1.74, and 1.78 points, respectively.
- Cross-benchmark transfer: DecoEvo improves over SkillOpt without transfer-benchmark updates by 1.4–4.5% on LLMEval-Med and 2.7–3.9% on Creative Writing.The transfer results support reusable strategies shaped by broader evaluation feedback rather than source-question corrections.
- Score coupling: Score-coupled co-evolution falls below SkillOpt in 13 of 15 cases and below Zero-shot in seven, showing that aggregate-score acceptance can erase solver-only gains.SkillOpt exceeds Zero-shot in 14 of 15 cases, indicating that solver optimization can help even with a fixed rubric generator.
- Ablations: Removing verification causes the largest ablation loss of 2.4–2.9 points, while removing contrastive and structural audits loses 1.7–2.1 and 1.3–1.7 points.Rubric blindness and near-tie sampling each contribute 1.0–1.5 points.
- Alternative explanations: Task priors improve SkillOpt by only 0.4–0.7 points and budget matching by 0.4–0.6, whereas DecoEvo gains 1.6–3.6 points; Direct Audit remains 0.8–1.6 points below it.At matched token use, Budget-Matched gains 0.4, SC-CoEvo loses 1.4, and DecoEvo gains 3.6 points, indicating that persistent generator evolution matters.
- Rubric alignment: DecoEvo raises F1 over SkillOpt by roughly 12 points on every source benchmark, while SC-CoEvo lowers F1 by 5.3–6.3 points.DecoEvo improves both precision and recall, supporting improved evaluation coverage rather than merely generating more criteria.
5. Conclusion
DecoEvo co-evolves solver and rubric-generator skills while decoupling generator updates from aggregate solver score, achieving the highest mean across all 15 evaluated settings. These results support score decoupling as a practical principle for evolving evaluation artifacts alongside black-box LLM skills.
- DecoEvo co-evolves solver and rubric-generator skills while decoupling generator updates from aggregate solver score.The method evolves both skills without tying generator updates to the solver’s aggregate score.
- 15 settings: DecoEvo achieves the highest mean across five benchmarks and three LLM backbones, including two within-domain cross-benchmark transfers.The reported settings cover the full combination of five benchmarks and three backbones.
- The results support score decoupling as a practical design principle for evolving evaluation artifacts alongside black-box LLM skills.This conclusion is drawn from DecoEvo’s results across the evaluated benchmarks, backbones, and transfers.