Qwen Team Introduces DecoEvo: Co-Evolving Solvers and Rubrics in Text Space
QwenBusinessUnit · hf · 2026-07-30
The Qwen team introduced DecoEvo (Decoupled Co-Evolution), a novel method that co-evolves a solver skill and a rubric-generator skill within text space.
- The Problem: Existing text-space optimization methods typically keep evaluation criteria fixed, which bottlenecks open-ended tasks. Simply evolving the rubric can lead to false progress if updates are selected by the current solver's score, making the rubric easier to satisfy.
- The Solution: DecoEvo co-evolves both skills under decoupled objectives without using gold rubrics. The solver is updated using criterion-level feedback, while the rubric-generator is revised via complementary audits (requirement coverage and response discrimination) independent of the aggregate solver score.
- Results: Across 5 benchmarks and 3 LLM backbones, DecoEvo outperforms all compared methods, yielding 2.8% to 5.0% relative gains over SkillOpt in the five-benchmark average.
More from Research
- SkillRise: A Unified RL Framework for Cross-Task Skill Evolution in LLM Agents — Zhiyuan Yao · 2026-07-30
- ByteDance's CoRT: Token-Level Credit Assignment for GRPO Optimization — ByteDance · 2026-07-30
- GuidedRAG: Semantic Steering Improves Retrieval Precision and Slashes RAG Overhead — _reachsumit · 2026-07-30
- Yandex Study: ID Embeddings Outperform GNN Embeddings in Large-Scale Recommenders — _reachsumit · 2026-07-30
- UCSB & LinkedIn Research: Agents Can Speculate Their Own Tool Calls — dair_ai · 2026-07-30
- evalstats: Open-Source Tool for LLM Judge Bias-Corrected Stats Tests — IanArawjo · 2026-07-30