AIAny
Icon for item

DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space

Co-evolves a solver skill and a rubric-generator skill for text-space LLM optimization under decoupled objectives to avoid rubric gaming without using gold rubrics. Solver updates use criterion-level feedback; generator updates use independent audits of requirement coverage and response discrimination.

Introduction

Most text-space LLM optimizers treat evaluation as fixed, which lets improvements focus on what the rubric already measures and hides other failure modes. DecoEvo flips that dynamic by co-evolving a solver and a rubric-generator under decoupled objectives so the rubric adapts to newly exposed weaknesses instead of being gamed by the solver.

Key Findings
  • Decoupled objectives: the solver is updated from criterion-level feedback while the rubric-generator is revised via audits that measure requirement coverage and response discrimination independent of the solver's aggregate score. This separation prevents generator updates from simply making the rubric easier to satisfy.
  • No gold rubrics required during optimization: the method trains generator and solver without relying on gold rubrics, using complementary audits as supervisory signals instead.
  • Empirical gains: across five benchmarks and three LLM backbones, DecoEvo outperforms compared text-space methods, reporting 2.8–5.0% relative gains over a strong prior (SkillOpt) on the five-benchmark average.
Who it's for and tradeoffs

Great fit if you optimize LLM behavior by editing external text artifacts (prompts, skills, rubrics) and need an inspectable, adaptive evaluation that resists Goodharting. Use DecoEvo when gold references are scarce but you can define criterion-level checks and lightweight audits. Look elsewhere if you require end-to-end guarantees from execution tests (e.g., unit-tested code) or if writing initial criterion ops and audits is infeasible—DecoEvo depends on a starting pool of detectable failure modes and audit procedures.

Information

  • Websitearxiv.org
  • OrganizationsTsinghua University, Qwen Business Unit of Alibaba, University of Chinese Academy of Sciences, Peking University
  • AuthorsJiangwang Chen, Zixin Song, Junlin Liu, Shuaiyu Zhou, Haiyan Wu, Haihan Shi, Chenxi Zhou, Hanqing Li, Xiao Yang, Da Zhu
  • Published date2026/07/28

More Items

Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.

Presents a 2.8T-parameter Mixture-of-Experts multimodal model with a 1-million-token context window and 104 billion activated parameters, targeting long-horizon agentic RL, coding, reasoning, and vision. Key innovations include Kimi Delta Attention, Attention Residuals, Stable LatentMoE (16 of 896 experts active per token), ~2.5× scaling efficiency over Kimi K2, and a public weight release.

Empirically studies how transformer-based native multimodal pre-training scales under fixed compute, deriving compute- and data-allocation power laws and an efficiency frontier for model size, token count, and data mixture; evaluates cross-modal transfer and multimodal in-context learning.