Reproduce “Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haochen Luo, Yi Huang, Sichun Luo, Fengyuan Liu, Lei Li, Zefa Hu, Junlan Feng, Qi Liu

The Harness-Aware Self-Evolving (HASE) framework enables a single large language model to co-evolve its policy, task solutions, and operational harness components, including both guidance and evaluation mechanisms. This allows a Qwen3-8B model to match or exceed the performance of larger models on tasks like text classification and alpha factor mining, and autonomously repair incomplete evaluators in geometric algorithm discovery tasks.

Key insights

  • Co-evolving model policy, task solutions, and the operational harness within a unified framework leads to more robust and higher-performing AI systems capable of continuous self-improvement.
  • Self-repair of evaluation components, driven by detection of divergences between local and real-world task objectives, is crucial for preventing reward hacking and aligning AI behavior with true task goals.
  • Intelligent harness co-evolution allows comparatively smaller LLMs (e.g., Qwen3-8B) to achieve performance competitive with or superior to much larger models or complex human-engineered systems, reducing computational overhead.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.