Reproduce “MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction” 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

Yixuan Li, Mingshu Cai, Ziyang Xiao, Wanyuan Wang, Yanchen Deng, Bo An

MIND-Skill introduces a multi-agent induction and deduction framework that automatically generates high-quality, generalizable agent skills from successful execution trajectories. The system employs a closed-loop optimization process, guided by three textual loss functions, to iteratively refine skill documentation and enhance an agent's ability to solve complex tasks, demonstrating up to an 8.1 percentage point improvement over non-optimized skill extraction methods.

Key insights

  • Iterative refinement through a closed-loop optimization process, guided by textual loss functions, is critical for generating high-quality, faithful, and generalizable agent skills.
  • Explicitly validating skill faithfulness by reconstructing source trajectories (via the deduction agent) and assessing documentation quality and abstraction level (via rubric loss) are crucial for robust skill generation.
  • The proposed framework can enable less capable language models to generate skills comparable in quality to those produced by frontier models, mitigating differences in base LLM capabilities.

Summary by alphaXiv.

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