MIND-Skill (arXiv 2605.08670): Quality-Guaranteed Skill Generation via Induction Agent + Deduction Agent — Reconstruction Loss Validates Learned Skills

arXiv·medium signal

MIND-Skill (submitted May 9, NTU/Waseda/Zhejiang) introduces a two-agent framework for automatically inducing reusable skills from successful agent trajectories. An induction agent abstracts skills, then a deduction agent reconstructs trajectories from those skills. Quality is enforced via three losses: reconstruction (input vs. reconstructed trajectory), outcome (correctness), and rubric (documentation quality). Outperforms concurrent methods on AppWorld and BFCL-v3 benchmarks.