3-6x Fewer Injected Tokens, Still +5.9 Points Over ACE

Notes on MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction (arXiv:2605.08670) — Yixuan Li, Mingshui Cai, Ziyang Xiao, Wanyuan Wang, Yanchen Deng, Bo An · May 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

MIND-Skill turns skill-writing into a closed loop with two roles instead of one static write-up. An induction agent abstracts a successful task trajectory into a skill document; a separate deduction agent then tries to redo the task using nothing but that document, and the gap between the original run and the blind reconstruction becomes the training signal. Three losses score that gap. A reconstruction loss checks whether the reconstruction follows the same procedural steps; an outcome loss checks whether it actually succeeds against the live environment; a rubric loss scores the write-up itself on five axes — ground-truth independence, actionability, transferability, completeness, conciseness. All three are optimized together with TextGrad, natural-language gradient descent over the skill text, with the outcome loss as the primary objective and the other two breaking ties. The setup is tested on AppWorld (90 training tasks held out against 168 normal and 417 harder "challenge" test tasks) and BFCL-v3 (50 training / 150 held-out multi-turn tool-call instances). Baselines span ReAct, in-context learning, a single-pass "skill-extract" ablation of their own induction agent, and two concurrent methods — ACE and trajectory distillation-based Trace2Skill.

Running everything through Qwen3.5-122B to both write and execute skills, the full loop scores 71.4%/55.4% (task/scenario goal completion) on AppWorld-Normal and 51.8%/39.6% on Challenge, and 77.3% on BFCL-v3 — ahead of ACE (65.5/55.4, 51.1/34.5, 74.0) and Trace2Skill (67.3/55.4, 46.8/33.1, 72.7) on every metric but one tie. The three losses aren't redundant: pulling the rubric loss costs the most on the easier Normal split (71.4% down to 64.3%), while pulling the reconstruction loss costs the most on Challenge (51.8% down to 45.8%) — one keeps documents from turning into unusable prose, the other keeps them from going technically vague. The resulting skills also stay small: MIND-Skill injects 3–6x fewer tokens per retrieved skill than ACE's single evolving playbook or Trace2Skill's skill directory, for equal or better completion. And skills authored by a weaker model close most of the gap to skills authored by a stronger one once both go through the same optimization loop — the loop compensates for part of the author model's own capability shortfall.