Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills
Learn how the framework builds transferable skills from raw traces and why it beats prior skill‑editing methods.
The authors aim to replace hand‑written agent skills and brittle LLM‑generated drafts with a system that learns reusable procedures from observed behavior. Their goal is a skill directory that can be applied to larger models, different model families, and out‑of‑distribution tasks.
Trace2Skill collects many execution traces in parallel, then applies inductive reasoning to infer the underlying procedural logic. It treats each trace as a sequence of actions and outcomes, extracts common patterns, and abstracts them into a skill definition that captures the essential steps while discarding noise.
The framework can both augment existing human‑authored skills and bootstrap new ones from weak drafts produced by an LLM. By consolidating these patterns, it creates standard operating procedures that replace ad‑hoc workarounds and failures observed during execution.
In experiments the derived skills improve performance dramatically: on the Wik‑iTableQuestions benchmark, skills distilled from a 35‑billion‑parameter Qwen3.5 model raise a 122‑billion‑parameter agent’s score by up to 57.65 points. The authors also report gains across office workflow, math reasoning, and vision QA tasks, and show that the skills compress repeated failures into reusable procedures.
The abstract notes that Trace2Skill outperforms sequential skill editing and retrieval‑based memories, but it leaves open how the approach scales with extremely long or highly stochastic traces and what overhead the induction step adds during skill creation.
TakeawayDistilling parallel execution traces yields portable skills that boost larger agents without extra parameter tuning.