Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Trace2Skill is a novel framework designed to automatically create and adapt domain-specific skills for Large Language Model (LLM) agents, addressing the scalability issues of manual authoring and limitations of sequential skill generation. It operates by deploying a parallel fleet of sub-agents to analyze diverse execution trajectories, extracting lessons, and then hierarchically consolidating them into a unified, conflict-free skill directory through inductive reasoning. This approach supports both refining existing human-written skills and generating new ones from scratch. Experiments across challenging domains like spreadsheet, VisionQA, and math reasoning demonstrate that Trace2Skill significantly improves performance over strong baselines, with skills evolved by Qwen3.5-35B enhancing a Qwen3.5-122B agent by up to 57.65 absolute percentage points on WikiTableQuestions. The framework ensures skill transferability across LLM scales and generalization to out-of-distribution tasks, requiring no parameter updates or external retrieval modules, and utilizing open-source models as small as 35B parameters.

Key takeaway

For MLOps Engineers and AI Scientists deploying LLM agents in complex, domain-specific environments, you should consider Trace2Skill's approach to skill evolution. Its parallel analysis and inductive consolidation of agent trajectories offer a robust method to create or deepen skills, outperforming sequential updates and retrieval-based systems. This framework enables highly transferable, declarative skills across LLM scales and out-of-distribution tasks, reducing dependency on manual authoring or proprietary models, and improving agent performance without parameter updates.

Key insights

Trace2Skill distills broad agent experience into transferable, declarative skills via parallel analysis and inductive consolidation.

Principles

• Holistic analysis of diverse trajectories yields generalizable skills.
• Parallel processing of experience prevents sequential drift and improves efficiency.
• Declarative skills offer better transferability than episodic retrieval.

Method

Trace2Skill generates trajectories, dispatches parallel sub-agents for patch proposals from successes/failures, then hierarchically consolidates these patches via inductive reasoning and programmatic conflict prevention into a unified skill.

In practice

• Deepen existing human-written skills for specific LLM agents.
• Create new, effective skills from scratch using parametric knowledge.
• Improve agent performance in spreadsheet, math, and VQA tasks.

Topics

• LLM Agents
• Skill Evolution
• Automated Skill Generation
• Trajectory Analysis
• Inductive Reasoning
• Qwen3.5

Best for: Research Scientist, AI Architect, NLP Engineer, AI Scientist, AI Engineer, MLOps Engineer

Related on AIssential

• Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents — Dynamically synthesizing task-specific skills from retrieved trajectories improves LLM agent performance without parameter updates.
• Don't Retrieve, Navigate: Distilling Enterprise Knowledge into Navigable Agent Skills for QA and RAG — Corpus2Skill distills document corpora into navigable hierarchical skills, enabling LLM agents to actively explore and combine evidence.
• 🥇Top AI Papers of the Week — Internalizing parallel reasoning and deliberation as a learned skill significantly boosts agentic model performance.
• TRACE: Trajectory Reasoning through Adaptive Cross-Step Evidence Aggregation for LLM Agents — Adaptive cross-step evidence aggregation substantially improves detection of temporally distributed sabotage.
• Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis — RL expands LLM agent capabilities for compositional tool use, unlike its role in static reasoning.
• Agent Skills — Part 2: Bridging Skills with Production Tool Ecosystems — SkillRegistry decouples tool discovery from execution, enabling scalable agent skill management with existing production tools.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.