Title: SkillAdaptor: Enabling Self-Adapting Capabilities for LLM Agents via Trajectory Analysis

Abstract

As Large Language Model (LLM) agents grow more dependent on reusable external skills to navigate complex, long-horizon interactive tasks, the need for effective adaptation mechanisms has become critical. Current training-free pipelines typically derive updates from entire trajectories or session-level feedback, a method that often leads to coarse failure attribution and results in revisions that are either unstable or excessively broad. To address these limitations, we introduce SkillAdaptor, a novel framework for step-level skill adaptation that operates without training and features explicit failure attribution. This system is designed to integrate seamlessly into OpenClaw-class agent harnesses.

When presented with a failed trajectory, SkillAdaptor pinpoints the initial actionable step where the fault occurred, assigns responsibility to specific candidate skills, and implements targeted updates. These modifications are strictly governed by explicit acceptance checks, ensuring that the underlying backbone model remains frozen throughout the process. We conducted evaluations using Kimi-K2.5, GLM-5, and GPT-5.2 across three benchmark suites: WebShop, PinchBench, and Claw-Eval. SkillAdaptor outperformed both the no-skill and standard skill-adaptation baselines on all three platforms. Notably, it achieved significant single-metric gains, including a +1.5 point increase in PinchBench Avg Score%, a +1.8 boost in Claw-Eval Avg Score, and a +1.7 improvement in WebShop success rate. These findings suggest that step-level attribution facilitates more stable and auditable maintenance of training-free skills.

\footnote{The code will be released at https://github.com/zjunlp/SkillAdaptor.}

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC