Textual skills enable agents built on large language models (LLMs) to accumulate reusable procedural knowledge without altering model parameters. Existing skill evolution is largely confined to the textual level: an optimizer can only diagnose success or failure from full execution trajectories and sparse task outcomes, then revise skills accordingly. This text‑only approach neglects the agent's internal representations, which capture fine‑grained execution states. We introduce Rep2Skill, a framework that leverages internal representations for skill self‑evolution. Concretely, given collected agent rollouts, Rep2Skill models the sequence of internal representations, pinpoints turns that deviate from successful execution dynamics, and translates these signals together with execution context into actionable textual feedback for targeted skill revision. Experiments on two agent environments and two open‑source LLMs show that, in the self‑evolution setting, Rep2Skill consistently outperforms text‑only baselines, with the same LLM acting as both executor and optimizer and no stronger external model required. This work demonstrates a promising direction for agent self‑improvement beyond pure textual reflection.
Review: Rep2Skill successfully incorporates internal state information into the skill refinement loop, enhancing the depth and efficiency of autonomous optimization and offering a practical path for open‑source models to improve themselves.