Rep2Skill: Representation-Guided Skill Self-Evolution for LLM Agents
Rep2Skill revises LLM agent skills from internal representation trajectories and beats text-only self-evolution.
Rep2Skill lets LLM agents revise external textual skills by using internal representation trajectories rather than execution text alone. From collected rollouts it localizes turns that deviate from successful execution dynamics and converts those signals, with context, into textual feedback for targeted skill edits. On two agent environments and two open-source LLMs, with the same model serving as executor and optimizer, it consistently outperformed text-only self-evolution.
- Skills stay external; model weights are not updated
- Representation trajectories flag turns that diverge from success
- The same LLM acts as both executor and optimizer
- It beat text-only skill evolution on two environments
From the original article
1 min read at arxiv.orgTextual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes.
By Kaixing Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.