From Experience to Ability: How Agentic Skills Form, Internalize, and Self‑Evolve

Wu Jinyang, a fourth‑year PhD student in the Automation department at Tsinghua University, researches large‑model inference, agentic reinforcement learning, and self‑evolution, with a focus on the formation, internalization, and continual update of Agentic Skills.

The talk titled “From Experience to Ability: The Formation, Internalization and Self‑Evolution of Agentic Skill” outlines three recent research efforts.

ThoughtICR shifts context learning from mimicking concrete examples to reusing abstract reasoning patterns, automatically extracting inference skills that can transfer across tasks.

TemplateRL incorporates these extracted reasoning skills into reinforcement learning, guiding high‑quality trajectory exploration so that explicit policy experience gradually internalizes as model capability.

SEED extracts skills from the trajectories generated by the current policy and converts the resulting behavior improvements into fine‑grained training signals, achieving a synergistic co‑evolution of skill extraction and policy learning.

Collectively, these works illustrate how Agentic Skills can bridge contextual reasoning, reinforcement learning, and autonomous self‑evolution, offering a pathway for agents to move from merely completing tasks to continuously learning from experience.