Automating SKILL.md Generation for Computer-Using Agents via Interaction Trajectory Mining

The paper presents a three-stage pipeline for automating the generation of skill libraries from interaction data of computer-using agents, focusing on segmenting GUI trajectories, clustering them into candidate skills, and training a skill-aware policy. The results indicate that while the mined clusters show high purity (0.95) against InteraSkill Workflows labels, the transfer to downstream tasks is limited, with GRPO only improving skill-step accuracy from 18.5% to 20.5%. This study highlights the challenges in using trajectory mining for reliable cross-domain policy improvement, revealing limitations in the current methodologies for skill detection and representation.