Progressive Agent Skill Generation via Reinforcement Learning
Skill-α trains skill writing as a sequence of edits judged by downstream agent performance.
The paper frames skill generation as reinforcement learning because the usefulness of a skill only shows up when an agent tries to use it. Its rollback reward compares execution with the original and edited skill on an anchored query, giving each edit a measurable signal. In experiments, Skill-α beat heuristic and pipeline baselines for both document-to-skill and experience-to-skill settings. With GPT-4o as the worker, it improved average success over the strongest baseline by 3.3 points on CL-Bench and 6.7 points on tau2-bench. HF Daily Papers' note
The paper frames skill generation as reinforcement learning because the usefulness of a skill only shows up when an agent tries to use it. Its rollback reward compares execution with the original and edited skill on an anchored query, giving each edit a measurable signal. In experiments, Skill-α beat heuristic and pipeline baselines for both document-to-skill and experience-to-skill settings. With GPT-4o as the worker, it improved average success over the strongest baseline by 3.3 points on CL-Bench and 6.7 points on tau2-bench. HF Daily Papers' note
score 4