OpenSkill: Open-World Self-Evolution for LLM Agents

OpenSkill: Open-World Self-Evolution for LLM Agents, from Lehigh University, the University of Illinois Chicago, Salesforce AI Research, the University of British Columbia, Vector Institute, and Harvard Medical School, reframes agent improvement as a no-supervision problem: starting from only a task prompt, a base model, and open-world resources, the agent must invent both reusable skills and its own verification signal without seeing ground-truth tests. The framework bootstraps this loop in three stages: open-world retrieval from documentation, repositories, and the web; leakage-free skill evolution against self-built virtual tasks anchored to independently verifiable facts; and zero-shot deployment of the frozen skill artifact to the target agent. On SkillsBench, SocialMaze, and ScienceWorld, OpenSkill is the best automated method across all reported target-agent settings, including 43.6% overall pass rate on Opus 4.6 and 42.1% on GPT 5.2 on SkillsBench, exceeding the strongest closed-world baseline by 8.9 and 8.8 points, respectively. Its virtual verifier is nontrivial despite never observing hidden tests: on 84 analyzed cases it reaches 56.9% precision, 80.5% recall, 60.7% agreement, and covers 88.9% of ground-truth test intents. Transfer experiments show the same skills, produced by Opus 4.6, generalize to weaker models without adaptation, while ablations show open-world retrieval and the virtual verifier each contribute about 6 percentage points and are largely complementary; too many refinement rounds overfit, with performance peaking at 3 iterations.