OpenSkill: Agents That Self-Evolve in the Open World With Zero Target-Task Supervision

HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

This HuggingFace-trending paper tackles 'open-world self-evolution,' where an agent gets only a task prompt — no curated skills, successful trajectories, or verifier signals. OpenSkill bootstraps both skills and its own verification anchors from documentation, repos, and the web, then refines them against self-built virtual tasks rather than target answers. Across three benchmarks and two target agents it attains the best automated pass rate under the no-supervision constraint, and its self-built verifier aligns with ground truth despite never accessing it — and skills transfer across models without re-adaptation.