OpenSkill: Open-World Self-Evolution for LLM Agents

In three linesOpenSkill introduces a framework for open-world self-evolution of LLM agents without target-task supervision. The agent acquires grounded knowledge and verification signals from documentation and web resources, synthesizes transferable skills, and refines them through self-built virtual tasks. Results across 3 benchmarks show best automated pass rates with cross-model skill transfer.

Summary generated by Claude — human-verified