OpenSkill: Open-World Self-Evolution for LLM Agents

OpenSkill introduces a framework enabling LLM agents to self-evolve in open-world environments without task-specific supervision, bootstrapping both skills and verification signals from public documentation and web resources. The approach demonstrates superior performance across benchmarks while maintaining transferability across different models, addressing a critical gap in autonomous agent deployment.

OpenSkill addresses a fundamental challenge in deploying autonomous LLM agents: the gap between controlled laboratory settings and real-world environments where curated training data, successful trajectories, and verification signals don't exist. Traditional self-improving agent frameworks rely on structured feedback loops—either human-provided or from specialized verifiers—assumptions that rarely hold in genuine open-world deployments. This research reframes the problem by treating the open world itself as both knowledge source and practice environment.

The framework's innovation lies in its bootstrapping approach: agents extract grounded knowledge from documentation, code repositories, and web content to build initial skills, then create self-supervising practice tasks anchored to this knowledge rather than target answers. This decoupling from target-task supervision during training while preserving it for final evaluation represents a meaningful shift in how autonomous systems can improve post-deployment.

For the AI industry, this work has significant implications for practical agent deployment at scale. Current production systems struggle with adaptation after launch; OpenSkill's supervision-independent approach could enable more robust, self-improving autonomous systems without requiring expensive human annotation or domain-specific infrastructure. The finding that synthesized skills transfer across different model architectures without model-specific retraining suggests genuine capability abstraction rather than overfitting to particular architectures.

The research trajectory points toward agents that become more capable over time through environmental interaction alone. Success here could accelerate autonomous agent adoption in domains where human oversight is impractical or expensive, though questions remain about safety mechanisms and failure mode detection in fully autonomous refinement loops.

• →OpenSkill enables LLM agents to self-evolve without any task-specific supervision, using only open-world resources for knowledge and practice.
• →The framework achieves best-in-class automated pass rates across three benchmarks while maintaining model-agnostic skill transferability.
• →Self-built verification signals from the system align with ground-truth outcomes despite never accessing target answers during training.
• →Open-world knowledge sources serve dual purpose: providing both learnable skills and supervision-independent practice environments.
• →Framework addresses critical deployment gap between controlled training and real environments lacking curated feedback infrastructure.