OpenSkill: Open-World Self-Evolution for LLM Agents
OpenSkill framework enhances LLM agents' skill transfer in open-world settings without supervision, achieving top automated pass rates.
Key Findings
Methodology
OpenSkill framework operates in three stages: acquiring knowledge from the open world, refining skills using self-built virtual tasks, and deploying them zero-shot on target agents. Each stage avoids target-task supervision.
Key Results
- OpenSkill achieved a 43.6% pass rate on SkillsBench, outperforming the strongest baseline by 8.9%, excelling on both Opus 4.6 and GPT 5.2.
- In SocialMaze and ScienceWorld, OpenSkill achieved 82.7% and 90.0% pass rates, surpassing all automated methods.
- The virtual verifier's proxy tests covered 88.9% of ground-truth test intents, despite not accessing real tests.
Significance
OpenSkill enables skill transfer across models under unsupervised conditions, significantly enhancing LLM agents' adaptability in open-world environments. It addresses the lack of reliable learning loops in existing methods.
Technical Contribution
OpenSkill provides a skill evolution method without target-task supervision through self-built virtual tasks and open-world knowledge acquisition, breaking the limitations of existing methods and offering new engineering possibilities.
Novelty
OpenSkill is the first to achieve open-world self-evolution of LLM agents without supervision, differing from previous methods that rely on supervision signals.
Limitations
- In the manufacturing domain, all automated methods scored 0%, indicating limitations of open-world acquisition.
- The virtual verifier's tests failed to cover deep semantic quality attributes.
Future Work
Future research could explore applying OpenSkill to more complex tasks and improving the virtual verifier to cover a broader range of test intents.
AI Executive Summary
The OpenSkill framework addresses the challenge of LLM agents lacking reliable learning loops in open-world settings. Existing methods often rely on supervision signals, while OpenSkill achieves self-evolution by acquiring knowledge from the open world and building virtual tasks.
OpenSkill operates in three stages: first, it acquires knowledge from the open world; then, it refines skills using self-built virtual tasks; finally, it deploys the refined skills zero-shot on target agents. Experimental results show that this framework achieves the best automated pass rates across multiple benchmarks.
While OpenSkill performs excellently in several domains, its performance in the manufacturing domain needs improvement. Future research can further optimize the virtual verifier to expand its test coverage. Overall, OpenSkill offers new possibilities for applying LLM agents in open-world scenarios.
Deep Analysis
Background
With the advancement of large language models (LLMs), the ability of agents to perform tasks in open-world settings has become a research focus. Traditional methods rely on predefined skills and supervision signals, limiting their adaptability in dynamic environments.
Core Problem
The core problem is how to achieve self-evolution of LLM agents in open-world settings, especially in constructing skills and verification signals without target-task supervision.
Innovation
OpenSkill achieves self-evolution through open-world knowledge acquisition and self-built virtual tasks. Its innovation lies in avoiding target-task supervision, enabling skill transfer across multiple models.
Methodology
- �� Open-world knowledge acquisition: Retrieve task-relevant knowledge from documents, code repositories, etc.
- �� Leakage-free skill evolution: Refine skills using self-built virtual tasks.
- �� Zero-shot target evaluation: Deploy the final skills on target agents.
Experiments
Experiments were conducted on SkillsBench, SocialMaze, and ScienceWorld benchmarks, using Opus 4.6 and GPT 5.2 as target agents. Baselines include Self-Gen, CoT, etc.
Results
OpenSkill achieved a 43.6% pass rate on SkillsBench, significantly outperforming baselines. The virtual verifier covered 88.9% of real test intents, demonstrating its effectiveness.
Applications
OpenSkill can be used in fields requiring dynamic adaptation, such as autonomous driving and intelligent customer service, reducing reliance on human supervision.
Limitations & Outlook
Performance in the manufacturing domain was poor, indicating limitations of open-world acquisition. Future improvements are needed for the virtual verifier's coverage.
Plain Language Accessible to non-experts
Imagine you're in a huge library, looking for a book on how to fix a bicycle. You don't have a teacher to tell you which book is right, so you have to rely on your own judgment. OpenSkill is like a smart assistant that helps you find the relevant books and practice fixing the bike based on the information in those books. Even if no one tells you exactly what to do, you can improve your skills through trial and practice.
ELI14 Explained like you're 14
Imagine you're playing a game where you have to figure out how to win on your own. No one tells you the answers, but you can look up guides and tutorials online. OpenSkill is like a super helper that finds all the relevant information for you, and then you can use that information to try and win. Even if no one tells you exactly what to do, you can get better through practice and trying different things!
Glossary
Open World
An environment without clear boundaries or restrictions, where agents need to explore and learn autonomously.
In OpenSkill, the agent acquires knowledge from the open world.
Self-Evolution
The process by which an agent improves its capabilities through autonomous learning and adaptation without external supervision.
The OpenSkill framework achieves self-evolution of LLM agents.
Virtual Verifier
An agent used to generate proxy tests for evaluating skill quality.
OpenSkill uses a virtual verifier to replace real tests.
Skill Transfer
The ability to apply skills generated by one model to other models.
OpenSkill enables cross-model skill transfer.
Zero-Shot
Applying skills directly without having seen specific task examples.
OpenSkill performs zero-shot evaluation on target agents.
Open Questions Unanswered questions from this research
- 1 How to apply OpenSkill to more complex tasks?
- 2 How can the virtual verifier cover deeper semantic tests?
Applications
Immediate Applications
Intelligent Customer Service
OpenSkill enables customer service systems to automatically adapt to customer needs without predefined answers.
Long-term Vision
Autonomous Driving
OpenSkill can help autonomous driving systems learn and adapt autonomously in dynamic environments, reducing reliance on human supervision.
Abstract
Self-evolving agents requires adaptation after deployment, but existing approaches assume a usable learning loop, such as curated skills, successful trajectories, or verifier signals. Real open-world deployments may provide none of these, offering only a task prompt. In this work, we study open-world self-evolution, where an agent must build both its skills and its own verification signals from scratch, using open-world resources but no target-task supervision. We propose OpenSkill, a framework that bootstraps this loop: it acquires grounded knowledge and verification anchors from documentation, repositories, and the web, synthesizes them into transferable skills, and refines those skills against self-built virtual tasks grounded in the anchors rather than in target answers. The open world thus supplies both the knowledge to be learned and a supervision-independent practice environment, with target-task supervision reserved for final evaluation. Across three benchmarks and two target agents, OpenSkill attains the best automated pass rate while satisfying the no-supervision constraint. Analysis shows its skills transfer across models without model-specific adaptation, and its self-built verifier aligns with ground-truth outcomes despite never accessing them.