OpenSkill: Open-World Self-Evolution for LLM Agents

2026-06-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

OpenSkill is a novel framework enabling open-world self-evolution for LLM agents, allowing them to build skills and verification signals from scratch using public resources like documentation and web pages, without target-task supervision. It bootstraps a learning loop by acquiring grounded knowledge and verification anchors, synthesizing them into transferable skills, and refining these skills against self-built virtual tasks. Across three benchmarks (SkillsBench, SocialMaze, ScienceWorld) and two target agents (Opus 4.6, GPT 5.2), OpenSkill achieved the best automated pass rate, improving by +8.9 and +8.8 points over baselines. Its skills transfer across models without adaptation, and its self-built verifier aligns with ground-truth outcomes, covering 88.9% of test intents. The framework involves open-world knowledge acquisition, leakage-free skill evolution, and zero-shot target evaluation.

Key takeaway

For AI Engineers developing LLM agents for dynamic, open-ended environments, OpenSkill offers a robust approach to continuous improvement. You should consider integrating open-world knowledge acquisition and self-verification mechanisms to enable agents to adapt post-deployment without relying on costly human-curated skills or target-task supervision. This method yields transferable skills and a reliable practice environment, significantly boosting agent performance and reducing dependency on explicit feedback.

Key insights

OpenSkill enables LLM agents to self-evolve skills and verification signals using open-world data, free from target-task supervision.

Principles

• Acquire knowledge from open-world resources.
• Refine skills against self-built virtual tasks.
• Ensure skills are model-agnostic artifacts.

Method

OpenSkill acquires knowledge and verification anchors from open-world resources, synthesizes them into skills, and iteratively refines these skills using virtual tests in a leakage-free environment.

In practice

• Use public documentation for skill grounding.
• Generate deterministic pytest suites for verification.
• Limit refinement iterations to prevent overfitting.

Topics

• LLM Agents
• Self-Evolution
• Open-World Learning
• Skill Acquisition
• Virtual Verification
• Model Transferability

Code references

• OpenLAIR/OpenSkill
• openai/codex

Best for: Research Scientist, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents — Self-improving LLM agents require regression-aware skill validation to prevent performance degradation.
• Deploy Self-Evolving Agents for Faster, More Secure Research with a Hermes Agent and NVIDIA NemoClaw — Self-evolving AI agents can securely synthesize public and private data for research, learning new skills that persist across deployments.
• Intelligence as Simulation: Why LLM Agents Need World Models — LLM agents need explicit world models to anticipate consequences and act coherently over time, beyond linguistic plausibility.
• AI Trends 2026: OpenClaw Agents, Reasoning LLMs, and More [Sebastian Raschka] - 762 — LLM evolution focuses on post-training reasoning, tool use, and inference scaling to enhance problem-solving and efficiency.
• antoinezambelli / forge — Forge boosts self-hosted LLM tool-calling reliability and performance through guardrails and VRAM-aware context management.
• GUIDE: Guided Updates for In-context Decision Evolution in LLM-Driven Spacecraft Operations — GUIDE enables LLMs to adapt in real-time by evolving natural-language decision rules without weight updates.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.