SkillMaster: A Framework for Autonomous Skill Mastery in LLM Agents
Researchers have introduced SkillMaster, a novel training framework designed to enable Large Language Model (LLM) agents to autonomously create, refine, and select skills during complex task solving. Unlike existing frameworks that rely on external teachers or hand-designed rules, SkillMaster allows agents to internalize capabilities through experience. The system employs three key mechanisms: trajectory-informed skill review to update skills based on past episodes, counterfactual utility evaluation for skill edits, and DualAdv-GRPO to stabilize joint training of task-solving and skill management. Experimental results on ALFWorld and WebShop benchmarks demonstrate that SkillMaster outperforms state-of-the-art baselines, improving success rates by 8.8% and 9.3% respectively. The study highlights a significant shift in agent capability, showing that trained agents can identify failures, refine procedural knowledge, and transfer improvements to future tasks with minimal edits. This advancement moves LLM agents beyond mere skill usage toward becoming self-improving entities capable of developing their own skill repertoires, marking a significant step forward in artificial intelligence research regarding autonomous agent adaptation and learning efficiency.
Wire timeline
SkillMaster: A Framework for Autonomous Skill Mastery in LLM Agents
Researchers have introduced SkillMaster, a novel training framework designed to enable Large Language Model (LLM) agents to autonomously create, refine, and select skills during complex task solving. Unlike existing frameworks that rely on external teachers or hand-designed rules, SkillMaster allows agents to internalize capabilities through experience. The system employs three key mechanisms: trajectory-informed skill review to update skills based on past episodes, counterfactual utility evaluation for skill edits, and DualAdv-GRPO to stabilize joint training of task-solving and skill management. Experimental results on ALFWorld and WebShop benchmarks demonstrate that SkillMaster outperforms state-of-the-art baselines, improving success rates by 8.8% and 9.3% respectively. The study highlights a significant shift in agent capability, showing that trained agents can identify failures, refine procedural knowledge, and transfer improvements to future tasks with minimal edits. This advancement moves LLM agents beyond mere skill usage toward becoming self-improving entities capable of developing their own skill repertoires, marking a significant step forward in artificial intelligence research regarding autonomous agent adaptation and learning efficiency.
cs.AI updates on arXiv.org