MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction
Researchers have introduced MIND-Skill, a novel framework designed to automate the generation of high-quality, reusable skills for Large Language Model (LLM) powered AI agents. Addressing the challenge where agents struggle with complex, multi-step tasks requiring domain-specific procedural knowledge, MIND-Skill eliminates the need for manual skill curation by human experts. The framework employs a multi-agent system consisting of an induction agent that abstracts generalizable skills from successful problem-solving trajectories, and a deduction agent that reconstructs these trajectories using the induced skills. To ensure robust quality, the system optimizes three textual losses: reconstruction loss, outcome loss, and rubric loss, utilizing TextGrad for joint optimization. Experimental evaluations on AppWorld and BFCL-v3 benchmarks demonstrate that MIND-Skill consistently outperforms existing concurrent skill generation methods. This advancement enables AI agents to build on prior experience more effectively, enhancing their autonomous problem-solving capabilities in real-world scenarios without extensive human intervention.
Wire timeline
MIND-Skill: Quality-Guaranteed Skill Generation via Multi-Agent Induction and Deduction
Researchers have introduced MIND-Skill, a novel framework designed to automate the generation of high-quality, reusable skills for Large Language Model (LLM) powered AI agents. Addressing the challenge where agents struggle with complex, multi-step tasks requiring domain-specific procedural knowledge, MIND-Skill eliminates the need for manual skill curation by human experts. The framework employs a multi-agent system consisting of an induction agent that abstracts generalizable skills from successful problem-solving trajectories, and a deduction agent that reconstructs these trajectories using the induced skills. To ensure robust quality, the system optimizes three textual losses: reconstruction loss, outcome loss, and rubric loss, utilizing TextGrad for joint optimization. Experimental evaluations on AppWorld and BFCL-v3 benchmarks demonstrate that MIND-Skill consistently outperforms existing concurrent skill generation methods. This advancement enables AI agents to build on prior experience more effectively, enhancing their autonomous problem-solving capabilities in real-world scenarios without extensive human intervention.
cs.AI updates on arXiv.org