SkillNet: Create, Evaluate, and Connect AI Skills (arxiv.org)

SkillNet has been introduced as a groundbreaking open infrastructure aimed at creating, evaluating, and organizing AI skills on a large scale. Currently, AI agents often struggle to build on previous knowledge, leading to unnecessary re-discoveries of solutions. SkillNet addresses this gap by offering a unified ontology that categorizes skills from diverse sources, supports intricate relational mapping, and allows multi-dimensional evaluations across criteria such as Safety, Completeness, and Cost-awareness. With a repository of over 200,000 skills and an interactive platform complemented by a Python toolkit, SkillNet enhances efficiency and collaboration among AI systems. The significance of SkillNet for the AI/ML community lies in its ability to facilitate the systematic accumulation and transfer of skills, ultimately transforming transient experiences into long-lasting expertise. Experimental results from testing on platforms like ALFWorld, WebShop, and ScienceWorld reveal substantial performance improvements—boosting average rewards by 40% and reducing execution steps by 30% across various backbone models. By defining skills as evolving, composable assets, SkillNet paves the way for smarter, more capable AI agents that can build upon their predecessors' successes, thereby accelerating innovation in the field.