Title: SkillPyramid: A Hierarchical Skill Consolidation Framework for Self-Evolving Agents

Abstract:

While contemporary AI agents demonstrate the flexibility to call upon various skills to tackle intricate tasks, their capacity for sustained enhancement is severely hampered by the absence of organized methods for building, storing, and transferring these capabilities. Specifically, the lack of a cohesive framework for consolidating skills leads agents to redundantly develop overlapping competencies for different objectives, fail to convert experiences into reusable knowledge, and find it difficult to adapt task-specific abilities to unfamiliar situations. To overcome these challenges, we introduce SkillPyramid, a consolidation framework designed to leverage prior skill experiences to enhance generalization across broader tasks. Utilizing a hierarchical skill structure, SkillPyramid incorporates a self-evolutionary process that allows agents to synthesize, verify, and integrate new skills while performing tasks. Evaluations conducted on ALFWorld, WebShop, and ScienceWorld, utilizing four distinct backbone models, reveal that SkillPyramid boosts average rewards by 38.0% and cuts execution steps by 27.7%. Ultimately, our approach shifts the paradigm of skill management from a static repository to a dynamic, evolving system.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC