SkillSpec: Consensus-Gated Agent Skill Evolution via Representation Specialization

2026-09-30

AI Summary

This study addresses the challenges of structural deficiency and unreliable updates in skill optimization for large language model agents by proposing a two-stage framework. The core innovation lies in decoupling “what to retain” from “how to organize.” Specifically, the method first introduces a pairwise evaluation consensus gating mechanism to rigorously validate the effectiveness of skill updates. Subsequently, it adaptively selects graph-based or hybrid representational structures guided by process sensitivity analysis to specialize knowledge organization. Experimental evaluations across six benchmarks demonstrate that the proposed approach improves average success rates by 6.89% over competitive baselines. Ultimately, this work establishes a novel paradigm for the reliable evolution of agent skills, offering a principled solution to enhance both the stability and structural coherence of skill acquisition in autonomous agents.

Abstract

Natural-language skills are textual procedural memories through which large language model (LLM) agents retain reusable task knowledge without updating model weights. Existing methods typically treat skills as either static artifacts or monolithic documents optimized using aggregate validation scores as feedback. However, representing a skill as a monolithic document restricts optimization to its textual content, without explicitly modeling the structure through which procedural knowledge is retrieved and executed. We identify a key distinction between learning what knowledge to retain and determining how to organize it: textual updates should first be validated through execution evidence, after which the retained knowledge should be structured according to its procedural dependencies and retrieval requirements. To this end, we introduce SkillSpec, a two-phase framework comprising consensus-gated evolution and representation specialization. In the consensus-gated phase, complementary editing intents generate complete candidate skills. An update is committed only when paired evaluations reach consensus, requiring sufficient overall improvement and non-negative aggregate paired gain in every repeated evaluation. In the specialization phase, signals of process and redundancy sensitivity derived from the full optimization trajectory, including accepted and rejected candidates, guide the selection of a flat, graph, or hybrid representation.Across six benchmarks and three target language models, SkillSpec improves average success rate over SkillOpt by 6.89%, averaged across the three models. These results demonstrate that reliable skill evolution and representation specialization address complementary objectives: deciding what knowledge to retain and how to structure it for inference.

Problem

• LLM agents
• skill evolution
• procedural knowledge
• representation specialization
• natural-language skills

Innovation

• Consensus-Gated Evolution
• Representation Specialization
• Skill Evolution
• Procedural Knowledge
• LLM Agents