SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

2026-08-07

AI Summary

This work addresses the limitations of existing large language model (LLM) agent skill evolution methods, which lack explicit diagnostic feedback mechanisms and fail to incorporate skill deletion into knowledge integration. The authors propose SkillProx, a novel framework that introduces proximal optimization into textual skill evolution for the first time, establishing a closed-loop process of diagnosis and refinement. In the forward phase, diagnostic signals drive skill editing while rolling back performance-degrading changes; in the backward phase, skills are decomposed into units whose utility is evaluated via a frozen leave-one-out strategy, enabling validation-gated integration, downgrading, or deletion. Experiments demonstrate that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline across multiple backbone LLMs on both in-distribution and out-of-distribution benchmarks.

Abstract

LLM agents increasingly adapt to recurring tasks by accumulating procedural knowledge in skills. These skills are lightweight, reusable textual artifacts that are loaded into the agent's context without weight updates. Recent methods refine skills through iterative task execution, failure diagnosis, and trajectory-guided text-space updates. However, existing frameworks lack explicit diagnosis--outcome feedback and treat deletion as a generic edit operation rather than a dedicated mechanism for consolidating accumulated knowledge. We introduce SkillProx, a proximal-gradient-inspired forward--backward framework that couples closed-loop diagnostic evolution with utility-aware proximal refinement. Motivated by a composite objective balancing task loss and skill complexity, the forward stage re-executes diagnosis-driven edits on the same task batch, rolls back regressions, and feeds measured outcomes into subsequent diagnoses. The backward stage decomposes the resulting skill into auditable knowledge units, estimates their contributions using a frozen leave-one-out utility audit, and applies validation-gated consolidation, demotion, or removal. Experiments on in-distribution and out-of-distribution benchmarks across multiple backbone LLMs show that SkillProx improves average accuracy by 3.0 percentage points over the strongest gradient-based baseline. Component ablations demonstrate the complementary effects of closed-loop diagnosis and proximal refinement.

Problem

• LLM agents
• skill evolution
• diagnostic feedback
• knowledge consolidation
• textual gradient

Innovation

• proximal gradient
• self-evolving skills
• closed-loop diagnosis
• utility-aware refinement
• knowledge consolidation