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

Imagine you're building an AI that needs to learn and adapt over time, like a personal assistant that gets better at managing tasks as it interacts with users. The challenge lies in how these agents accumulate knowledge and improve their skills without constant retraining. Current methods often fail because they lack a structured way to diagnose failures and refine skills effectively, leading to issues like losing context or not adapting to new information — this is what's called a lack of explicit diagnosis. To address this, the authors propose SkillProx, a framework that combines iterative task execution with a focus on diagnosing failures and refining skills based on outcomes. The approach involves a forward stage that re-evaluates and adjusts skills based on feedback, and a backward stage that breaks down skills into manageable units for better auditing and consolidation. This dual-stage process allows the agent to learn from its mistakes and improve its performance over time. Compared to previous methods, SkillProx not only enhances the accuracy of LLMs by 3 percentage points but also provides a more structured way to manage and refine skills, making it a valuable tool for anyone looking to build more reliable and adaptive AI systems.

SkillProx introduces a novel framework for skill refinement in LLMs.

The experiments show solid improvements over existing methods with clear metrics.