SkillProx: Self-Evolving Agent Skills via Proximal...

SkillProx: Self-Evolving Agent Skills via Proximal... | AI Research

Key Takeaways

SkillProx is a framework designed to improve how LLM agents refine their procedural knowledge.
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.

Paper AbstractExpand

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.

SkillProx is a framework designed to improve how LLM agents refine their procedural knowledge. Instead of treating skill development as a one-time task, it uses a two-stage process to evolve textual skills—reusable instructions loaded into an agent's context—by balancing task performance with the complexity of the skill itself.

The Problem with Current Skill Evolution

LLM agents often store task-solving strategies as "skills." Existing methods typically use "open-loop" evolution, where an agent diagnoses a failure and immediately updates its skill. This approach has two main flaws: it often commits updates without verifying if they actually improve performance, and it allows skills to grow indefinitely, leading to the accumulation of redundant, conflicting, or overly specific instructions that can hinder the agent.

How SkillProx Works

SkillProx uses a "forward-backward" framework inspired by proximal gradient descent to manage skill evolution:

Forward Stage (Diagnostic Evolution): When an agent encounters a task, it proposes a patch to its skill. Unlike previous methods, SkillProx re-executes the updated skill on the same task batch. It only accepts the update if the skill shows improved accuracy. If the update fails, the system rolls it back and uses the failed attempt as feedback for future diagnoses.
Backward Stage (Proximal Refinement): This stage manages the "health" of the skill. The system breaks the skill into smaller, auditable units and performs a "leave-one-out" utility audit to see which parts contribute to success and which are redundant. It then uses a validation-gated process to consolidate, demote, or remove harmful or unnecessary content while ensuring the skill remains effective.

Performance and Results

Researchers Mingxuan Zheng, Yujin Zhou, and their colleagues evaluated SkillProx across multiple backbone LLMs using one in-distribution and two out-of-distribution benchmarks. The framework achieved an average accuracy improvement of 3.0 percentage points over the strongest gradient-based baseline. Component ablations indicated that the combination of closed-loop diagnosis and utility-aware refinement is responsible for these gains, as each component addresses a distinct aspect of skill degradation.

Key Considerations

The framework relies on a specific trade-off between task performance and textual complexity. While it effectively prunes negative-utility knowledge, the process is constrained by a soft compression threshold (set at 10% in the study) to prevent the removal of essential information. Because the forward stage operates on specific task batches, the authors note that performance is not guaranteed to increase monotonically across every iteration, as the system is constantly adapting to new, diverse task environments.