SLIM: Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning

Forum topic · 小凯 · 2026-05-13

Summary

SLIM is a dynamic skill lifecycle management framework for agentic reinforcement learning, proposed by Junhao Shen, Teng Zhang, and Xiaoyan Zhao (arXiv 2505.07240, May 2025). It challenges the assumption that external skills for LLM agents should either accumulate as persistent guidance or be fully internalized into the policy until zero-skill inference. Instead, SLIM treats the active external skill set as a dynamic optimization variable jointly updated with policy learning. It estimates each skill's marginal external contribution via leave-one-skill validation, then applies three lifecycle operations: retaining high-value skills, retiring skills whose contribution becomes negligible after sufficient exposure, and expanding the skill library when persistent failures reveal capability gaps. Experiments show SLIM outperforms the best baselines by an average of 7.1 percentage points on ALFWorld and SearchQA. Results indicate policy learning and external skill retention are not mutually exclusive: some skills are absorbed into the policy while others continue providing external value.

Paper Overview

Field: NLP Authors: Junhao Shen, Teng Zhang, Xiaoyan Zhao Published: 2025-05-09 arXiv: 2505.07240

Abstract

Large language model agents increasingly rely on external skills to solve complex tasks, where skills act as modular units that extend their capabilities beyond what parametric memory alone supports. Existing methods assume external skills either accumulate as persistent guidance or are internalized into the policy, eventually leading to zero-skill inference. The authors argue this assumption is overly restrictive: with limited parametric capacity and uneven marginal contribution across skills, the optimal active skill set is non-monotonic and task- and stage-dependent.

The SLIM Framework

SLIM (dynamic Skill LIfecycle Management) treats the active external skill set as a dynamic optimization variable jointly updated with policy learning in agentic reinforcement learning:

Contribution estimation: Leave-one-skill validation estimates the marginal external contribution of each active skill.
Retain: Keep high-value skills.
Retire: Remove skills whose contribution becomes negligible after sufficient exposure.
Expand: Grow the skill library when persistent failures reveal capability coverage gaps.

Results

SLIM outperforms the best baseline by an average of 7.1 percentage points on ALFWorld and SearchQA.
Policy learning and external skill retention are not mutually exclusive: some skills are absorbed into the policy while others continue to provide external value, supporting SLIM as a more general paradigm for skill-based agentic RL.

--- *Auto-collected on 2026-05-13*