SkillForge Proposes Dynamic Skill Lifecycle to Prevent Agent Performance Degradation
New research SkillForge uses dynamic retirement and mutation of skills to keep LLM agent libraries compact and effective during reinforcement learning, reporting up to 7.8% improvement.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Researchers from the Institute of Computing Technology at the Chinese Academy of Sciences and other institutions have proposed SkillForge, a reinforcement learning method for Large Language Model (LLM) agents. The approach addresses the issue where obsolete or harmful entries accumulate in the skill library as the policy improves, misleading the agent.
Traditional methods often treat the skill library as an append-only collection, leading to increased retrieval noise and context pollution. SkillForge introduces a fitness-driven skill lifecycle with four states: trial, active, stable, and retired. Before training, it pre-retires low-fitness skills using the base model's own rollouts. During reinforcement learning, selective retirement, stabilization, and LLM-guided mutation allow skills and the model to co-evolve.
According to the paper's abstract, SkillForge achieves the highest aggregate success rate across multiple interactive agent benchmarks, delivering up to 7.8% relative improvement over the strongest baseline while keeping the skill library compact throughout training. The study has been accepted at NeurIPS 2026.