SkillForge Co-Evolves LLM Agents and Skill Libraries via Trial-Active-Stable-Retired Lifecycle

Chinese Academy of Sciences researchers introduce a fitness-driven skill lifecycle — trial, active, stable, retired — that prunes obsolete entries during agent RL. SkillForge hits 92.4% on ALFWorld and 78.4% on WebShop, topping SkillRL and surpassing GPT-4o and Gemini-2.5-Pro baselines, with 57.2% of retired skills previously stable. Pre-retirement from base-model rollouts drives the biggest ablation gain.