SkillForge Evolves Agent Skill Libraries via Lifecycle Management, Gaining Up to 7.8% Success Rate

ict-cas · hf · 2026-10-08

ICT-CAS proposes SkillForge, an agentic RL method that co-evolves skill libraries and models through a fitness-driven skill lifecycle.

Problem: in memory-augmented RL, retaining every skill as the policy improves lets obsolete or harmful entries accumulate and mislead agents.
Method: skills cycle through trial, active, stable, and retired states. A pre-RL phase uses the base model's own rollouts to pre-retire low-fitness skills, seeding SFT; during RL, selective retirement, stabilization, and LLM-guided mutation keep forging the library alongside policy optimization.
Results: highest aggregate success rate across multiple interactive agent benchmarks, up to 7.8% relative improvement while keeping the library compact. Ships SkillFurnace, a 5k+ record dataset with retirement-filtered SFT trajectories, fitness-annotated skill libraries, and human-annotated failure categories.

More from coding & agent