SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles
SkillForge treats an agent’s skill memory as something to prune and revise during training, not just accumulate.
The method moves skills through trial, active, stable, and retired states based on measured fitness. It first filters weak skills before supervised fine-tuning, then keeps retiring, stabilizing, and mutating skills during reinforcement learning. The authors report the best aggregate success rate across multiple interactive agent benchmarks, with up to a 7.8% relative gain over the strongest baseline. They also release SkillFurnace, a 5k+ record dataset covering filtered trajectories, evolved skill libraries, and annotated retirement events. HF Daily Papers' note
The method moves skills through trial, active, stable, and retired states based on measured fitness. It first filters weak skills before supervised fine-tuning, then keeps retiring, stabilizing, and mutating skills during reinforcement learning. The authors report the best aggregate success rate across multiple interactive agent benchmarks, with up to a 7.8% relative gain over the strongest baseline. They also release SkillFurnace, a 5k+ record dataset covering filtered trajectories, evolved skill libraries, and annotated retirement events. HF Daily Papers' note
score 4