SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation
SkillGym turns scraped skills into offline-verifiable training environments for agent finetuning.
The paper says its pipeline filters internet-sourced skills for reproducible offline workflows, then builds difficulty-controlled tasks with reference solutions and executable verifiers. The authors report 6.8k environments and 19k verified successful trajectories for supervised finetuning. Finetuned models improved across four skill-use benchmarks, with a Qwen3.5-9B SFT model beating an untrained 397B model on two of them. The analysis attributes much of the gain to agents learning to read and invoke the relevant skill more often, from 28% to 96%. HF Daily Papers' note
The paper says its pipeline filters internet-sourced skills for reproducible offline workflows, then builds difficulty-controlled tasks with reference solutions and executable verifiers. The authors report 6.8k environments and 19k verified successful trajectories for supervised finetuning. Finetuned models improved across four skill-use benchmarks, with a Qwen3.5-9B SFT model beating an untrained 397B model on two of them. The analysis attributes much of the gain to agents learning to read and invoke the relevant skill more often, from 28% to 96%. HF Daily Papers' note
score 5