AISkill2Env reportedly generated 7,971 reinforcement-learning tasks from public Agent Skills
DAIR.AI says NVIDIA’s Skill2Env generated tasks across 13 domains at an API cost above $90,000. It reports benchmark gains after training, though adding a behavioral rubric to the reward produced smaller gains.
TLDR
According to DAIR.AI, NVIDIA’s Skill2Env uses a Codex planner to read SKILL.md instruction bundles and split skills into workflows. A Codex creator then builds executable terminal tasks with programmatic tests and a behavioral rubric drawn from each skill’s quality criteria.
DAIR.AI says roughly 3,400 crawled skills yielded 7,971 tasks across 13 domains. Generating them with GPT-5.6 Sol cost over $90,000 in API usage.
It reports that after 300 steps of outcome-only reinforcement learning, Qwen3.8-27B improved from 49.4% to 54.1% on Terminal-Bench 2.1 and from 33.4% to 37.7% pass@1 on S2EBench, the researchers’ hand-verified, held-out benchmark. Adding the rubric to the reward produced a smaller Terminal-Bench gain, to 50.1%. But given the source SKILL.md, a judge preferred the rubric-trained model’s trajectories over the base model’s on 73.0% of tasks, versus 54.5% for the outcome-only model.
