XNews

NVIDIA Skill2Env: Collective Skills to Train RL Agents

NVIDIA Skill2Env: Collective Skills to Train RL Agents

NVIDIA researchers converted 3,400 public agent skills into 7,971 terminal tasks with programmatic tests and behavior rubrics. Just 300 steps of RL training on Qwen3.8-27B improved Terminal-Bench 2.1 by 4.7 points and aligned agent behavior with skill methodologies. The dataset costs $90,000 to generate using GPT-5.6 Sol.

Source posts