Prime Intellect and NVIDIA post-train Nemotron 3.5 Super VL for enterprise agents
Prime Intellect and NVIDIA built a 2,500-task verified enterprise dataset and ran 100 GRPO steps (about 11 hours on 128 NVIDIA GB200 GPUs) to lift NVIDIA Nemotron 3.5 Super VL from 34.0% at its SFT checkpoint to 57.7% on the public split of AutomationBench-AA, a 23.7-point gain. The tasks span 47 simulated SaaS applications such as CRMs, inboxes, help desks and spreadsheets, and scoring inspects only the final application state: partial credit for objectives met, zeroed if any guardrail is violated. The post says every task is attacked with scripted shortcuts (blanket actions, doing nothing, claiming success) before admission, with about 18,000 attempts in the latest batches getting through none, and a second independent grader matching scores byte for byte on roughly 69,000 episodes.
It frames enterprise workflows as an RL problem checkable from final system state, offering a reproducible path to post-train open-weight models on each company's own processes.