Prime Intellect and NVIDIA raise Nemotron 3.5 Super VL's AutomationBench-AA score to 57.7%

AIPrime Intellect says NVIDIA used tasks validated on its verifiers, renderers, and prime-rl to train Nemotron 3.5 Super VL with GRPO. The post says enterprise agent failures can seed new verified tasks for further post-training and evaluation.

AIPrime Intellect, working with NVIDIA, built 2,500 hard enterprise tasks modeled on AutomationBench. The post says the tasks were tested for shortcuts and answer leakage, and that graders reward valid solutions without accepting incomplete work.

AIPrime Intellect says its AutomationBench tests agents across simulated SaaS applications, from CRMs to inboxes, on finding context, reasoning through policies, and completing multi-turn work. The post says the team raised the score from 34.0% to 57.7% and describes how it did so.

AIPrime Intellect, working with NVIDIA, post-trained Nemotron 3.5 Super VL on thousands of tasks reflecting enterprise workflows in sales, operations, marketing, and finance. The post says this raised its AutomationBench-AA score on the public split by about 70%.