Prime Intellect and NVIDIA lift Nemotron 3.5 Super VL to 57.7% on AutomationBench-AA
Original titlePost-Training Nemotron 3.5 Super VL for Long-Horizon Enterprise Work
AISummary
Prime Intellect and NVIDIA post-trained Nemotron 3.5 Super VL with GRPO on 2,500 verified enterprise tasks, raising its AutomationBench-AA public-split score from 34.0% to 57.7%. Training took about 11 hours on 128 NVIDIA GB200 GPUs, with step 100 as the best checkpoint. Scoring gives partial credit on objectives but zeroes a task if any guardrail is violated.
Source: Prime Intellect Blog · primeintellect.aiPublished · added here