NVIDIA TAO boosts Cosmos 3 Nano video Q&A accuracy from 54.41% to 93.35%

By NVIDIA · July 28, 2026.

NVIDIA has published educational content on how to post-train NVIDIA Cosmos 3 in a single day using NVIDIA TAO agent skills. According to the email, using these TAO agent skills together with two natural-language prompts, it is possible to raise NVIDIA Cosmos 3 Nano's video Q&A accuracy from 54.41% to 93.35% on the Woven Traffic Safety (WTS) dataset.

The email links to two main resources: a blog post explaining the process ('Learn How To Post-Train NVIDIA Cosmos 3 in a Day') and a step-by-step video tutorial showing how to install the TAO skills, run post-training with LoRA, optimize the results with AutoML, and review the model's final performance.

NVIDIA also announces a series of livestreams called 'Cosmos Labs', in which company experts will show live how to use natural-language prompts combined with AutoML to significantly improve Cosmos 3 Nano's video Q&A accuracy on the same WTS dataset. An 'office hour' session is also scheduled for July 30, dedicated specifically to explaining how to post-train Cosmos 3 in a day.

NVIDIA communications of this kind are generally part of its technical outreach strategy for its AI ecosystem (Cosmos, TAO, and tools related to world models and agents for robotics and autonomous vehicles), although the email itself gives no details beyond the figures and links mentioned.