Nvidia Codex Agent Raises Cosmos 3 Nano Accuracy to 93.35% in One Day
Nvidia said its Codex agent increased the accuracy of the Cosmos 3 Nano video reasoning model from 54.41% to 93.35% in a single day using two text prompts. The system autonomously processed more than 8,000 video samples, generated machine-learning configurations, and executed 43 parallel training trials across Nvidia A100 hardware.
The top-performing trial finished in 19.5 hours, requiring roughly seven times fewer GPU hours than full-parameter training, Nvidia said. The work utilized a traffic safety dataset from Toyota to test video reasoning capabilities. Nvidia noted the results demonstrate how coding tools are evolving into the primary interface for managing specialized artificial intelligence infrastructure.
From the sources (1 posts)
@kimmonismusNVIDIA says Codex post-trained Cosmos 3 Nano from 54.41% to 93.35% accuracy in one day - with two prompts. The experiment used Toyota’s Woven Traffic Safety dataset: 8,000+ training and validation samples for four-choice video reasoning. U