NVIDIA Cosmos 3 post-training with agent skills achieves over 90% accuracy

NVIDIA demonstrates that using autonomous coding AI agents to post-train the Cosmos 3 vision reasoning model can push accuracy above 90% with minimal manual effort.

This approach allows developers to adapt vision reasoning models to production video tasks without spending days on data formatting, container setup, training scripts, baseline evaluation, and hyperparameter sweeps.

By automating these preparatory steps, the method significantly reduces the time required to determine if post-training improves accuracy.