NVIDIA Codex 用两个提示词一天内将模型准确率从54%提升至93%

NVIDIA 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. Using NVIDIA TAO agent skills, Codex autonomously:

Detected and patched missing video metadata Ran the zero-shot baseline Generated LoRA configurations Launched training and evaluation Ran an AutoML hyperparameter sweep Reported the best model

One LoRA run reached 87.14% after roughly 30 minutes on eight A100 GPUs.

A second prompt launched 43 parallel AutoML trials across multiple A100 nodes, reaching 93.35% after 19.5 hours. NVIDIA says LoRA required roughly seven times fewer GPU-hours than full-parameter training. Agent skills are becoming the interface through which general coding agents operate highly specialized ML infrastructure.