In a technical experiment, Nvidia increased the accuracy of the Cosmos 3 Nano model on a multiple-choice test from 54.41% to 93.35% in less than a day. The main retraining stages were performed by the AI agent Codex using pre-prepared instructions from TAO.
The result was achieved using a specialized dataset—Woven Traffic Safety by Woven by Toyota. It contains videos of road situations and multiple-choice questions. Over 8,000 examples were used for training and testing.
Without additional adaptation, Cosmos 3 Nano correctly answered 54.41% of the questions. Developers then tasked Codex with evaluating the base model and conducting retraining using the LoRA method.
The agent selected a specialized instruction, Cosmos-reason, checked the dataset annotations, and identified a missing frame rate parameter. After correcting the configuration, it uploaded the model weights, prepared the settings, and launched the training container.
LoRA does not alter all model parameters but trains small additional adapters. According to Nvidia, this method required approximately seven times fewer GPU hours than full retraining of Cosmos 3 Nano.
One LoRA run took about 30 minutes on eight NVIDIA A100 accelerators with 80 GB of memory. Accuracy increased to 87.14%.
In a second prompt, developers launched TAO AutoML to optimize learning rate, batch size, LoRA parameters, and other settings. The system conducted 43 parallel trials. The best result with Bayesian optimization reached 93.35%. This stage took 19.5 hours on multiple A100 nodes in Oracle Cloud Infrastructure.
In the experiment, Codex went beyond analyzing results. It selected workflows, verified data, corrected configurations, launched containers, and compared metrics. The agent’s autonomy was limited, acting according to pre-prepared instructions.
The 93.35% figure reflects the accuracy of answers in the test portion of a single research dataset. It does not measure autonomous driving safety or confirm the model’s ability to make real-time decisions.
In June, Nvidia researchers, along with Carnegie Mellon University and the University of California, Berkeley, introduced ENPIRE—a framework that allows AI agents for programming to improve robot control policies on real equipment.
