In a technical experiment, Nvidia increased the accuracy of its Cosmos 3 Nano model on a multiple-choice test from 54.41% to 93.35% in less than a day. The key stages of fine-tuning were carried out by the AI agent Codex, following pre-prepared instructions from TAO.
This result was achieved using a specialized dataset — the Woven Traffic Safety dataset from Woven by Toyota, which includes videos of traffic situations and questions with four answer options. Over 8,000 examples were used for training and validation.
Without additional adaptation, Cosmos 3 Nano answered 54.41% of the questions correctly. Developers then asked Codex to evaluate the base model and perform fine-tuning using the LoRA method.
The agent selected the specialized instruction Cosmos-reason, checked the dataset annotations, and identified a missing parameter for frame rate. After correcting the configuration, it loaded the model weights, prepared the settings, and launched the training container.
https://youtu.be/9AQkVbx3fKA?si=6y1p3qZHRgeqzuEq
LoRA does not alter all model parameters but trains a small set of additional adapters. According to Nvidia, this method required about seven times fewer GPU hours than fully retraining Cosmos 3 Nano.
One run of LoRA took approximately 30 minutes on eight NVIDIA A100 accelerators with 80 GB of memory, resulting in an accuracy of 87.14%.
In the second prompt, developers launched TAO AutoML to optimize learning rate, batch size, LoRA parameters, and other settings. The system conducted 43 parallel trials, achieving the best result of 93.35% through Bayesian optimization. This stage took 19.5 hours across multiple A100 nodes in Oracle Cloud Infrastructure.
In the experiment, Codex was not limited to analyzing results. It selected workflows, checked data, corrected configurations, launched containers, and compared metrics. The agent's autonomy remained limited, as it acted according to pre-prepared instructions.
The 93.35% figure reflects the accuracy of responses in the validation portion of a single research dataset. It does not measure the safety of autonomous driving or confirm the model's ability to make real-time decisions.
As a reminder, in June, researchers from Nvidia, Carnegie Mellon University, and the University of California, Berkeley introduced ENPIRE — a framework that allows AI agents to improve robot control policies on real hardware.