Nvidia Corporation conducted a demonstrative experiment showcasing a new level of automation in machine learning. Using its own AI agent Codex and the TAO toolkit, engineers managed to improve the accuracy of the Cosmos 3 Nano model from 54.41% to an impressive 93.35% on Toyota's specialized Woven Traffic Safety dataset in less than 24 hours. This is not just optimization — it is a paradigm shift in the approach to model post-training.
How the experiment was conducted
The base Cosmos 3 Nano, without any adaptation, achieved a result of 54.41% on a four-option test. Developers tasked the Codex agent with analyzing the model and performing fine-tuning using the LoRA (Low-Rank Adaptation) method. The agent did not simply run a script — it independently selected the Cosmos-reason instruction, identified the absence of a frame rate parameter in the dataset annotations, fixed the configuration, loaded the weights, and launched the training container.
The first LoRA run on eight NVIDIA A100 accelerators (80 GB) took only 30 minutes and boosted accuracy to 87.14%. This is already an impressive leap, but Nvidia went further.
Automatic hyperparameter optimization
In the second stage, developers activated TAO AutoML for automatic hyperparameter tuning: learning rate, batch size, LoRA parameters, and other settings. The system conducted 43 parallel trials with Bayesian optimization. The best configuration yielded 93.35% — and this in less than 20 hours of computation on multiple A100 nodes in Oracle Cloud Infrastructure.
The key advantage of the method is resource savings. According to Nvidia's calculations, LoRA required approximately seven times fewer GPU hours than full model fine-tuning. This is critically important for companies looking to adapt large models without renting entire clusters.
What this means for the industry
It is important to understand the context: the 93.35% figure applies exclusively to the validation portion of the research Woven Traffic Safety dataset. This is not a test for autonomous driving safety nor confirmation of the model's ability to make real-time decisions. However, the principle itself — using an AI agent for autonomous fine-tuning — opens new horizons.
My analysis: we are witnessing a transition from manual model tuning to agent-oriented pipelines. In this experiment, Codex did not just execute commands — it analyzed data, fixed configuration errors, and selected optimal strategies. Although its autonomy is limited to pre-prepared instructions, this is a first step toward fully automated ML development cycles. In the coming years, such agents will become the standard for post-training, especially in niche domains requiring rapid adaptation of base models to specific datasets.