Nvidia Corporation conducted a technical experiment demonstrating how autonomous AI agents can dramatically accelerate the fine-tuning process of large language models. The experiment involved the Cosmos 3 Nano model, designed for analyzing road scenes. The results are impressive: accuracy on a four-choice test soared from 54.41% to 93.35% — all in less than 24 hours.
The key executor was the Codex AI agent, operating within Nvidia's TAO toolkit. The agent did not just run scripts; it made independent decisions: it selected the specialized Cosmos-reason instruction, verified the annotations of the Woven Traffic Safety dataset from Woven by Toyota, detected a missing frame rate parameter, corrected the configuration, and launched the training container.
Initially, the base model without adaptation showed only 54.41% correct answers. Codex applied the Low-Rank Adaptation (LoRA) method. Unlike full fine-tuning, LoRA trains only small additional adapters without affecting all model parameters. According to Nvidia's calculations, this approach required approximately seven times fewer GPU hours than a full training cycle.
The first LoRA run on eight NVIDIA A100 accelerators with 80 GB of memory took about 30 minutes and raised accuracy to 87.14%. The developers then issued a command to launch TAO AutoML for hyperparameter tuning: learning rate, batch size, and LoRA parameters. The system conducted 43 parallel trials with Bayesian optimization. The best result was 93.35%, and this stage took 19.5 hours on several A100 nodes in Oracle Cloud Infrastructure.
It is important to note that Codex in this experiment was not fully autonomous — it acted according to pre-prepared instructions. However, it independently selected the workflow, verified data, corrected the configuration, and compared metrics. This is already a significant step toward automating complex ML pipelines.
A crucial caveat should be made: the 93.35% figure reflects accuracy on the validation portion of a single research dataset. It does not measure autonomous driving safety nor confirm the model's ability to make real-time decisions. Nevertheless, the very fact that an AI agent could adapt a model to a specific task so quickly and efficiently opens new horizons for accelerating R&D in the field of artificial intelligence.
Expert opinion: This experiment is a vivid demonstration of how AI agents can transform the AI development process itself. If previously fine-tuning a model for a specific dataset required days of manual work by engineers, now this task can be delegated to an autonomous agent. In the long term, this will lead to lower costs and democratized access to fine-tuning large models, which is especially critical for niche and demanding fields such as road safety or medical diagnostics.