Meta-affiliated researchers have introduced Autodata, a new approach that treats an AI system like a data scientist tasked with producing synthetic training and evaluation data. The paper argues that this kind of agentic data creation can improve model performance by making data generation more deliberate and higher quality than traditional synthetic dataset methods.

The work, posted to arXiv, describes a general framework in which an agent is trained to design better data rather than simply produce more of it. The authors say the system can be meta-optimized, meaning the data-generating agent itself is trained to improve over time. In the paper’s framing, that leads to stronger datasets and, in turn, better downstream model results.

The researchers also outline a practical implementation they call Agentic Self-Instruct. While the paper presents the broader method as applicable to multiple domains, the experiments they report focus on computer science research tasks, legal reasoning, and reasoning with mathematical objects. In those tests, the authors say Autodata outperformed conventional synthetic data creation approaches.

A key finding in the paper is that tuning the data-generating agent itself can produce an additional performance boost beyond the gains from the basic framework. The authors describe this as meta-optimizing the data scientist agent, a step that appears to make the generated datasets even more useful for training and evaluation.

The paper positions the method as a way to turn more inference compute into better data. Rather than using extra compute only to scale model outputs, the approach channels that capacity into creating more informative training examples and evaluation sets. The authors suggest that could become an important lever for future AI development as systems grow more capable and more expensive to train.

Synthetic data has become an increasingly important part of AI development, especially when human-labeled examples are scarce, costly, or difficult to produce for specialized tasks. But the quality of synthetic data can vary widely, and low-quality examples can limit model performance. Autodata aims to address that problem by making the data creation process more agent-driven and adaptive.

According to the paper, the idea is not limited to a single benchmark or task type. The authors frame it as a general method that could be applied wherever higher-quality synthetic examples are useful, including both training and evaluation. That breadth may make it relevant to researchers and developers exploring how to improve model reliability without relying entirely on larger datasets or more manual labeling.

The paper was submitted on June 24, 2026, and revised the following day. It is authored by Ilia Kulikov and colleagues, including Jason Weston. As with other arXiv preprints, the work has not yet been peer reviewed.

Even so, the researchers argue that the direction could reshape how AI data is built. By training agents to behave like better data scientists, the method suggests a future in which dataset quality becomes a learned capability rather than a fixed input.