Data Synthesis

High-quality task data is often difficult to collect at the scale or in the styles needed for AI security applications. This project studies how to generate synthetic examples that can support downstream tasks, post-training, and the training of small models.

I lead a team of four PhD researchers on controllable data synthesis. Our current approach pairs a diffusion language model generator with an encoder-only discriminator. The discriminator provides a signal for steering generated text toward target styles, giving us a way to study control over the output rather than generation alone.

The research focuses on the generation method and on how the resulting data can be used in security-related model development. This is ongoing work.