Arani Roy

I am a Ph.D. researcher in Electrical and Computer Engineering at Purdue University, advised by Prof. Kaushik Roy.

My research spans generative AI across multimodal and language models, including diffusion models, diffusion language models, and large language models. I work on representation geometry, model unlearning, efficient reasoning, and efficient deep learning, with a focus on developing training-free and geometry-aware methods for controlling, analyzing, compressing, and selectively modifying foundation models.

My current work spans four closely connected directions:

1. Controllable and geometry-aware generative models — understanding representation geometry for precise, training-free control across diffusion and foundation models.
2. Machine unlearning and model editing — selectively modifying concepts, features, and knowledge in diffusion models and language models while preserving unrelated capabilities.
3. Efficient generative AI and reasoning — compression and acceleration of diffusion models, efficient reasoning in diffusion language models, and model-aware data selection.
4. Efficient learning beyond backpropagation — scalable local-learning and hardware-aware methods for reducing training and inference cost.

A recurring theme across these projects is to understand where knowledge lives, how it is geometrically organized, and how it can be changed efficiently without disrupting unrelated capabilities.