Pranav Ajit Nair

About

Pranav Ajit Nair is a first-year PhD student in the ML Foundations Group advised by Prof. Sham Kakade. Before this, he was a predoctoral researcher at Google DeepMind India, working on efficient LLM inference. He finished his integrated Bachelor’s and Master’s from the Indian Institute of Technology (BHU), Varanasi, during which he interned with Prof. Chris Biemann’s group at the University of Hamburg.

Research Focus

MoEs and test time compute are taking the bulk of the LLM serving load. However, the communication overhead involved with MoE, the on-policy nature of RL training, and the memory boundness of long CoT chains at inference result in substantial underutilization of the hardware. Pranav is interested in designing architectures and algorithms that maximally use the hardware’s compute power, improving both throughput and quality of the generated responses. He is also interested in understanding the training dynamics and routing in MoEs, and developing fast and robust LLM optimizers.