Ashok Vardhan Makkuva
 |
Associate Professor of Mathematical Data Science Looking for strongly motivated students for exciting projects on Reasoning and Interpretable AI! |
About Me
I am an Associate Professor of Mathematical Data Science in the Department of Computer Science, Data and Artificial Intelligence at Télécom Paris, Institut Polytechnique de Paris (IP Paris). I'm also a frequent visiting researcher at Stanford-ISL and Berkeley-BASICS. Prior to Télécom, I was a postdoctoral researcher at EPFL, hosted by Michael Gastpar, and in close collaboration with Martin Jaggi and Caglar Gulcehre. Before that, I got my Ph.D. in Electrical and Computer Engineering from UIUC, where I worked with Pramod Viswanath and Sewoong Oh. During my PhD, I've also had the pleasure of collaborating with Sreeram Kannan, Founder and CEO of EigenLayer. Prior to that, I graduated from IIT Bombay with a B. Tech. (Honors) in Electrical Engineering and Minors in Mathematics, where I worked with Vivek Borkar.
My Research
My research focuses on building efficient and interpretable AI systems, with a particular emphasis on reasoning models. While AI has undoubtedly made remarkable strides across almost every aspect of modern life, often this has come at the expense of enormous computational and monetary costs, and driven by gigantic black-box-like models, whose inner workings we know little of. My research goal is to precisely address these important gaps and to design AI systems that not only achieve state-of-the-art performance, but are also grounded in cost-and-energy-efficient algorithms as well as principled and interpretable models. Along these themes, few representative works are: (i) Efficient AI : TERMINATOR, Fundamental limits of prompt compression, and Efficient OT, and (2) Interpretable AI : Markov to Laplace via Mamba, Two layers is all you need, and Attention with Markov. Please check out my publication page for more details!
News
May 2026: Exicted to share TERMINATOR, our new state-of-the-art method for early-exit in reasoning models to stymie overthinking. Reduces CoT length by 14%–55% and model latency by 2x. Project page.
Apr 2026: Glad to receive the EuroTech visiting researcher award.
Feb 2026: Our work on demonstrating how Mamba learns the optimal Laplacian estimators is accepted for an ICLR Oral.
Nov 2025: Started my new role as an Associate Professor at Télécom Paris.
Sep 2025: Our recent paper showcasing that two layers are enough to represent any $k$-th order induction head will appear at NeurIPS 2025-Spotlight!
Aug 2025: Honored to be an invited speaker at ICTS, Bangalore, presenting a tutorial on recent advances in LLMs' (representation, learning, generalization).
Apr 2025: Honored to receive the prestigious DAAD AInet Fellowship, awarded to outstanding international AI researchers for an exclusive postdoc research visit to top German universities.
Apr 2025: Excited to share that my body of work on Markovian analysis of transformers will be the centerpiece for an upcoming invited article in the IEEE BITS Magazine!
Feb 2025: Attention with Markov is accepted for an ICLR Spotlight (5% out of 11,670 papers).