Kartik Ahuja

Kartik Ahuja Research Scientist FAIR Google Scholar    Github Email: ahujak@ucla.edu

About

I am a Research Scientist at FAIR (Meta AI). Prior to joining FAIR, I was a postdoctoral fellow at Mila – Quebec AI Institute, where I was hosted by Yoshua Bengio, Irina Rish, and Ioannis Mitliagkas. I earned my Ph.D. in Electrical and Computer Engineering from the University of California, Los Angeles, and a B.Tech–M.Tech Dual Degree in Electrical Engineering from the Indian Institute of Technology, Kanpur. I also spent a year at IBM Research, Thomas J. Watson Research Center in Yorktown Heights, NY, as an AI Resident in the Trustworthy AI team led by Kush Varshney.

My research spans machine learning, causality, and optimization. I am particularly interested in questions around generalization in machine learning, especially generalization beyond the training distribution. In the LLM era, I find it especially intriguing to explore how large language models generalize, the conditions under which they fail to generalize, and to identify the essential ingredients of a new, significantly more sample-efficient training paradigm.

In earlier work, I also focused on optimization methods for resource allocation problems.

Selected Publications and Recent Preprints

• Konstantin Donhauser, Charles Arnal, Mohammad Pezeshki, Vivien Cabannes, David Lopez-Paz, Kartik Ahuja
  Unveiling Simplicities of Attention: Adaptive Long-Context Head Identification.
  Submitted

• Kartik Ahuja, Amin Mansouri
  On Provable Length and Compositional Generalization.
  Submitted

• Divyat Mahajan, Mohammad Pezeshki, Charles Arnal, Kartik Ahuja⁺, Pascal Vincent⁺
  Compositional Risk Minimization.
  Submitted,⁺ Joint last author.

• Kartik Ahuja, Divyat Mahajan, Yixin Wang, Yoshua Bengio
  Interventional Causal Representation Learning.
  ICML 2023 (Oral)

• Kamalika Chaudhuri^*, Kartik Ahuja^*, Martin Arjovsky, David Lopez-Paz
  Why Does Throwing Away Data Improve Worst-Group Error?
  ICML 2023 (Oral), ^*Equal contribution.

• Kartik Ahuja, Jason Hartford, Yoshua Bengio.
  Weakly Supervised Representation Learning with Sparse Perturbations.
  NeurIPS 2022.

• Kartik Ahuja^*, Jason Hartford^*, Yoshua Bengio.
  Properties from Mechanisms: An Equivariance Perspective on Identifiable Representation Learning.
  ICLR, 2022 (Spotlight) ^*Equal contribution.

• Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Audet, Yoshua Bengio, Ioannis Mitliagkas, Irina Rish.
  Invariance Principle Meets Information Bottleneck for Out-of-Distribution Generalization.
  NeurIPS, 2021 (Spotlight) (code)

• Kartik Ahuja, Jun Wang, Amit Dhurandhar, Karthikeyan Shanmugam, Kush R. Varshney.
  Empirical or Invariant Risk Minimization? A Sample Complexity Perspective.
  ICLR, 2021. (video)(code)

• Kartik Ahuja, Karthikeyan Shanmugam, Kush R. Varshney, Amit Dhurandhar.
  Invariant Risk Minimization Games.
  ICML, 2020 (video)(code).

Curriculum Vitae

Here is a link to my CV last updated April, 2025.