Sofian Zalouk

Contact: saz43 [at] cornell [dot] edu

prof_pic.jpg

I’m a third-year PhD student in Computer Science at Cornell University, focusing on machine learning, specifically diffusion language models. I am fortunate to be advised by Prof. Chris De Sa and Prof. Kilian Q. Weinberger. Previously, I was an MS student in Computer Science at Stanford University, where I worked closely with Prof. Stefano Ermon and Prof. Andrew Ng. I completed my B.A.Sc. in Electrical Engineering at the University of Toronto, where I worked with Prof. Andreas Moshovos.

My research focuses on generative modeling for language, particularly the design of efficient and scalable diffusion language models. In LOOT, we introduce the Leave-One-Out transformer that enables masked diffusion models to: 1) learn from every token during training, 2) self-correct during generation. Beyond diffusion, I am interested in separating factual knowledge from reasoning in language models. Along this line, LMLM externalizes factual knowledge into an explicit, editable knowledge base rather than memorizing it in model weights. Please feel free to reach out if you’re interested in collaborating!

news

selected publications

  1. loot.svg
    Empowering Masked Diffusion Models to Self-Correct with Leave-One-Out Transformers
    Sofian Zalouk, Vincent Counathe, Paul Jünger, Daniel Cao, Daniel Zlotnick, Kilian Q. Weinberger, and Christopher De Sa
    NeurIPS 2026
  2. lmlm.jpg
    Pre-training Limited Memory Language Models with Internal and External Knowledge
    Linxi Zhao, Sofian Zalouk, Christian K. Belardi, Justin Lovelace, Jin Peng Zhou, Ryan Thomas Noonan, Dongyoung Go, Kilian Q. Weinberger, Yoav Artzi, and Jennifer J. Sun
    ICLR 2026 (Top 2%)
    CCFM @ NeurIPS 2025 Best Paper Runner-up
  3. kernel_calibration_mmd.png
    Calibration by Distribution Matching: Trainable Kernel Calibration Metrics
    Charles Marx*, Sofian Zalouk*, and Stefano Ermon
    NeurIPS 2023
    *Equal Contribution