Hi! I'm Themis! I am a PhD Candidate at Boston University, jointly advised by Krzysztof Onak and Venkatesh Saligrama. I did my undergrad at Dartmouth College advised by Amit Chakrabarti. I am currently a Research Intern at Google Research.

My research develops efficient and reliable foundation models, spanning sparse attention, inference optimization, Transformer theory, robustness, and generative 3D modeling.

I am seeking Research Scientist and Research Engineer positions beginning January 2027.

Current Research

I am developing sparse Transformer architectures that combine sparsity while matching or outperforming dense baselines.

I am also working on generative modeling for efficient 3D generation.

Selected Publications

  • 1
    Layerwise Dynamics for In-Context Classification in Transformers
    Lutz, Haris, Chandra, Gangrade, Saligrama
    We show how Transformers geometrically separate classes layer-by-layer during in-context learning.
    ICML 2026 ★ Spotlight arXiv
  • 2
    Noise Stability of Transformer Models
    Haris, Zhang, Yoshida
    We introduce noise stability to measure model simplicity and accelerate Transformer training and grokking.
    ICLR 2026arXiv
  • 3
    Compression Barriers for Autoregressive Transformers
    Haris, Onak
    We prove fundamental limits on KV cache compression, showing when sublinear memory is impossible.
    COLT 2025arXiv
  • 4
    $k$NN Attention Demystified: A Theoretical Exploration for Scalable Transformers
    Haris
    We provide the first theoretical guarantees and fast sub-quadratic algorithms for $k$NN attention.
    ICLR 2025 arXiv
All publications →