Lorenz Hufe Berlin  ·  Updated June 2026

LorenzHufe.

About

I am an ELLIS PhD candidate at Fraunhofer Heinrich Hertz Institute under Prof. Wojciech Samek, with a secondary affiliation at the University of Oxford, where I am co-supervised by Dr. Christian Schröder de Witt and Prof. Philip Torr. My current research is focused on AI for history: building machine-learning systems that help historians read, connect, and reason about historical sources at scale.

Previously I worked on mechanistic interpretability of multimodal foundation models, with a focus on typographic understanding. I co-founded Bliss e.V., a Berlin-based non-profit research collective, and remain a core contributor to the ViT-Prisma toolkit for vision and video interpretability.

Publications
  1. 2026
  2. 2026
    ICLR 2026·project page·arXiv
    Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP
    L. Hufe, C. Venhoff, E. Purelku, M. Dreyer, S. Lapuschkin, W. Samek
  3. 2025
    The TUB Sign Language Corpus Collection
    E. Avramidis, V. Czehmann, F. Deckert, L. Hufe, et al.
  4. 2025
    From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance
    M. Dreyer, L. Hufe, J. Berend, T. Wiegand, S. Lapuschkin, W. Samek
  5. 2025
    CVPR Workshop · Oral·arXiv
    Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video
    S. Joseph, P. Suresh, L. Hufe, et al.
  6. 2025
    CVPR Workshop·arXiv
    Steering CLIP's Vision Transformer with Sparse Autoencoders
    S. Joseph, P. Suresh, E. Goldfarb, L. Hufe, et al.
  7. 2025
    SCAM: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models
    J. Westerhoff, E. Purelku, J. Hackstein, J. Loos, L. Pinetzki, L. Hufe
  8. 2022