Thomas Dooms
Bio
Thomas Dooms is a PhD student in Computer Science at the University of Antwerp (IDLab), where he focuses on compositional interpretability — understanding how neural networks encode complex behaviors and knowledge by treating model weights as compositional systems rather than reducing them to simplistic feature sets. His key publication, "Bilinear MLPs enable weight-based mechanistic interpretability" (co-authored with Michael T. Pearce, Alice Rigg, Jose M. Oramas, and Lee Sharkey), was accepted as a spotlight paper at ICLR 2025 and demonstrates that bilinear MLP weights can be fully analyzed via third-order tensor decomposition, enabling circuit discovery and adversarial example construction directly from weights. As of January 2026, he is also a fellow at Goodfire, an AI interpretability company. He co-mentors applicants in the MARS AI safety fellowship program alongside Logan Riggs, and has received funding from the Long-Term Future Fund in support of his interpretability research.
Links
- Personal Website
- https://tdooms.github.io/
- Twitter / X
- LessWrong
- tdooms
Grants
No grants recorded.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.
Details
- Last Updated
- Mar 23, 2026, 1:31 AM UTC
- Created
- Mar 20, 2026, 3:00 AM UTC