Alexander Detkov
PhD student, Computation and Neural Systems, Caltech
I study how neural networks build world models: how local observations are glued into a coherent representation of topological, geometric, and semantic structure.
I'm advised by Matt Thomson. Before Caltech, I studied Engineering Physics and Mathematics at the University of Alberta.
Papers
-
Do World Models Learn Global Understanding?
Tests whether models learn global constraints from local transitions.
A. Detkov, M. Thomson. arXiv 2026.
-
Prompt Baking
Memory consolidation by baking LLM context into weights.
A. Detkov*, A. Bhargava*, C. Witkowski*, M. Thomson. arXiv 2024.
-
Reparameterization through Spatial Gradient Scaling
Explains why multi-branch blocks improve CNN training.
A. Detkov, M. Salameh, M. Fetrat, J. Zhang, W. Lui, S. Jui, D. Niu. ICLR 2023.
-
Imbibition of a Single Polymer into a Nanocapillary
Molecular and Langevin dynamics study of polymers entering nanopores.
A. Detkov, W. Stroberg. CSME 2022.
