Thomas Fel

Block Recurrent Dynamics in Vision Transformers

Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry

Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders

Priors in time: Missing inductive biases for language model interpretability

Cross-Modal Redundancy and the Geometry of Vision–Language Embeddings

From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit

Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models

Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry

Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment

Interpreting the linear structure of vision-language model embedding spaces

Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models

An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures

One Wave To Explain Them All: A Unifying Perspective On Feature Attribution