Ekdeep Singh Lubana

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

Priors in time: Missing inductive biases for language model interpretability

Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders

Belief Dynamics Unify In-Context Learning and Activation Steering

Emergence of Hierarchical Emotion Organization in Large Language Models

From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?

From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit

Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry

Competition Dynamics Shape Algorithmic Phases of In-Context Learning

A Percolation Model of Emergence: Analyzing Transformers Trained on a Formal Language

Detecting High-Stakes Interactions with Activation Probes

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

ICLR: In-Context Learning of Representations

Swing-by Dynamics in Concept Learning and Compositional Generalization

In-Context Learning Strategies Emerge Rationally

Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing

Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing