CS Researcher · ML Engineer · UW Allen School

Sulayman
Yusuf

CS student at UW researching sparse autoencoders, mechanistic interpretability, and representation geometry.

Accepted · ICLR 2026

RT-TopKSAE
Improving Top-k SAEs with the Rotation Trick

Re-Align Research Workshop

Adapting the rotation trick to TopK sparse autoencoders — rotating the k-activated subspace to preserve principal components across gradient updates, eliminating dead features entirely.

100%

dictionary utilization vs 47% baseline

Read the Paper

Research Interests

Geometric Deep Learning

Neural architectures that respect the symmetry groups of data — SO(3), SE(3), and beyond. Equivariance as inductive bias.

Sparse Autoencoders

Mechanistic interpretability via geometric sparsity. Recovering the true over-complete dictionary of representation space.

SSL Geometry

How CNN inductive biases shape self-supervised representation geometry — alignment, uniformity, and dimensional collapse.