Beyond Sparse Dictionaries: Measuring Subspace Geometry in Diffusion Transformers
I applied MCR² coding rate metrics to a frozen PixArt-α and found progressive subspace separation, dense activations, and text-independent structure — evidence that a subspace lens may complement SAE-based interpretability.
April 13, 2026 14 min read machine-learning interpretability diffusion representation-learning mcr2 
