Meenakshi Khosla

@andre-longon.bsky.social led/executed this project beautifully—he's applying to PhD programs this fall and would be an incredible addition to any lab!

also thanks to @david-klindt.bsky.social
for an incredible collaboration.

The takeaway: superposition isn’t just an interpretability issue—it warps alignment metrics too. Disentangling reveals the true representational overlap between models and between models and brains.

Across toy models, ImageNet DNNs (ResNet, ViT), and even brain data (NSD), alignment scores jump once we replace base neurons with their disentangled SAE latents—showing that superposition can mask shared structure.

We develop a theory showing how superposition arrangements deflate predictive-mapping metrics. Then we test it: disentangling with sparse autoencoders (SAEs) reveals hidden correspondences.

Superposition disentanglement of neural representations reveals hidden alignment The superposition hypothesis states that a single neuron within a population may participate in the representation of multiple features in order for the population to represent more features than the ...

Superposition has reshaped interpretability research. In our @unireps.bsky.social paper led by @andre-longon.bsky.social we show it also matters for measuring alignment! Two systems can represent the same features yet appear misaligned if those features are mixed differently across neurons.