Elhage et al.

A replication by the University of Chicago xLab confirms claims from Stevinson et al. (arXiv:2510.11709) that adversarial examples in toy models exploit interference between features in superposition,…

Researchers at an undisclosed institution trained a neural network that performs computation in superposition (CiS), implementing more nonlinear functions than it has neurons, by switching from L² to …

New research reveals that up to 77% of features in degraded Sparse Autoencoders (SAEs) are functionally inert, and even in well-trained models 9% of matched features are dead, challenging the interpre…