LatentScalpel

Sparse autoencoders for mechanistic interpretability of diffusion language models. Train SAEs on residual stream activations, automatically interpret learned features, and run causal interventions during the reverse diffusion chain to measure their impact on generation.

Installation

Related servers