Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models

Name: Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models
Uploaded: 2026-10-08T20:58:17.132910+00:00
Duration: 641 s
Description: This work introduces a residualized temporal Sparse Autoencoder (SAE) to study how diffusion model activations change over time. It separates linearly predictable patterns from residual information in activation trajectories, allowing sparse latents to capture complex structures beyond simple linear trends.

This work introduces a residualized temporal Sparse Autoencoder (SAE) to study how diffusion model activations change over time. It separates linearly predictable patterns from residual information in activation trajectories, allowing sparse latents to capture complex structures beyond simple linear trends.

Open the full episode