ezyang: sparse autoencoders can be applied per residual add, not per block

ezyang · x · 2026-09-26

ezyang (PyTorch lead) shares a small observation: SAC is conventionally applied at transformer block granularity, but since many SAC policies end up always saving the result of a residual add — the sole input into attn/mlp — you can apply SAC to each residual add individually.

Related event: PyTorch Maintainer: SAC Can Be Applied at Finer Granularity Than Whole Blocks(2 posts)→

Original post →

More from Research

Research channel →