sae-middle-layer-residual-stream-target

IN premisesummaries/2026/08/24/templeton-2024-scaling-monosemanticity-chunk-3.md

Created 2026-08-25T02:58:38+00:00

The SAE is applied to residual stream activations at the middle layer specifically to reduce dimensionality versus MLP output and to mitigate cross-layer superposition.

Summary

The sparse decoder is pointed at the residual stream in the middle of the network rather than at the raw MLP output, because the residual stream is a smaller, cleaner signal to decompose. This design choice avoids the mess of features from different layers overlapping and interfering with each other, which keeps the learned sparse features more interpretable and less tangled.

Dependents

These beliefs depend on this one: