Attention Residuals may expose cross-layer information flow directly for interpretability

tokenbender · x · 2026-07-28

The post points out an overlooked use of Attention Residuals (AttnRes) for interpretability work.

Because AttnRes gives each layer a learned pseudo-query and an explicit depth-wise weight matrix over earlier blocks, the model effectively exposes cross-layer information flow instead of forcing researchers to infer it indirectly through activation patching, path attribution, or causal tracing.

The author’s takeaway is that this could make interpretability much easier: instead of reverse-engineering who contributed what to the residual stream, you can inspect the layer-to-layer read distribution directly. They note the idea is still preliminary and may have caveats.

Original post →

Attention Residuals may expose cross-layer information flow directly for interpretability

More from Research