Would 100+ layers tip the scales between attnres and full attention residual?

stochasticchasm · x · 2026-08-28

A researcher raises a question about attention residual (attnres) vs. full attention residual: their benchmark results are already very close, and at more extreme depth (100+ layers) the rankings might flip, since attnres can always pull in full-magnitude residuals in early layers. In a follow-up reply, they note two design details: the 2x factor in the residual write-back op seems tied to init dynamics — sigmoid(0)=0.5 with zero-centered init, so a 2×sigmoid restores normal-scale writeback — and the \overline{R} notation appears unused anywhere, despite group-RMSNorm being called out as important in the prior section.

Related event: Residual Write-Back Design and AttnRes Depth Debate Draw Attention(2 posts)→

Original post →

More from Research

Research channel →