Full Bandwidth Transformer could preserve information in final layers, says Langford

JohnCLangford · x · 2026-09-06

Continuing his thread, John Langford argues a Full Bandwidth Transformer could use early layers for deep compute while final layers preserve relevant information instead of discarding it — yielding better depth utilization and more depth, with gradients free to quench irrelevant prior-token information.

Related event: Langford Explores Information Dropping in Transformer Architectures(3 posts)→

Original post →

More from Research

Research channel →