The Two Information Channels of Transformers
repligate · x · 2026-07-14
This piece explains how information flows within a Transformer, pointing out that many popular explainer pages fail to clarify the core mechanics.
The core ideas are:
- Transformers have two distinct information channels: the residual stream and the K/V stream.
- The residual stream passes information down across layers at the same position.
- The K/V stream propagates across positions within the same layer, enabling information exchange between different tokens.
The author details the steps occurring at each layer and position: first, the input residual stream computes the K/V for that layer, which are then used for attention calculation. Overall, it helps readers understand the Transformer through the lens of an "information highway."
More from Research
- Structural ensembles beat single predictions in TCR:pMHC generalization study — quaidmorris · 2026-07-22
- RSS launches under OMSF to push structural biology data modeling at scale — MoAlQuraishi · 2026-07-22
- enFoldX tops 8 neoantigen scans and an unseen-peptide benchmark — quaidmorris · 2026-07-22
- enFoldX reaches AUC 0.82 on human VDJdb and transfers to mouse at 0.76 — quaidmorris · 2026-07-22
- enFoldX gains accuracy as AF3 ensemble disagreement rises for non-binders — quaidmorris · 2026-07-22
- A 3D ray plot shows how hard this Jacobian counterexample is to read — moultano · 2026-07-22