Can Cross-Layer Transcoders Replace Vision Transformer Activations?

An Interpretable Perspective on Vision

Gerasimos Chatzoudis1, Konstantinos D. Polyzos2, Zhuowei Li1*, Difei Gu1
Gemma E. Moran1, Hao Wang1, Dimitris N. Metaxas1
open to work Gerasimos Chatzoudis & Konstantinos D. Polyzos
Copied
1Rutgers University, 2University of California San Diego
*Work done outside of Amazon
CVPR 2026 Spotlight · XAI4CV Workshop · Denver, Colorado · June 2026
Paper arXiv Poster
Overview of the Cross-Layer Transcoder (CLT) framework: training CLTs to predict post-MLP activations, using them as a replacement model at inference, and using them as an interpretable proxy for ViTs.

Figure 1. Overview of the Cross-Layer Transcoder (CLT) framework. Left: each CLT encodes LN2 activations into sparse codes zℓ and reconstructs MLP outputs yℓ via triangular decoders. Middle: at inference, CLTs replace MLPs across layers, preserving zero-shot performance. Right: CLTs serve as an interpretable proxy for ViTs to understand the cross-layer contributions of different token types and identify the most significant layers in shaping the ViT's final representations.

Cross-Layer Transcoders replace ViT MLPs without hurting zero-shot accuracy, and reveal which layers actually build the final representation.

Functional Replacement.

CLTs can replace MLP blocks, especially in later layers for patches or for the [CLS] token across all layers, preserving and in some cases even improving zero-shot classification performance.

Patch granularity matters.

ViT-B/16 reconstruction is more faithful than ViT-B/32. Smaller patches distribute information across more tokens, yielding simpler per-token activations that are easier to approximate.

[CLS] integrates across depth; patches remain local.

Patch tokens show diagonal attribution, while [CLS] draws credit from many preceding layers.

Necessary and sufficient attribution layers.

The top-4 attributed layers recover accuracy, while removing the highest-scored layer causes substantial degradation.

BibTeX

@inproceedings{chatzoudis2026clt,
  author    = {Chatzoudis, Gerasimos and Polyzos, Konstantinos D. and Li, Zhuowei and Gu, Difei and Moran, Gemma E. and Wang, Hao and Metaxas, Dimitris N.},
  title     = {Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision},
  booktitle = {CVPR Workshop on Explainable AI for Computer Vision (XAI4CV)},
  year      = {2026},
  note      = {Spotlight}
}