Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Cross-Layer Transcoders (CLTs) are introduced as sparse, depth-aware proxy models for MLP blocks within Vision Transformers (ViTs), aiming to enhance interpretability and trustworthiness. Unlike Sparse Autoencoders, CLTs capture cross-layer computational structure by reconstructing post-MLP activations from sparse embeddings of preceding layers. The researchers trained CLTs on CLIP ViT-B/32 and ViT-B/16 across CIFAR-100, COCO, and ImageNet-100 datasets, demonstrating high reconstruction fidelity (cosine similarities 0.92–0.97, R² 0.89–0.95) and preserved or improved CLIP zero-shot classification accuracy. CLTs provide faithful attribution through cross-layer contribution scores, revealing that the final representation is concentrated in a small subset of dominant layers. Ablation experiments showed removing the single highest-scored layer degraded accuracy by up to 5.6%, while retaining only the top-4 layers recovered near-baseline performance.

Key takeaway

Research scientists working on Vision Transformer interpretability should consider integrating Cross-Layer Transcoders (CLTs) into their analysis workflows. CLTs provide a robust method for understanding cross-layer contributions and identifying critical layers, especially for the [CLS] token, without significantly impacting zero-shot classification performance. This enables a more granular, depth-aware understanding of how information flows and aggregates within ViTs, which is crucial for building more transparent and controllable models.

Key insights

Cross-Layer Transcoders offer a depth-aware, interpretable alternative to Vision Transformer MLP blocks without performance compromise.

Principles

• Cross-layer context improves interpretability over single-layer analysis.
• Final ViT representations are shaped by a few dominant layers.
• CLS tokens aggregate information broadly across layers.

Method

CLTs reconstruct post-MLP activations from sparse features of preceding layers using an encoder-decoder scheme. Three sparsity functions (JumpReLU, ReLU-Top-k, Abs-Top-k) were tested, with ReLU-Top-k often performing best.

In practice

• Replace ViT MLP blocks with CLTs for interpretability.
• Focus on [CLS] token or late-layer substitutions for minimal accuracy impact.
• Use projection-based attribution scores to identify critical layers.

Topics

• Cross-Layer Transcoders
• Vision Transformers
• Model Interpretability
• Sparse Autoencoders
• Zero-Shot Classification

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision — Cross-Layer Transcoders provide interpretable, layer-resolved representations for Vision Transformers by modeling cross-layer computational structure.
• UNBOX: Unveiling Black-box visual models with Natural-language — UNBOX enables black-box visual model interpretability by using LLMs and diffusion models for semantic search.
• Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models — CA-LIG provides context-aware, layer-wise explanations for Transformer models by integrating token attributions with attention gradients.
• Sparsity as a Key: Unlocking New Insights from Latent Structures for Out-of-Distribution Detection — Sparse Autoencoders can disentangle Vision Transformer features, revealing structural invariants for robust OOD detection.
• Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability — Unimodal VLM explainability metrics fail due to cross-modal redundancy; synergistic metrics are crucial for faithful evaluation.
• Vision Transformer-Conditioned UNet for Domain-Adaptive Semantic Segmentation — ViTC-UNet combines frozen ViT global priors with UNet local inductive biases for efficient, high-precision biomedical semantic segmentation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.