Decoding Vision Transformers: the Diffusion Steering Lens

2025/04/18 by Ryota Takatsuki, Takatsuki, Ryota, Sonia Joseph +5 · 1 voice · 2 citations
Computer Science · Engineering · #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Infrared Target Detection Methodologies #cs.AI #cs.CV

paper · pdf · doi:10.48550/arxiv.2504.13763

openalex publication_date 2025/04/18 · arxiv published 2025/04/18 · arxiv updated 2025/04/23 · openalex created_date 2025/10/11 · openalex updated_date 2026/07/28

Abstract

Logit Lens is a widely adopted method for mechanistic interpretability of transformer-based language models, enabling the analysis of how internal representations evolve across layers by projecting them into the output vocabulary space. Although applying Logit Lens to Vision Transformers (ViTs) is technically straightforward, its direct use faces limitations in capturing the richness of visual representations. Building on the work of Toker et al. (2024)~\citeToker2024-ve, who introduced Diffusion Lens to visualize intermediate representations in the text encoders of text-to-image diffusion models, we demonstrate that while Diffusion Lens can effectively visualize residual stream representations in image encoders, it fails to capture the direct contributions of individual submodules. To overcome this limitation, we propose Diffusion Steering Lens (DSL), a novel, training-free approach that steers submodule outputs and patches subsequent indirect contributions. We validate our method through interventional studies, showing that DSL provides an intuitive and reliable interpretation of the internal processing in ViTs.

Citations

• Enhancing Automated Interpretability with Output-Centric Feature Descriptions
• Transformer Layers as Painters
• The Remarkable Robustness of LLMs: Stages of Inference?
• Interpreting the Second-Order Effects of Neurons in CLIP
• The Missing Curve Detectors of InceptionV1: Applying Sparse Autoencoders to InceptionV1 Early Vision
• Mechanistic Interpretability for AI Safety -- A Review
• Automatic Discovery of Visual Circuits
• Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models
• Interpreting CLIP's Image Representation via Text-Based Decomposition
• Kandinsky: an Improved Text-to-Image Synthesis with Image Prior and Latent Diffusion
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• The Hydra Effect: Emergent Self-repair in Language Model Computations
• Don't trust your eyes: on the (un)reliability of feature visualizations
• Towards Automated Circuit Discovery for Mechanistic Interpretability
• Eliciting Latent Predictions from Transformers with the Tuned Lens
• How Well do Feature Visualizations Support Causal Understanding of CNN Activations?
• Learning Transferable Visual Models From Natural Language Supervision
• Probing Classifiers: Promises, Shortcomings, and Advances
• Of Non-Linearity and Commutativity in BERT
• Transformer Interpretability Beyond Attention Visualization
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Understanding Neural Networks via Feature Visualization: A survey
• Residual Connections Encourage Iterative Inference
• A Unified Approach to Interpreting Model Predictions
• Network Dissection: Quantifying Interpretability of Deep Visual Representations
• Highway and Residual Networks learn Unrolled Iterative Estimation
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
• Understanding intermediate layers using linear classifier probes
• "Why Should I Trust You?": Explaining the Predictions of Any Classifier

Discussions

• This work was done as my internship project at Araya. Huge thanks to my supervisors, Ippei Fujisawa & Ryota Kanai, and my external mentor @soniajoseph.bsky.social for making this happen! 🙏 Link to th [bsky, 2 points, 0 comments]

Related

• Exploring Vision Transformers as Diffusion Learners
• Rare Text Semantics Were Always There in Your Diffusion Transformer
• Exploring the Role of Large Language Models in Prompt Encoding for Diffusion Models
• Diffscaler: Enhancing the Generative Prowess of Diffusion Transformers
• Interpreting vision transformers via residual replacement model