A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle
On the map: Mechanistic interpretability → Interpreting vision and multimodal models
Abstract (arXiv)
Most current paradigms in visual mechanistic interpretability (MI) remain confined to interpreting internal units of the vision model via heuristic methods (e.g., top-K activation retrieval or optimization with regularization). In this work, we establish a theoretical distributional view for visual MI, which models the influence of a feature activation on the natural image distribution, thereby formulating a Kullback-Leibler (KL)-minimal optimization problem to model the MI task. Under this framework, statistical biases are identified within previous MI paradigms, which reveal that they may either be perceptually uninterpretable to humans (i.e., deviate from the natural image distribution), or mechanistically unfaithful to the vision models (i.e., unable to activate model features). To resolve the biases under the distributional view, we propose a model with a KL-minimal soft-constraint principle for visual MI that theoretically balances interpretability and faithfulness. We realize this principle via energy-guided diffusion posterior sampling. Extensive experiments validate the theoretical soundness of the proposed distributional view and demonstrate the practical effectiveness of our paradigm on the DINOv3 vision model.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Interpreting vision and multimodal models, one of 13 directions in Mechanistic interpretability. Builds on 32 of its 38 references that are in the map. First posted on arXiv 17 May 2026.
Builds on
32 of 38 references.
Closest ancestors · 2
Same direction, newest first
- 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment6 Feb
- 2024Sparse autoencoders reveal selective remapping of visual concepts during adaptation6 Dec
Other references · 30
- 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,863
- 2020Score-Based Generative Modeling through Stochastic Differential Equationscited by 9,122
- 2021Diffusion Models Beat GANs on Image Synthesiscited by 8,255
- 2015You Only Look Once: Unified, Real-Time Object Detectioncited by 4,852
- 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,371
- 2013Visualizing and Understanding Convolutional Networkscited by 2,290
- 2022Diffusion Posterior Sampling for General Noisy Inverse Problemscited by 1,416
- 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,319
- 2014Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Imagescited by 776
- 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
- 2015Understanding Neural Networks Through Deep Visualizationcited by 334
- 2024Transcoders Find Interpretable LLM Feature Circuitscited by 200
18 more
- 2023Fundamental Limitations of Alignment in Large Language Modelscited by 155
- 2023Contrastive Energy Prediction for Exact Energy-Guided Diffusion Sampling in Offline Reinforcement Learningcited by 149
- 2016Learning a No-Reference Quality Metric for Single-Image Super-Resolutioncited by 144
- 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretabilitycited by 124
- 2023A Holistic Approach to Unifying Automatic Concept Extraction and Concept Importance Estimationcited by 109
- 2023Concept-based Explainable Artificial Intelligence: A Surveycited by 98
- 2015Visualizing Deep Convolutional Neural Networks Using Natural Pre-Imagescited by 93
- 2024Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discoverycited by 91
- 2022Low-Rank Tensor Function Representation for Multi-Dimensional Data Recoverycited by 46
- 2025Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometrycited by 34
- 2022Making Sense of Dependence: Efficient Black-box Explanations Using Dependence Measurecited by 33
- 2024Quality-Aware Image-Text Alignment for Opinion-Unaware Image Quality Assessmentcited by 32
- 2025Advancing LLM Safe Alignment with Safety Representation Rankingcited by 16
- 2025Towards Understanding the Nature of Attention with Low-Rank Sparse Decompositioncited by 9
- 2024Diffexplainer: Towards Cross-modal Global Explanations with Diffusion Modelscited by 5
- 2025DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Modelscited by 4
- 2023Saliency strikes back: How filtering out high frequencies improves white-box explanationscited by 3
- 2025VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flowcited by 3
Built on by
Citing papers per year: 2026: 1.
1 paper.
Similar, not cited
Close in wording, no citation either way
- 2023Scale Alone Does not Improve Mechanistic Interpretability in Vision Models
- 2026From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models
- 2025Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video
- 2026Capability neq Interpretability: Human Interpretability of Vision Foundation Models
- 2026Mechanistically Interpretable Neural Encoding Reveals Fine-Grained Functional Selectivity in Human Visual Cortex
- 2025A Mathematical Philosophy of Explanations in Mechanistic Interpretability -- The Strange Science Part I.i
- 2025A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima
- 2026Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers