The 5th Explainable AI for Computer Vision (XAI4CV) Workshop at CVPR 2026

Date & Time: June 3rd, 2026, 8:00 AM - 1:00 PM (MDT local time)
Location: Mile High 1AB, Colorado Convention Center, Denver, CO, USA

Motivation

Computer vision for high-stakes, real-world applications necessitates robust explanation and transparency to ensure trust, accountability, and ethical deployment. Celebrating its 5th Anniversary, the Explainable AI for Computer Vision (XAI4CV) workshop provides a premier forum for the entire spectrum of XAI research, from interpretable-by-design models to challenges in multimodal foundational models. The program includes invited talks, spotlight papers, a poster session, and a tutorial. XAI4CV accepts paper and demo submissions to define the future of trustworthy visual AI.

Schedule

The schedule is in MDT local time.

08:15–08:30	Opening
08:30–09:00	Invited Talk 1 – Chaofan Chen Learning by Comparison: Case-based Reasoning for Interpretable Vision Models. Deep neural networks have achieved remarkable success in computer vision, yet their decision-making processes often remain opaque. This lack of transparency often limits their use in high-stakes domains such as healthcare. In this talk, I will present a line of work focused on interpretable vision models that make predictions using case-based reasoning, by comparing inputs with learned prototypical examples. I will also highlight how these interpretable models could be used in real-world applications, such as medical imaging, to improve transparency and user trust in AI-assisted decision-making.
09:00–09:30	Invited Talk 2 – Anh Totti Nguyen Vision Language Models with Explainable Bottleneck Layers. By design, VLM is a challenging type of neural network to "explain" due to (1) the parallel processing of multi-head attention; (2) dual image-text input streams; (3) the lack of a natural interface to point users to where evidence is in the input image. Empirically, VLMs tend to be heavily biased towards the knowledge in the language model and sometimes ignoring the signals from the image itself. In this talk, I'd advocate for explaining VLMs by first defining a bottleneck layer where we want to 2-way exchange the explanatory information between the user and the neural network. Depending on where in the network we want to extract information and insert information back into the network (for ablation study or further processing), we want to build a different explainability component. First, I'll talk about the phenomenon that VLMs are often acting like a short-sighted person (VLMs are Blind) and heavily biased to the text data (VLMs are Biased). Then, I'll share distinct attempts of building an explainable bottleneck in various points in the VLM: the attention layer (Transformer Attention Bottleneck); pen-ultimate layer (PEEB); and output layer or VLMs (Highlighted Chain of Thoughts, PageGuide, SketchVLM). I'll share our most recent, exciting demos of how frontier VLMs (Gemini 3 Pro) can be turned into a VLM that annotates your input image or computer screen to guide users through local-computer or web tasks.
09:30–10:00	Spotlight Session 1 Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision. Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li, Difei Gu, Gemma Elyse Moran, Hao Wang, Dimitris Metaxas. Multi-Granularity Concept Whitening for Neural Network Interpretability. Russell Barton, Hung Le, Yunhong Shan, Alexander Katopodis, Jonathan Donnelly, Eric Chen, Chaofan Chen, Cynthia Rudin. Activation-Based Concept Extraction for Explainability in Image Classification. Matteo Bianchi, Riccardo Campi, Antonio De Santis, Sara Merengo, Marco Brambilla.
10:00–10:30	Coffee Break
10:30–11:00	Invited Talk 3 – Elizabeth Barnes The Final Model Is Not Enough: Training Dynamics as a Form of XAI. AI weather and climate models now rival operational forecast systems, but interpreting these models remains a major challenge. Standard explainable AI methods, while effective for simpler climate prediction networks, have proven difficult to apply to large autoregressive emulators operating on high-dimensional spatiotemporal fields. In this talk, I argue that the training trajectory itself offers a complementary and underutilized diagnostic lens. Using AI weather emulators as a testbed, I show that models actively learn and then unlearn specific extreme weather events during training, that learning dynamics differ qualitatively across forecast tasks, and that designed perturbation experiments can pinpoint when models acquire knowledge of physical relationships between variables. These results suggest that for complex AI models where post-hoc explanation is intractable, studying how a model arrives at its final state may be as revealing as explaining what that final state does.
11:00–11:30	Spotlight Session 2 DINO-QPM: Adapting Visual Foundation Models for Globally Interpretable Image Classification. Robert Zimmermann, Thomas Norrenbrock, Bodo Rosenhahn. FaCT: Faithful Concept Traces for Explaining Neural Networks. Amin Parchami-Araghi, Sukrut Rao, Jonas Fischer, Bernt Schiele. Interpretable 3D Neural Object Volumes for Robust Conceptual Reasoning. Nhi Pham, Artur Jesslen, Bernt Schiele, Adam Kortylewski, Jonas Fischer.
11:30–12:00	Tutorial – Maximilian Dreyer From Concepts to Control: Diagnosing and Steering Vision Foundation Models. This tutorial shows how to turn concept-based explanations into diagnosis and control for vision(-language) models. We train a sparse autoencoder (SAE) on CLIP embeddings to discover interpretable components, explore them via textual search, and compute component-level attributions for single examples to see which concepts drive predictions. We then use these concepts to reveal biases (e.g., gender associations for “nurse”) and demonstrate lightweight control strategies: suppressing or amplifying specific concepts to test and correct the model. We briefly replicate the workflow for Qwen3‑VL, noting the differences for multimodal hidden states. Participants leave with a practical recipe, pitfalls to avoid, and guidance for diagnosing and correcting when models rely on undesired concepts. The tutorial is available on GitHub
12:00–12:15	Closing Remarks
12:15–13:00	Poster Session

Invited Speakers

Elizabeth Barnes

Boston University

Chaofan Chen

University of Maine

Anh Totti Nguyen

Auburn University

Maximilian Dreyer

Fraunhofer Heinrich Hertz Institute

Accepted Papers

Spotlight Papers

P01 How does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders. Yiming Tang, Abhijeet Sinha, Dianbo Liu.
P05 Multi-Granularity Concept Whitening for Neural Network Interpretability. Russell Barton, Hung Le, Yunhong Shan, Alexander Katopodis, Jonathan Donnelly, Eric Chen, Chaofan Chen, Cynthia Rudin.
P07 DINO-QPM: Adapting Visual Foundation Models for Globally Interpretable Image Classification. Robert Zimmermann, Thomas Norrenbrock, Bodo Rosenhahn.
P09 Activation-Based Concept Extraction for Explainability in Image Classification. Matteo Bianchi, Riccardo Campi, Antonio De Santis, Sara Merengo, Marco Brambilla.
P13 Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision. Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li, Difei Gu, Gemma Elyse Moran, Hao Wang, Dimitris Metaxas.
P15 Concepts in Motion: Temporal Bottlenecks for Interpretable Video Classification. Patrick Knab, Sascha Marton, Philipp Schubert, Drago Guggiana Nilo, Christian Bartelt.
P16 FaCT: Faithful Concept Traces for Explaining Neural Networks. Amin Parchami-Araghi, Sukrut Rao, Jonas Fischer, Bernt Schiele.
P19 Interpretable 3D Neural Object Volumes for Robust Conceptual Reasoning. Nhi Pham, Artur Jesslen, Bernt Schiele, Adam Kortylewski, Jonas Fischer.

Proceedings Track

P01 How does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders. Yiming Tang, Abhijeet Sinha, Dianbo Liu.
P02 FAME: Feature Activation Map Explanation on Image Classification and Face Recognition. Xinyi Zhang, Manuel Günther.
P03 From Attribution to Action: A Human-Centered Application of Activation Steering. Tobias Labarta, Maximilian Dreyer, Sebastian Lapuschkin, Katharina Weitz, Nhi Hoang, Jim Berend, Wojciech Samek.
P04 Faces in Focus: Evaluating Generalizability and Transferability of Facial Recognition Explainability Techniques for CNN-based Models. Paweł Borsukiewicz, El-hacen Diallo, Abdoul Aziz Bonkoungou, Wendkûuni C. Ouédraogo, Jacques Klein, Charles Beumier, Tegawendé F. Bissyandé.
P05 Multi-Granularity Concept Whitening for Neural Network Interpretability. Russell Barton, Hung Le, Yunhong Shan, Alexander Katopodis, Jonathan Donnelly, Eric Chen, Chaofan Chen, Cynthia Rudin.
P06 Investigating Anisotropy in Visual Grounding under Controlled Counterfactual Perturbations. Gabriele Lombardo, Luigi Maiorana, Liliana Lo Presti, Marco La Cascia.
P07 DINO-QPM: Adapting Visual Foundation Models for Globally Interpretable Image Classification. Robert Zimmermann, Thomas Norrenbrock, Bodo Rosenhahn.
P08 Why Fake? Unveiling the Semantic Vocabulary of Deepfake Detectors. Vazgken Vanian, Alexandros Doumanoglou, Dimitris Zarpalas.
P09 Activation-Based Concept Extraction for Explainability in Image Classification. Matteo Bianchi, Riccardo Campi, Antonio De Santis, Sara Merengo, Marco Brambilla.
P10 Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions? Kamalasankari Subramaniakuppusamy, Jugal Gajjar.
P11 A Benchmark Study on the Reliability of Explainability Methods. Ibna Kowsar, Shahbaz Rezaei, Fanyu Meng, Xin Liu.
P12 A Mechanistic Analysis of Training-Time Image Protection in Diffusion Models. Michael Martin, Garrick Chan, Kwan-Liu Ma.
P13 Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision. Gerasimos Chatzoudis, Konstantinos D. Polyzos, Zhuowei Li, Difei Gu, Gemma Elyse Moran, Hao Wang, Dimitris Metaxas.
P14 GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension. Bochu Ding, Augustus Wendell, Brinnae Bent.

Non-Proceedings Track

P15 Concepts in Motion: Temporal Bottlenecks for Interpretable Video Classification. Patrick Knab, Sascha Marton, Philipp Schubert, Drago Guggiana Nilo, Christian Bartelt.
P16 FaCT: Faithful Concept Traces for Explaining Neural Networks. Amin Parchami-Araghi, Sukrut Rao, Jonas Fischer, Bernt Schiele.
P17 Hierarchical Concept Embedding & Pursuit for Interpretable Image Classification. Ho Huu Nghia Nguyen, Tianjiao Ding, Rene Vidal.
P18 Interpretable and Steerable Concept Bottleneck Sparse Autoencoders. Akshay Kulkarni, Tsui-Wei Weng, Vivek Narayanaswamy, Shusen Liu, Wesam Sakla, Kowshik Thopalli.
P19 Interpretable 3D Neural Object Volumes for Robust Conceptual Reasoning. Nhi Pham, Artur Jesslen, Bernt Schiele, Adam Kortylewski, Jonas Fischer.
P20 Object-Level Explanations for Image Geolocation Models: a GeoGuessr use-case. Emilie Durrieu, Christophe Hurter, Philippe Muller, Victor Boutin.
P21 HyperPG: Probabilistic Prototypes on Hyperspheres for Interpretable Deep Learning. Maximilian Li, Korbinian Rudolf, Paul Mattes, Nils Blank, Rudolf Lioutikov.
P22 Concept Regions Matter: Benchmarking CLIP with a New Cluster-Importance Approach. Aishwarya Agarwal, Srikrishna Karanam, Vineet Gandhi.
P23 Evaluation Cards for XAI Metrics. Rokas Gipiškis, Olga Kurasova.
P24 Extended Abstract: CHiQPM: Calibrated Hierarchical Interpretable Image Classification. Thomas Norrenbrock, Timo Kaiser, Sovan Biswas, Neslihan Kose, Bodo Rosenhahn.
P25 Explaining CLIP Zero-shot Predictions Through Concepts. Onat Ozdemir, Anders Christensen, Stephan Alaniz, Zeynep Akata, Emre Akbas.
P26 What does CLIP see per token?. Wajahat Khan, Seungkyu Lee.
P27 How Many Visual Levers Drive Urban Perception? Interventional Counterfactuals via Multiple Localised Edits. Jason Tang.
P28 Explainable Visual Anomaly Detection via Concept Bottleneck Models. Arianna Stropeni, Valentina Zaccaria, Francesco Borsatti, Davide Dalle Pezze, Manuel Barusco, Gian Antonio Susto.
P29 Concept-Aware Pruning via Disentangled Subspaces for Robust Convolutional Networks. Kirin Danek, Vikram Ramaswamy.

Organizers

Miguel-Ángel Fernández-Torres

Universidad Carlos III de Madrid

Yuhui Zhang

Stanford University

Jon Donnelly

Duke University

Maximilian Dreyer

Fraunhofer Heinrich Hertz Institute

Marina L. Gavrilova

University of Calgary

Dahye Kim

Boston University

Indu Panigrahi

University of Illinois Urbana-Champaign

Sukrut Rao

Max Planck Institute for Informatics

Avinab Saha

UT Austin, Google Research

Lenka Tětková

Technical University of Denmark

Program Committee

We thank our great Program Committee members who made this workshop possible!

Alina Barnett, Quentin Bouniot, Thea Brüsch, Tzoulio Chamiti, Jinwoo Choi, Joseph-Paul Cohen, Fernando Díaz-De-María, Maximilian Dreyer, Jonathan Donnelly, Miguel-Ángel Fernández-Torres, Marina Gavrilova, John Gkountouras, Maria Gonzalez-Calabuig, Iván González-Díaz, Ada Görgün, Sadaf Gulshad, Shashank Gupta, Adrian Höhl, Nils Huetten, Dahye Kim, Tobias Labarta, Lorenz Linhardt, Manxi Lin, Raphael Maser, Miguel Molina-Moreno, Elisa Nguyen, Ivica Obadic, Matthew Olson, Indu Panigrahi, Amin Parchami-Araghi, Paraskevas Pegios, Nhi Pham, Vipin Pillai, Sukrut Rao, Fawaz Sammani, Simone Schaub-Meyer, Mayank Singh, Shreya Tendulkar, Lenka Tětková, Navneet Tyagi, Kristoffer Wickstrøm, Romain Xu-Darme, Xiwei Xuan, Luna Zhang, Mengxue Zhang

(Closed) Call for Papers and Demos

We welcome paper and demo submissions:

Papers should describe high-quality, original research. Contributions can include novel XAI methods; applications of existing methods on new domains, models, and tasks; evaluation or analysis of existing methods; and practical toolboxes.
Demos should consist of static or interactive presentations of XAI for CV models and tasks, accompanied by a description. Contributions can include visualizations, explanations, and explorations of novel XAI systems; novel visualizations, explanations, and explorations of existing XAI systems; studies of how different visualizations, explanations, and explorations of XAI systems are perceived by people; among others.

We have two tracks of submissions:

Proceedings track: We welcome max 8-page submissions of papers and demos. Submissions accepted to this track will be published in the CVPR workshop proceedings.
Non-proceedings track: We welcome max 4-page submissions (commonly referred to as "extended abstracts") of papers and demos. For the non-proceedings track, we encourage submissions of published or accepted work (e.g., papers and demos accepted to the CVPR main program). Submissions accepted to this track will not be published in the CVPR workshop proceedings.

Topics

We encourage submissions on topics including, but not limited to:

Interpretable-by-design computer vision (CV) models, including transparent CNNs, Vision Transformers (ViTs), and hybrid architectures designed for intrinsic interpretability.
Post-hoc explanation methods for CV models, such as saliency and activation mapping, feature visualization, and counterfactual reasoning.
Mechanistic interpretability, including reverse-engineering network behavior, analyzing layer-wise and concept-level representations, and understanding learned mechanisms.
Multimodal XAI, covering multimodal explanations of CV models (e.g., vision-language, vision-audio) and unimodal explanations of multimodal systems.
Evaluation and benchmarking of XAI methods, including metrics, robustness analysis, human evaluations, and comparative studies.
Datasets for XAI, supporting benchmarking, reproducibility, and human-centered evaluations.
Open-source frameworks and tools for XAI, enabling transparent and scalable research and deployment.
Human-centered XAI, including user studies, human-in-the-loop explanation systems, and the assessment of trust, usability, and decision support.
XAI applications across domains, including healthcare, autonomous systems, robotics, geosciences, and remote sensing.
Explainability in relation to broader constructs, such as fairness, transparency, interpretability, accountability, causality, and trust, and its implications for society.
Emerging directions, including counterfactual and causal explanations, interpretability of foundational and generative models, concept discovery, interactive and adaptive explanations, and evaluation of XAI in real-world deployments.

Timeline

Proceedings Track

Submission Deadline: February 27, 2026 (Anywhere on Earth)
Notification to Authors (Accept as Spotlight, Poster, or Reject): March 20, 2026 (Anywhere on Earth)
Camera Ready Deadline: April 7, 2026 (Anywhere on Earth)

Non-Proceedings Track

Submission Deadline (to be considered for Spotlights): ~~March 6, 2026~~ March 20, 2026 (Anywhere on Earth)
Notification to Authors (Accept as Spotlight, Poster, or Reject): ~~March 27, 2026~~ April 7, 2026 (Anywhere on Earth)
Rolling Submissions and Notifications (Accept as Poster or Reject): Until April 7, 2026 (Anywhere on Earth)

Submission Instructions

All submissions should be in the anonymized CVPR 2026 format available at https://cvpr.thecvf.com/Conferences/2025/AuthorGuidelines.
The page limits do not include references.
You may optionally upload supplementary material. Reviewers will be encouraged to look at it, but are not obligated to do so.
Submissions can be done at https://cmt3.research.microsoft.com/XAICV2026

The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support.

Attendance and Presentation

Posters: All accepted submissions will be invited to participate in an in-person poster session at our workshop. Additionally, the authors will be asked to upload their posters which will be hosted on our webpage.
Spotlights: We will pick several works among the submissions to be presented as spotlights. Presentations can either be in-person or pre-recorded.
Abiding by the CVPR guidelines, all accepted papers must be presented by one of the authors.