The schedule is a half-day event on either Sept 9, 2026, from 8.30 am to 12.30 pm at Malmö Arena and Malmömässan, Malmö, Sweden. The room is Malmömässan E2

The program is finally available.

📃 Extended abstracts (non-archival)

• 10. Anupam Srivastava, Shubham Chakraborty, Sneh Nandu, "Selective Rigidity: Online Model-Editing Decisions Under Adversarial Identity Pressure"

• 13. Yusuke Kuwana, Takashi Shibata, Kiyoharu Aizawa, Go Irie, "Hard Prompt Search for Class Unlearning"

• 14. Sahil Kale, "ConceptGuard: Benchmarking Concept-Level Selective Unlearning in Large Language Models"

• 23. Xing Han, Shravan S Chaudhari, Tanvi Ranade, Rama Chellappa, Suchi Saria, "Continual Multimodal Multi-Task Learning"

• 24. Poornima Jain, Sairam VCR, Vineeth N. Balasubramanian, "Safe Visual Counterfactuals Can Make Multimodal LLMs Safer"

•  26. Ashok Urlana, L D M S Sai Teja, Vivek Hruday Kavuri, Ponnurangam Kumaraguru, "MapRoute++: Surrogate-Guided Semantic Routing for Visual Concept Unlearning" -  🏆  Winner of the visual concept track in the GenMu Challenge

• 29. Komal Kumar, Ankan Deria, Abhishek Basu, Fahad Shamshad, Hisham Cholakkal, Karthik Nandakumar, "SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training"

• 33. Yuma Sano, Takashi Matsubara, "Not All Attention Weights Interfere Alike: A Role-Aware Analysis of LoRA Merging"

Title: Evaluating AI (and AI Evaluation) as a Science: Towards Science-guided Unlearning and Model Editing (U&ME)

Abstract: A prevailing narrative holds that AI model architectures have largely converged, but convergence is only the beginning. AI progress now hinges on science-driven evaluation and iterative refinement. In this keynote, YZ will present two new works from the Active Perception Research Group: an evaluation framework that pinpoints the circuit-level origins of hallucinations in vision-language models, and a Schrödinger Bridge-based framework for optimal reference selection in generative AI. The talk will also reflect on the research journey that led to these directions, including WOUAF (CVPR 2024), RACE (ECCV 2024), VOILA (ICLR 2025), and EraseFlow (NeurIPS 2025), and some learnings from the speaker’s engagement in a GenAI for creativity project. 

Title: Where Do Concepts Live? Discrete, Linear, and Curved Geometries for Unlearning and Steering

Abstract: Every method for removing a concept from a generative model encodes an assumption — usually unstated — about where that concept lives in the model's representation. This talk argues that the assumption is the method, and traces it through three training-free interventions.

We start with discrete structures: in VQ-VAE text-to-motion models, violence concentrates in identifiable codebook entries, and replacing them matches a retrained oracle while leaving retained-motion quality intact. We then treat concepts as linear directions, using multimodal difference-in-means with contrastive PCA to sharply cut unsafe video generation at near-zero fidelity cost. Finally we consider the case of curved and field-based steering, and the correction depends on where a representation sits.

Title: Controllable Forgetting in Foundation Models: Neural Activation Redirection and the Geometry of Unlearning

Abstract: As foundation models become infrastructure across mission-critical sectors, the ability to selectively remove undesirable content is becoming a core requirement for alignment and trustworthy deployment. Removing private, copyrighted, unsafe, or outdated information must do more than suppress an answer: the model should eliminate the targeted influence without sacrificing general capabilities, hallucinating, or revealing the deletion through unnatural behavior. This talk argues that reliable unlearning is not merely an optimization problem, but a question of how knowledge and behavior are organized in representation space, and how that geometry can be modified with precision.

The talk develops this perspective through two complementary ideas. First, targeted activations can be redirected toward regions associated with epistemic abstention, enabling controlled and minimally invasive unlearning. Second, output generation relies on geometric pathways shared across otherwise distinct contexts; effective unlearning must preserve this common structure while selectively displacing only the targeted representations. Together, these principles recast unlearning as constrained geometric transport: relocating unwanted representations while preserving the subspaces and output structures that sustain retained competence. This suggests a broader framework for reliable unlearning in language, vision-language, and multimodal foundation models.

Title: Advancing Computer Vision in the Era of Vision Foundation Models

Abstract: Over the past several decades, computer vision has witnessed remarkable progress across a broad range of tasks. Recently, the field has entered a new era with the emergence of large-scale Vision Foundation Models (VFMs). By learning general-purpose visual representations from massive data, VFMs offer new opportunities to address several long-standing challenges in computer vision, including generalization, open-vocabulary understanding, and the development of unified models. In this talk, we will explore the evolving role of VFMs in vision, spanning several paradigms: leveraging frozen VFMs, adapting VFMs, developing domain-specific VFMs, and, ultimately, using VFMs as agents. Together, these approaches point toward a broader shift in how we build, adapt, and deploy computer vision systems.