The paper introduces MedForget, a hierarchy‑aware multimodal unlearning benchmark that models medical data as nested institutional structures, allowing fine‑grained evaluation of retain and forget splits. Experiments reveal that existing unlearning methods struggle to achieve effective hierarchy‑aware forgetting without harming retained utility on clinically relevant tasks. To address this, the authors propose CHIP, a training‑free method that deletes target‑specific weight subspaces while preserving sibling‑shared information, achieving superior forget‑retain performance across all hierarchy levels while maintaining competitive downstream utility.
arXiv:2607. 07907v1 Announce Type: cross Abstract: With the growing adoption of VLMs, DMs, LLMs, and AFMs, these multimodal foundation models can inadvertently encode sensitive, copyrighted, biased, or unsafe cross-modal associations that originate from their training data.
MLLMEraser is a training‑free, input‑aware framework that achieves test‑time unlearning for multimodal large language models by steering activations. It constructs an erasure direction through adversarial contrast between knowledge‑recall and knowledge‑erasure image‑text pairs, and applies this direction selectively to preserve retained knowledge while forgetting specified content. Experiments on LLaVA‑1.5 and Qwen‑2.5‑VL show that MLLMEraser outperforms existing unlearning baselines with lower computational cost and minimal utility loss.
The paper introduces a model‑agnostic framework that lets frozen large language models (LLMs) and vision‑language models (VLMs) learn from deployment experience without retraining. It does so by adding three external expertise components—a Skill for guiding reasoning and tool use, a Knowledge Memory for storing reliable facts, and a Multimodal Knowledge Base for visual examples—and by using a validation strategy that updates only when new cases improve performance without harming past performance. Experiments on six benchmarks with both open‑weight and closed‑source models show up to a 34.2% performance gain on medical tasks, generalization to unseen cases, and transferability to other models and non‑medical domains.
arXiv:2608.01008v2 Announce Type: replace Abstract: Multimodal large language models (MLLMs) can memorize and disclose sensitive information, motivating machine unlearning to remove designated knowle...
arXiv:2608. 03791v1 Announce Type: new Abstract: Vision-Language Models (VLMs), like Large Language Models (LLMs), may memorize sensitive, copyrighted, or harmful knowledge from their pretraining corpora.
arXiv:2607. 18615v1 Announce Type: cross Abstract: Machine unlearning for vision-language models (VLMs) remains underexplored.
arXiv:2606. 12809v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) are trained on massive multimodal data, making data unlearning increasingly important as data owners may request the removal of specific content.
arXiv:2606. 07141v1 Announce Type: cross Abstract: Language models trained for clinical disease inference are trained on patient data, which may include sensitive and private information, and data owners may request the removal of their data from a trained model due to privacy or copyright concerns.
UniH$^3$ is a new framework for all-in-one medical image restoration that unifies hierarchical homogeneity and heterogeneity. It introduces a Hierarchical Homogeneity Memory (H2M) module to distill and retrieve shared anatomical priors, and a Hierarchical Heterogeneity Balancer (H2B) to mitigate inter- and intra-task conflicts during training. Experiments on MedIR-2D-500K and MedIR-3D-3K show that UniH$^3$ achieves state‑of‑the‑art performance for both multi‑task and single‑task restoration.
Machine unlearning for vision-language models (VLMs) remains underexplored. Unlike language models, VLMs combine a language backbone with visual components, which makes unlearning more complex.
The paper discusses the rapid progress of machine unlearning for multimodal large language models (MLLMs) and the need for rigorous evaluation. It identifies three pitfalls in current benchmarks: lack of held‑out evaluation, reliance on synthetic fine‑tuned knowledge, and unfair comparisons across methods. To address these issues, the authors propose GEMMU, a benchmark using real‑world knowledge and a principled protocol, and show that re‑evaluating existing methods under GEMMU uncovers significant generalization gaps.
Vision-Language Models (VLMs), like Large Language Models (LLMs), may memorize sensitive, copyrighted, or harmful knowledge from their pretraining corpora. Removing such knowledge is essential for building trustworthy AI systems.