Survey Explores Multimodal Unlearning for Foundation Models

The increasing adoption of large multimodal foundation models, including those for vision, language, video, and audio, brings a significant challenge: these models can unintentionally embed sensitive, copyrighted, biased, or unsafe information from their training data. Traditional retraining to remove such data after deletion requests or policy changes is often impractical due to the sheer scale and complexity of these models. Multimodal unlearning emerges as a solution, enabling the selective removal of specific knowledge across different modalities while preserving the model's overall utility. This survey offers a unified, system-oriented perspective on this nascent field, covering methods, datasets, and benchmarks relevant to vision, language, audio, and video. The paper introduces a taxonomy that facilitates systematic comparison of unlearning techniques across various model architectures and modalities. It clarifies the trade-offs involved, such as deletion strength, knowledge retention, computational efficiency, reversibility, and robustness. By outlining current advances, emerging applications, and persistent open problems, the survey aims to guide future research and practical deployment of multimodal unlearning.