Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks

Jul 16, 2026

Multimodal unlearning for multimodal foundation models addresses the need for selective removal of sensitive, copyrighted, biased, or unsafe cross-modal associations learned from web-scale training data. Retraining after deletion requests or policy updates is often impractical, and targeted forgetting is difficult because knowledge is distributed across shared representations. The survey systematizes vision, language, audio, and video unlearning approaches with a unified taxonomy, comparing deletion strength, retention, efficiency, reversibility, and robustness, and outlines open problems for deployment. A curated repository is released.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages29

Reading Time1 h 13 min

Introduction

Unlearning Intervention Points

Data-Side Interventions

Data-Path Perturbation Unlearning

Data Hygiene and Prompt Normalization

Training-Time Edits

Direct Gradient

Constrained Updates

Mask-Driven Selective Unlearning

Distillation-Based

Architecture Constrained

Training-Free Unlearning

Weight-space Linear Unlearning

Representation Projection

Decoding Time

Guidance-Path Control

Conditioning-Path Control

Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks Nobin Sarwar , Shubhashis Roy Dipta , Zheyuan Liu , Vaidehi Patil University of Maryland, Baltimore County University of Notre Dame UNC Chapel Hill {sms2, [sroydip1}@umbc.edu](sroydip1}@umbc.edu) [zliu29@nd.edu](zliu29@nd.edu) , [vaidehi@cs.unc.edu](vaidehi@cs.unc.edu) arXiv :2607 .07907v 1 [ cs .LG] 8 Jul 2026 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. Retraining after deletion requests or policy updates is often impractical, and targeted forgetting remains difficult because knowledge is distributed across shared representations. Multimodal unlearning addresses this challenge by enabling selective removal across modalities while retaining overall utility. This survey offers a unified, system-oriented view of multimodal unlearning across vision, language, audio, and video, grounded in recent advances, emerging applications, and open problems. Our taxonomy enables systematic comparison across model architectures and modalities, clarifying trade-offs among deletion strength, retention, efficiency, reversibility, and robustness. This survey highlights open problems and practical considerations to support future research and deployment of multimodal unlearning. We release a curated repository.1 1 Introduction Multimodal foundation models, including Vision Language Models (VLMs), Diffusion Models (DMs), Large Language Models (LLMs) and Audio Foundation Models (AFMs)-based (Ho et al., 2020 ; Team et al., 2023 ; Yang et al., 2025a ; Chu et al., 2023 ; Huang et al., 2024c) generators, support image, text, video, and audio understanding and generation at scale. Training on web-scale multimodal data improves generalization, but it can also induce memorization and undesired associations involving sensitive, copyrighted, biased, or unsafe content across modalities. As a result, deployed models may need to forget specific items 1 [https://smsnobin77.github.io/](https://smsnobin77.github.io/)[ ](https://smsnobin77.github.io/)Awesome-Multimodal-Unlearning/ Unlearning Intervention Points Figure 1: Unlearning intervention points for a Multimodal Foundation Model (MFM) . Methods intervene at the data side, during training, via architectureconstrained edits, or at decoding time, producing an updated model (MFM′) with reduced influence from targeted content. Training-free methods use closed-form parameter or representation edits (denoted by ∆) to directly transform the model without retraining. | Survey | Venue & Year | System-first | Text | Image | Video | Audio | | --- | --- | --- | --- | --- | --- | --- | | Si et al., 2023 | arXiv’23 | | ✔ | | | | | Liu et al., 2024f | arXiv’24 | ✔ | ✔ | ✔ | | | | Blanco-Justicia et al., 2025 | AIR’25 | ✔ | ✔ | | | | | Liu et al., 2025b | NMI’25 | | ✔ | | | | | Feng et al., 2025b | arXiv’25 | ✔ | ✔ | ✔ | | ✔ | | Geng et al., 2025 | arXiv’25 | ✔ | ✔ | ✔ | | | | Ours | ACL’26 | ✔ | ✔ | ✔ | ✔ | ✔ | Table 1: Comparison of multimodal unlearning surveys across modalities and system-first taxonomy coverage. or concepts, such as a copyrighted artwork, a private face, or a harmful trope, while retaining performance on the remaining data (Fan et al., 2023 ; Gandikota et al., 2023 ; Zhang et al., 2024d ; Sun et al., 2024 ; Chen et al., 2025d,b ; Facchiano et al., 2025) . When deletion requests or policy updates affect only part of the training signal, retraining from scratch is often impractical (Voigt and Von dem Bussche, 2017 ; Goldman, 2020) . Targeted removal is challenging because knowledge is distributed in shared representations, so eliminating one association can disrupt unrelated behavior. These challenges have driven growing interest in multimodal unlearning as a mec

FAQs

How does the survey organize multimodal unlearning methods across modalities?

It provides a unified, system-oriented taxonomy spanning vision, language, audio, and video, comparing approaches by deletion strength, retention, efficiency, reversibility, and robustness.

Jul 16, 2026

Upload by Theodore