arxiv: 2605.05909 · v1 · submitted 2026-05-07 · 💻 cs.AI

Recognition: unknown

Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning

Yuhang Wang , Zhenxing Niu , Haoxuan Ji , Guangyu He , Linlin Zhang , Haichang Gao

Authors on Pith no claims yet

Pith reviewed 2026-05-08 11:01 UTC · model grok-4.3

classification 💻 cs.AI

keywords machine unlearningmultimodal large language modelsvisual forgettingnull space constraintcontrastive learningknowledge retentioncontinual unlearning

0 comments

The pith

Freezing the LLM backbone and constraining visual module updates to the null space of retained knowledge lets MLLMs forget target visual concepts while keeping all other knowledge intact.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tackles the intertwined visual and textual knowledge in multimodal large language models by proposing a targeted unlearning method. It freezes the language model backbone and updates only the visual module to remove specific visual knowledge. A contrastive visual forgetting step separates target concepts in feature space, while null space constraints keep those updates from affecting retained visuals or any text. The same technique handles both one-off and sequential forgetting requests. A sympathetic reader would care because it offers a modular way to address privacy or safety issues in deployed models without full retraining or broad capability loss.

Core claim

We introduce Null Space Constrained Contrastive Visual Forgetting, which achieves unlearning by fine-tuning only the visual module while the LLM backbone remains frozen; contrastive visual forgetting guides target visual representations toward appropriate feature-space regions, and the null space of retained knowledge constrains the updates so that target visual knowledge is removed without degrading non-target visuals or any textual knowledge, with the method also extending to continual unlearning.

What carries the argument

Null Space Constrained Contrastive Visual Forgetting, which uses contrastive separation of target visual representations together with projection of all updates into the null space associated with retained knowledge.

Load-bearing premise

Fine-tuning only the visual module while freezing the LLM backbone is sufficient to remove target visual knowledge without degrading non-target visual knowledge or any textual knowledge.

What would settle it

A controlled test in which the unlearned model still produces accurate outputs for queries about the target visual concepts or shows measurable accuracy drops on non-target visual or textual benchmarks would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.05909 by Guangyu He, Haichang Gao, Haoxuan Ji, Linlin Zhang, Yuhang Wang, Zhenxing Niu.

**Figure 1.** Figure 1: Overview of our CVF mechanism in MLLM unlearning. In our approach, we use the original model as a reference model and take its Intermediate Visual Representations (IVRs) as supervision signals. As illustrated in view at source ↗

**Figure 2.** Figure 2: Evaluation of continual unlearning with five sequential forgetting tasks. (a) Average VQA view at source ↗

**Figure 3.** Figure 3: Stage-wise performance heatmaps of different unlearning methods. The color intensity view at source ↗

**Figure 4.** Figure 4: Hyperparameter sensitivity of α and β. Performance trends on Forget/Retain/Real-World VQA when varying α and β while keeping other settings fixed. Since λ is an inner-level balancing factor that primarily affects the stability of CVF, we select λ via an automated validation-based tuning procedure. Concretely, we perform a lightweight search over a small candidate set and choose the value that best maintain… view at source ↗

**Figure 5.** Figure 5: Qualitative results of GA, MANU, MMUNLEARNER, and our method on Forget/Retain view at source ↗

read the original abstract

The core challenge of machine unlearning is to strike a balance between target knowledge removal and non-target knowledge retention. In the context of Multimodal Large Language Models (MLLMs), this challenge becomes even more pronounced, as knowledge is further divided into visual and textual modalities that are tightly intertwined. In this paper, we introduce an MLLM unlearning approach that aims to forget target visual knowledge while preserving non-target visual knowledge and all textual knowledge. Specifically, we freeze the LLM backbone and achieve unlearning by fine-tuning the visual module. First, we propose a Contrastive Visual Forgetting (CVF) mechanism to separate target visual knowledge from retained visual knowledge, guiding the representations of target visual concepts toward appropriate regions in the feature space. Second, we identify the null space associated with retained knowledge and constrain the unlearning process within this space, thereby significantly mitigating degradation in knowledge retention. Third, beyond static unlearning scenarios, we extend our approach to continual unlearning, where forgetting requests arrive sequentially. Extensive experiments across diverse benchmarks demonstrate that our approach achieves a strong balance between effective forgetting and robust knowledge retention.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper tries to unlearn specific visual knowledge in MLLMs by fine-tuning only the vision module with contrastive separation and null-space projection while freezing the LLM, but the isolation claim looks fragile on the evidence given.

read the letter

The main thing to know is that this work targets visual-only unlearning in multimodal models by keeping the language backbone frozen and updating just the visual encoder through a contrastive forgetting step plus a null-space constraint on retained knowledge. They also sketch an extension to sequential unlearning requests. The combination of those two mechanisms for the visual modality and the continual setting is the clearest new piece relative to prior unlearning work on single-modality models. The motivation section does a straightforward job explaining why modality entanglement makes standard forgetting harder here, and the high-level experimental claim is that forgetting succeeds while retention on other visuals and all text holds up across the benchmarks they test. That kind of practical framing is useful for people who actually deploy these models and face removal requests. The soft spot is the central premise that freezing the LLM and constraining updates to the visual null space will keep side effects from reaching non-target visuals or text. The LLM was trained on the original visual feature distribution, so even small shifts in the encoder outputs can change the statistics the frozen weights see and degrade downstream behavior. In high-dimensional visual space with limited unlearning samples, recovering an exact null space is also unlikely to be clean, which opens the door to leakage. If the full paper has strong ablations and retention metrics that directly test this isolation, the results would carry more weight; otherwise the balance they report may not generalize. This is for researchers working on machine unlearning and safety for deployed multimodal systems who need concrete methods rather than theory. A reader who wants empirical techniques for continual forgetting would get value from the setup even if they end up tweaking the constraints. It deserves a serious referee because the problem is timely and the proposal is specific enough to evaluate, though reviewers will need to press on whether the freezing step actually delivers the claimed separation.

Referee Report

3 major / 2 minor

Summary. The paper proposes Null Space Constrained Contrastive Visual Forgetting (CVF) for unlearning target visual knowledge in Multimodal Large Language Models (MLLMs). It freezes the LLM backbone and fine-tunes only the visual module: CVF separates target from retained visual representations via contrastive guidance in feature space, while updates are constrained to the null space of retained-knowledge gradients to protect non-target visual and all textual knowledge. The method is extended to continual unlearning with sequential forgetting requests, and the abstract claims extensive experiments on diverse benchmarks demonstrate an effective balance between forgetting and retention.

Significance. If the empirical claims hold, the work offers a computationally efficient, modality-aware unlearning technique for MLLMs that avoids full-model retraining. The null-space projection idea provides a principled way to mitigate retention degradation during targeted forgetting and could generalize to other continual or selective unlearning settings in large multimodal models.

major comments (3)

[§3 (Method)] The load-bearing assumption (stated in the abstract and §3) that freezing the LLM while fine-tuning only the visual module isolates target visual forgetting without side-effects is not obviously true. Because the LLM was trained on the original visual feature distribution, any shift induced by CVF—even when projected into a null space—can alter input statistics to the frozen weights and degrade non-target visual or textual performance; this requires explicit ablation on retention metrics for both modalities.
[§3.2] §3.2 (Null-space identification): The claim that the null space of retained-knowledge gradients can be accurately recovered from limited unlearning data is central to preventing leakage, yet high-dimensional visual feature spaces make exact orthogonal complement recovery unlikely. Without a concrete procedure (e.g., how many samples, how gradients are aggregated, or error bounds), it is unclear whether the constraint actually blocks updates into retained directions.
[Experiments section] The abstract asserts 'extensive experiments across diverse benchmarks' and a 'strong balance,' but supplies no quantitative metrics, baselines, ablation tables, or implementation details. Without these, the central empirical claim cannot be evaluated and the soundness of the method remains unverifiable.

minor comments (2)

[Abstract / §3.1] The description of 'appropriate regions in the feature space' for target concepts is vague; a figure or equation showing the target vs. retained separation would clarify the CVF objective.
[§3] Notation for the null-space projection operator and the contrastive loss should be defined consistently before use in equations.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, clarifying aspects of the method and committing to revisions that strengthen the presentation and empirical support.

read point-by-point responses

Referee: [§3 (Method)] The load-bearing assumption (stated in the abstract and §3) that freezing the LLM while fine-tuning only the visual module isolates target visual forgetting without side-effects is not obviously true. Because the LLM was trained on the original visual feature distribution, any shift induced by CVF—even when projected into a null space—can alter input statistics to the frozen weights and degrade non-target visual or textual performance; this requires explicit ablation on retention metrics for both modalities.

Authors: We appreciate the referee's emphasis on this potential issue. The design of CVF with null-space projection is intended to avoid updates that would degrade retained directions, and the manuscript already reports retention results on both non-target visual tasks and textual benchmarks to support that side-effects are limited. To make the isolation explicit, we will add a dedicated ablation subsection and table in the revised §3 and Experiments that directly compares retention metrics (for both modalities) with and without the null-space constraint. revision: yes
Referee: [§3.2] §3.2 (Null-space identification): The claim that the null space of retained-knowledge gradients can be accurately recovered from limited unlearning data is central to preventing leakage, yet high-dimensional visual feature spaces make exact orthogonal complement recovery unlikely. Without a concrete procedure (e.g., how many samples, how gradients are aggregated, or error bounds), it is unclear whether the constraint actually blocks updates into retained directions.

Authors: This is a valid request for implementation details. We will expand §3.2 in the revision to specify the exact procedure: the number of retained samples used to compute gradients, the aggregation method (averaging), the numerical technique for recovering the orthogonal complement, and a short analysis of approximation quality via singular-value thresholds. This will allow readers to assess the reliability of the constraint. revision: yes
Referee: [Experiments section] The abstract asserts 'extensive experiments across diverse benchmarks' and a 'strong balance,' but supplies no quantitative metrics, baselines, ablation tables, or implementation details. Without these, the central empirical claim cannot be evaluated and the soundness of the method remains unverifiable.

Authors: We acknowledge that the experimental presentation must be fully self-contained. The manuscript's Experiments section contains quantitative results, baseline comparisons, and ablation studies across the reported benchmarks; however, we will reorganize and expand this section in the revision to include all requested elements—complete metric tables, baseline descriptions, hyperparameter details, and data splits—so that the claims in the abstract are directly verifiable from the text. revision: yes

Circularity Check

0 steps flagged

Empirical method with no derivation chain or self-referential reductions

full rationale

The paper describes an empirical unlearning technique: freezing the LLM backbone, fine-tuning only the visual module via Contrastive Visual Forgetting (CVF) to separate target from retained visual knowledge, plus a null-space constraint on retained-knowledge directions. No equations, closed-form derivations, or first-principles predictions appear in the provided abstract or method summary. Claims of balance between forgetting and retention are supported solely by experimental results on benchmarks rather than any reduction of outputs to fitted inputs or self-citations by construction. The approach is presented as a practical algorithm validated externally, with no load-bearing steps that equate to their own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no technical equations or implementation details, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5509 in / 1071 out tokens · 36812 ms · 2026-05-08T11:01:01.867386+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 15 canonical work pages · 3 internal anchors

[1]

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

work page internal anchor Pith review arXiv 2024
[2]

Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

2023
[3]

Pencil: Long thoughts with short memory.arXiv preprint arXiv:2503.14337, 2025

Chenxiao Yang, Nathan Srebro, David McAllester, and Zhiyuan Li. Pencil: Long thoughts with short memory.arXiv preprint arXiv:2503.14337, 2025

work page arXiv 2025
[4]

On path to multimodal generalist: General-level and general-bench

Hao Fei, Yuan Zhou, Juncheng Li, Xiangtai Li, Qingshan Xu, Bobo Li, Shengqiong Wu, Yaoting Wang, Junbao Zhou, Jiahao Meng, et al. On path to multimodal generalist: General-level and general-bench. InForty-second International Conference on Machine Learning, 2025

2025
[5]

Mmunlearner: Reformulating multimodal machine unlearning in the era of multimodal large language models.arXiv preprint arXiv:2502.11051, 2025

Jiahao Huo, Yibo Yan, Xu Zheng, Yuanhuiyi Lyu, Xin Zou, Zhihua Wei, and Xuming Hu. Mmunlearner: Reformulating multimodal machine unlearning in the era of multimodal large language models.arXiv preprint arXiv:2502.11051, 2025

work page arXiv 2025
[6]

Towards safer large language models through machine unlearning.arXiv preprint arXiv:2402.10058, 2024

Zheyuan Liu, Guangyao Dou, Zhaoxuan Tan, Yijun Tian, and Meng Jiang. Towards safer large language models through machine unlearning.arXiv preprint arXiv:2402.10058, 2024

work page arXiv 2024
[7]

Clear: Character unlearning in textual and visual modalities

Alexey Dontsov, Dmitrii Korzh, Alexey Zhavoronkin, Boris Mikheev, Denis Bobkov, Aibek Alanov, Oleg Rogov, Ivan Oseledets, and Elena Tutubalina. Clear: Character unlearning in textual and visual modalities. InFindings of the Association for Computational Linguistics: ACL 2025, pages 20582–20603, 2025

2025
[8]

International conference on machine learning.Transactions on machine learning research, 2023

Wenjie Li, Chi-hua Wang, Guang Cheng, and Qifan Song. International conference on machine learning.Transactions on machine learning research, 2023

2023
[9]

Towards making systems forget with machine unlearning

Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463–480. IEEE, 2015

2015
[10]

Adaptive machine unlearning.Advances in Neural Information Processing Systems, 34: 16319–16330, 2021

Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning.Advances in Neural Information Processing Systems, 34: 16319–16330, 2021

2021
[11]

Remember what you want to forget: Algorithms for machine unlearning.Advances in Neural Information Processing Systems, 34:18075–18086, 2021

Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning.Advances in Neural Information Processing Systems, 34:18075–18086, 2021

2021
[12]

Machine unlearning via algorithmic stability

Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. InConference on Learning Theory, pages 4126–4142. PMLR, 2021

2021
[13]

TOFU: A Task of Fictitious Unlearning for LLMs

Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. Tofu: A task of fictitious unlearning for llms.arXiv preprint arXiv:2401.06121, 2024

work page internal anchor Pith review arXiv 2024
[14]

Direct preference optimization: Your language model is secretly a reward model

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023

2023
[15]

Negative preference optimization: From catastrophic collapse to effective un- learning.arXiv preprint arXiv:2404.05868,

Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning.arXiv preprint arXiv:2404.05868, 2024. 10

work page arXiv 2024
[16]

Gru: Miti- gating the trade-off between unlearning and retention for llms.arXiv preprint arXiv:2503.09117, 2025

Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, and Bo Han. Gru: Miti- gating the trade-off between unlearning and retention for llms.arXiv preprint arXiv:2503.09117, 2025

work page arXiv 2025
[17]

Mllm machine unlearning via visual knowledge distillation.arXiv preprint arXiv:2512.11325, 2025

Yuhang Wang, Zhenxing Niu, Haoxuan Ji, Guangyu He, Haichang Gao, and Gang Hua. Mllm machine unlearning via visual knowledge distillation.arXiv preprint arXiv:2512.11325, 2025

work page arXiv 2025
[18]

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, and Tat-Seng Chua. Alphaedit: Null-space constrained knowledge editing for language models. arXiv preprint arXiv:2410.02355, 2024

work page arXiv 2024
[19]

Cmmlu: Measuring massive multitask language understanding in chinese

Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. Cmmlu: Measuring massive multitask language understanding in chinese. InFindings of the Association for Computational Linguistics: ACL 2024, pages 11260–11285, 2024

2024
[20]

Evaluating large language models in class-level code generation

Xueying Du, Mingwei Liu, Kaixin Wang, Hanlin Wang, Junwei Liu, Yixuan Chen, Jiayi Feng, Chaofeng Sha, Xin Peng, and Yiling Lou. Evaluating large language models in class-level code generation. InProceedings of the IEEE/ACM 46th International Conference on Software Engineering, pages 1–13, 2024

2024
[21]

Unrolling sgd: Understanding factors influencing machine unlearning

Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors influencing machine unlearning. In2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303–319. IEEE, 2022

2022
[22]

Commonsense knowledge editing based on free-text in llms.arXiv preprint arXiv:2410.23844, 2024

Xiusheng Huang, Yequan Wang, Jun Zhao, and Kang Liu. Commonsense knowledge editing based on free-text in llms.arXiv preprint arXiv:2410.23844, 2024

work page arXiv 2024
[23]

Understanding multimodal llms: the mechanistic inter- pretability of llava in visual question answering.arXiv preprint arXiv:2411.10950, 2024

Zeping Yu and Sophia Ananiadou. Understanding multimodal llms: the mechanistic inter- pretability of llava in visual question answering.arXiv preprint arXiv:2411.10950, 2024

work page arXiv 2024
[24]

Representation Learning with Contrastive Predictive Coding

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

work page internal anchor Pith review arXiv 2018
[25]

Lora: Low-rank adaptation of large language models.ICLR, 1 (2):3, 2022

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1 (2):3, 2022

2022
[26]

Protecting privacy in multimodal large language models with mllmu-bench

Zheyuan Liu, Guangyao Dou, Mengzhao Jia, Zhaoxuan Tan, Qingkai Zeng, Yongle Yuan, and Meng Jiang. Protecting privacy in multimodal large language models with mllmu-bench. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages ...

2025
[27]

Umu-bench: Closing the modality gap in multimodal unlearning evaluation

Chengye Wang, Yuyuan Li, XiaoHua Feng, Chaochao Chen, Xiaolin Zheng, and Jianwei Yin. Umu-bench: Closing the modality gap in multimodal unlearning evaluation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025

2025
[28]

Continual learning and private unlearning

Bo Liu, Qiang Liu, and Peter Stone. Continual learning and private unlearning. InConference on Lifelong Learning Agents, pages 243–254. PMLR, 2022

2022
[29]

Variational bayesian unlearning.Advances in Neural Information Processing Systems, 33:16025–16036, 2020

Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning.Advances in Neural Information Processing Systems, 33:16025–16036, 2020

2020
[30]

Modality-aware neuron pruning for unlearning in multimodal large language models.arXiv preprint arXiv:2502.15910, 2025

Zheyuan Liu, Guangyao Dou, Xiangchi Yuan, Chunhui Zhang, Zhaoxuan Tan, and Meng Jiang. Modality-aware neuron pruning for unlearning in multimodal large language models.arXiv preprint arXiv:2502.15910, 2025

work page arXiv 2025
[31]

Rouge: A package for automatic evaluation of summaries

Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004

2004
[32]

Mmneuron: Discovering neuron-level domain-specific interpretation in multimodal large language model.arXiv preprint arXiv:2406.11193, 2024

Jiahao Huo, Yibo Yan, Boren Hu, Yutao Yue, and Xuming Hu. Mmneuron: Discovering neuron-level domain-specific interpretation in multimodal large language model.arXiv preprint arXiv:2406.11193, 2024. 11

work page arXiv 2024
[33]

Fast exact unlearning for in-context learning data for llms

Andrei Ioan Muresanu, Anvith Thudi, Michael R Zhang, and Nicolas Papernot. Fast exact unlearning for in-context learning data for llms. InForty-second International Conference on Machine Learning
[34]

Large language model unlearning via embedding-corrupted prompts.Advances in Neural Information Processing Systems, 37: 118198–118266, 2024

Chris Liu, Yaxuan Wang, Jeffrey Flanigan, and Yang Liu. Large language model unlearning via embedding-corrupted prompts.Advances in Neural Information Processing Systems, 37: 118198–118266, 2024

2024
[35]

Performance gap in entity knowledge extraction across modalities in vision language models

Ido Cohen, Daniela Gottesman, Mor Geva, and Raja Giryes. Performance gap in entity knowledge extraction across modalities in vision language models. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29095–29108, 2025

2025
[36]

Single image unlearning: Efficient machine unlearning in multimodal large language models.Advances in Neural Information Processing Systems, 37:35414–35453, 2024

Jiaqi Li, Qianshan Wei, Chuanyi Zhang, Guilin Qi, Miaozeng Du, Yongrui Chen, Sheng Bi, and Fan Liu. Single image unlearning: Efficient machine unlearning in multimodal large language models.Advances in Neural Information Processing Systems, 37:35414–35453, 2024

2024
[37]

Sauce: Selective concept unlearning in vision-language models with sparse autoencoders

Jiahui Geng and Qing Li. Sauce: Selective concept unlearning in vision-language models with sparse autoencoders. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3023–3033, 2025

2025
[38]

arXiv preprint arXiv:2310.12508 (2023)

Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation.arXiv preprint arXiv:2310.12508, 2023. 12 A Motivation and Problem Formulation Recent literature presents two competing definitions of MLLM unlearning: (i) forget...

work page arXiv 2023