mlsanityData to 14 Oct 2026

Stochastic Meta-Unlearning: Bridging Language Backbone and Multimodal Unlearning

arXiv 2607.18615 ↗First posted 21 Jul 2026cs.CL cs.LGcited by 0

On the map: Multimodal LLMs / vision-language models → MLLM jailbreaks and adversarial robustness

also in 1 more
Abstract (arXiv)

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. There is a surprising phenomenon when moving from single-modality unlearning to VLM unlearning: a target forgotten by the standalone language backbone can still be recovered when image information is given to the full VLM. This shows that text-only feedback is not enough for reliable VLM unlearning. Motivated by this observation, we propose Stochastic Meta-Unlearning (SMU), a bilevel framework that uses VLM-level feedback to learn an unlearning-ready initialization. In the inner loop, SMU applies a few unlearning steps to the language backbone using text data. In the outer loop, SMU recomposes the updated backbone with the frozen VLM and evaluates forgetting and utility at the VLM level. This design makes the unlearning update aware of the final multimodal behavior, while still keeping the update local to the language backbone. Experiments on two VLMs, two multimodal meme datasets, and three baselines show that SMU achieves the best overall forget-retain trade-off. Compared with the strongest baseline for each metric, SMU reduces average Forget accuracy by 10.52 points and improves average Retain and Test accuracy by 20.10 and 17.01 points, respectively. More importantly, SMU also transfers to new forgetting targets and to different meta-test unlearning methods. These results suggest that VLM-level feedback can make language-backbone unlearning more reliable and more transferable for VLMs.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 43 of its 49 references that are in the map. First posted on arXiv 21 Jul 2026.

Builds on

43 of 49 references.

Closest ancestors · 9

Same direction, newest first

  1. 2025PULSE: Practical Evaluation Scenarios for Large Multimodal Model Unlearning2 Jul
  2. 2025Unlearning Sensitive Information in Multimodal LLMs: Benchmark and Attack-Defense Evaluation1 May
  3. 2025SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoders16 Mar
  4. 2025Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Models21 Feb
  5. 2025SafeEraser: Enhancing Safety in Multimodal Large Language Models through Multimodal Machine Unlearning18 Feb
  6. 2025MMUnlearner: Reformulating Multimodal Machine Unlearning in the Era of Multimodal Large Language Models16 Feb
  7. 2024Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset5 Nov
  8. 2024Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench29 Oct
  9. 2024Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models21 May

Other references · 34

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2023Visual Instruction Tuningcited by 7,869
  3. 2023BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Modelscited by 6,137
  4. 2017Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networkscited by 4,897
  5. 2022Flamingo: a Visual Language Model for Few-Shot Learningcited by 4,386
  6. 2019Language Models as Knowledge Bases?cited by 1,936
  7. 2019Machine Unlearningcited by 1,061
  8. 2018On First-Order Meta-Learning Algorithmscited by 1,018
  9. 2019How Can We Know What Language Models Know?cited by 959
  10. 2024TOFU: A Task of Fictitious Unlearning for LLMscited by 533
  11. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 509
  12. 2024Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearningcited by 504
22 more
  1. 2020The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memescited by 490
  2. 2022Knowledge Unlearning for Mitigating Privacy Risks in Language Modelscited by 437
  3. 2023Who's Harry Potter? Approximate Unlearning in LLMscited by 381
  4. 2020Descent-to-Delete: Gradient-Based Methods for Machine Unlearningcited by 311
  5. 2023Large Language Model Unlearningcited by 302
  6. 2024MUSE: Machine Unlearning Six-Way Evaluation for Language Modelscited by 261
  7. 2024Rethinking Machine Unlearning for Large Language Modelscited by 257
  8. 2024Machine Unlearning of Pre-trained Large Language Modelscited by 119
  9. 2024Large Language Model Unlearning via Embedding-Corrupted Promptscited by 116
  10. 2024SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearningcited by 104
  11. 2024Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Differencecited by 88
  12. 2024Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearningcited by 84
  13. 2024RWKU: Benchmarking Real-World Knowledge Unlearning for Large Language Modelscited by 78
  14. 2024A Closer Look at Machine Unlearning for Large Language Modelscited by 56
  15. 2024Soft Prompting for Unlearning in Large Language Modelscited by 46
  16. 2021BERTnesia: Investigating the capture and forgetting of knowledge in BERTcited by 40
  17. 2024Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Conceptscited by 40
  18. 2024Does Unlearning Truly Unlearn? A Black Box Evaluation of LLM Unlearning Methodscited by 26
  19. 2024Learning to Unlearn for Robust Machine Unlearningcited by 19
  20. 2024Data Selection for Transfer Unlearningcited by 5
  21. 2024Unlearning Personal Data from a Single Imagecited by 4
  22. 2024Towards Transfer Unlearning: Empirical Evidence of Cross-Domain Bias Mitigationcited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026On the Robustness of Machine Unlearning for Vision-Language Models
  2. 2026Visual-Noise Guided In-Context Distillation for Multimodal Large Language Model Unlearning
  3. 2026One Modality to Forget Them All: Enhancing Cross-Modal Unlearning in Vision-Language Models
  4. 2026ICED: Concept-level Machine Unlearning via Interpretable Concept Decomposition
  5. 2026PPU-Bench:Real World Benchmark for Personalized Partial Unlearning in Vision Language Models
  6. 2025MLLMEraser: Achieving Test-Time Unlearning in Multimodal Large Language Models through Activation Steering
  7. 2025Adapting Vision-Language Models Without Labels: A Comprehensive Survey
  8. 2026Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks