Elamur Research⌕
☰
home / papers / 2026 / OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models

OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models

Paper · 2026 · Hao Zheng, Zirui Pang, Ling Li et al.

Summary

OFFSIDE is a benchmark for checking whether multimodal large language models (MLLMs) can be made to forget misinformation. The authors argue that earlier multimodal unlearning benchmarks rely on synthetic images, can only delete everything tied to an image, and ignore what happens after unlearning, for example during further training. They build their resource from real photos of football players, gathered via Google, with transfer-market rumors carried by text or embedded directly in images. The data is split into Forget, Retain, Test and Relearn sets. Four scenarios are tested: Complete Unlearning, Fine-grained Unlearning (removing image-specific private facts while keeping shared ones), Corrective Relearning and Unimodal (text-only) Unlearning. Five standard baselines are applied to Qwen2.5-VL models. No method reliably removes rumors. Forgotten content returns after brief relearning or under prompt attacks. Text-only unlearning misses multimodal rumors, and rumors inside images are hardest to remove. The apparent forgetting looks statistically like catastrophic forgetting. These results show that current multimodal unlearning is fragile and needs more robust, modality-aware methods.

Key points

  • Dataset scale: OFFSIDE contains 15.68K manually curated records covering 80 players from 20 clubs, with 640 images. Each image has 8 shared and 6 private VQA pairs.
  • Models and evaluation: The base models are Qwen2.5-VL-3B and Qwen2.5-VL-7B, first fine-tuned on Forget ∪ Retain. Evaluation uses classification, generation, fact scoring, MM-Bench and a GPT evaluator.
  • New problem settings: The authors say this is the first benchmark to target visual rumors, fine-grained targets, corrective relearning and multi-image entity association.
  • Method stability: Gradient-ascent-based methods behave inconsistently between the 3B and 7B models. PO, which avoids gradient ascent, is more stable across both.
  • Limitations: Visual rumors are scarce, with only one visual rumor per player. The paper diagnoses failures of existing methods but proposes no new algorithm. Code is at github.com/zh121800/OFFSIDE.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Harbin Institute of TechnologyHarbin · China158 works · 28,841 citations🎓University of Science and Technology of ChinaHefei · China108 works · 112,698 citations🎓The Hong Kong University of Science and Technology (Guangzhou)Guangzhou · China53 works · 1,367 citations🎓Zhejiang University of TechnologyHangzhou · China6 works · 98 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 works

🧭 Topics 1Research subjects and languages the work is about.Vision & language · 657 works

Cite

Cite as Hao Zheng, Zirui Pang, Ling Li, Zhijie Deng, Yuhan Pu, Zhaowei Zhu, Xiaobo Xia and Jiaheng Wei. 2026. OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.613

@inproceedings{zheng2026offside,
  title = {{OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models}},
  author = {Hao Zheng, Zirui Pang, Ling Li, Zhijie Deng, Yuhan Pu, Zhaowei Zhu, Xiaobo Xia and Jiaheng Wei},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.613},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.613}
}

References

53 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In ICCV, pages 2425–2433. find
  2. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. arXiv preprint arXiv:25…
  3. Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. find
  4. Junkai Chen, Zhijie Deng, Kening Zheng, Yibo Yan, Shuliang Liu, PeiJun Wu, Peijie Jiang, Jia Liu, and Xuming Hu. 2025. Safeeraser: Enhancing safety in multimodal large language models through multimodal machine unlearning. Preprint, arXiv:2502.12520. find
  5. Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. 2023. Can we edit multimodal large language models? arXiv preprint arXiv:2310.08475. find
  6. Quang-Vinh Dang. 2021. Right to be forgotten in the age of machine learning. In ICADS, pages 403–411. find
  7. Zhijie Deng, Chris Yuhao Liu, Zirui Pang, Xinlei He, Lei Feng, Qi Xuan, Zhaowei Zhu, and Jiaheng Wei. 2025. Guard: Generation-time llm unlearning via adaptive restriction and detection. arXiv preprint arXiv:2505.13312. — [arXiv:2505.13312](https://arxiv.org/abs/2505.13312) find
  8. Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. 2020. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397. — [arXiv:2006.07397](https://arxiv.org/abs/2006.07397) find
  9. Alexey Dontsov, Dmitrii Korzh, Alexey Zhavoronkin, Boris Mikheev, Denis Bobkov, Aibek Alanov, Oleg Y Rogov, Ivan Oseledets, and Elena Tutubalina. 2024. Clear: Character unlearning in textual and visual modalities. arXiv preprint arXiv:2410.18057. find
  10. R Eldan and M Russinovich. 2023. Who’s harry potter? approximate unlearning in llms, arxiv. arXiv preprint arXiv:2310.02238. find
  11. Chongyang Gao, Lixu Wang, Kaize Ding, Chenkai Weng, Xiao Wang, and Qi Zhu. 2024. On large language model continual unlearning. find
  12. Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904– 6913.
  13. Weiyang Guo, Zesheng Shi, Zeen Zhu, Yuan Zhou, Min Zhang, and Jing Li. 2026. Backdoors in rlvr: Jailbreak backdoors in llms from verifiable reward. Preprint, arXiv:2604.09748.
  14. Yanming Guo, Yu Liu, Ard Oerlemans, Songyang Lao, Song Wu, and Michael S Lew. 2016. Deep learning for visual understanding: A review. Neurocomputing, 187:27–48. find
  15. Xiangdong Hu, Yangyang Jiang, Qin Hu, and Xiaojun Jia. 2026. Gambit: A gamified jailbreak framework for multimodal large language models. Preprint, arXiv:2601.03416.
  16. Peichun Hua, Hao Li, Shanghao Shi, Zhiyuan Yu, and Ning Zhang. 2025. Rethinking jailbreak detection of large vision language models with representational contrastive scoring. arXiv preprint arXiv:2512.12069.
  17. Jiahao Huo, Yibo Yan, Xu Zheng, Yuanhuiyi Lyu, Xin Zou, Zhihua Wei, and Xuming Hu. 2025. Mmunlearner: Reformulating multimodal machine unlearning in the era of multimodal large language models. arXiv preprint arXiv:2502.11051. find
  18. Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, pages 2901–2910.
  19. Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. 2020. Analyzing and improving the image quality of stylegan. In IEEE/CVF conference, pages 8110–8119. find
  20. Tatsuki Kawakami, Kazuki Egashira, Atsuyuki Miyai, Go Irie, and Kiyoharu Aizawa. 2025. Pulse: Practical evaluation scenarios for large multimodal model unlearning. arXiv preprint arXiv:2507.01271. — [arXiv:2507.01271](https://arxiv.org/abs/2507.01271) find
  21. Black Forest Labs. 2024. Flux. https://github.com/ black-forest-labs/flux. ↗
  22. Jiaqi Li, Miaozeng Du, Chuanyi Zhang, Yongrui Chen, Nan Hu, Guilin Qi, Haiyun Jiang, Siyuan Cheng, and Bozhong Tian. 2024a. Mike: A new benchmark for fine-grained multimodal entity knowledge editing. arXiv preprint arXiv:2402.14835. find
  23. Jiaqi Li, Qianshan Wei, Chuanyi Zhang, Guilin Qi, Miaozeng Du, Yongrui Chen, Sheng Bi, and Fan Liu. 2024b. Single image unlearning: Efficient machine unlearning in multimodal large language models. In NeurIPS, pages 35414–35453. find
  24. Ling Li, Yu Ye, Bingchuan Jiang, and Wei Zeng. 2024c. Georeasoner: Geo-localization with reasoning in street views using a large vision-language model. In ICML. find
  25. Ling Li, Yao Zhou, Yuxuan Liang, Fugee Tsung, and Jiaheng Wei. 2025a. Recognition through reasoning: Reinforcing image geo-localization with large vision-language models. arXiv preprint arXiv:2506.14674. — [arXiv:2506.14674](https://arxiv.org/abs/2506.14674) find
  26. Xiaoyuan Li, Keqin Bao, Yubo Ma, Moxin Li, Wenjie Wang, Rui Men, Yichang Zhang, Fuli Feng, Dayiheng Liu, and Junyang Lin. 2025b. Mtr-bench: A comprehensive benchmark for multi-turn reasoning evaluation. arXiv preprint arXiv:2505.17123.
  27. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In ACL, pages 74–81.
  28. Bo Liu, Qiang Liu, and Peter Stone. 2022. Continual learning and private unlearning. In CoLLAs), pages 243–254. find
  29. Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, and 1 others. 2024a. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216–233.
  30. Yujian Liu, Yang Zhang, Tommi Jaakkola, and Shiyu Chang. 2024b. Revisiting who’s harry potter: Towards targeted unlearning from a causal intervention perspective. arXiv preprint arXiv:2407.16997. find
  31. Zheyuan Liu, Guangyao Dou, Mengzhao Jia, Zhaoxuan Tan, Qingkai Zeng, Yongle Yuan, and Meng Jiang. 2024c. Protecting privacy in multimodal large language models with mllmu-bench. arXiv preprint arXiv:2410.22108.
  32. Zheyuan Liu, Guangyao Dou, Zhaoxuan Tan, Yijun Tian, and Meng Jiang. 2024d. Machine unlearning in generative ai: A survey. arXiv preprint arXiv:2407.20516. — [arXiv:2407.20516](https://arxiv.org/abs/2407.20516) find
  33. Ximing Lu, Sean Welleck, Jack Hessel, Liwei Jiang, Lianhui Qin, Peter West, Prithviraj Ammanabrolu, and Yejin Choi. 2022. Quark: Controllable text generation with reinforced unlearning. In NeurIPS, pages 27591–27609.
  34. Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. 2024. Tofu: A task of fictitious unlearning for llms. arXiv preprint arXiv:2401.06121.
  35. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2023a. Locating and editing factual associations in gpt. find
  36. Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023b. Mass-editing memory in a transformer. find
  37. Zirui Pang, Hao Zheng, Zhijie Deng, Ling Li, Zixin Zhong, and Jiaheng Wei. 2025. Label smoothing improves gradient ascent in llm unlearning. arXiv preprint arXiv:2510.22376. — [arXiv:2510.22376](https://arxiv.org/abs/2510.22376) find
  38. Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Jiankang Deng, Bernhard Kainz, and Stefanos Zafeiriou. 2024. Arc2face: A foundation model for id-consistent human faces. In ECCV, pages 241–261. Springer. find
  39. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311–318.
  40. Vaidehi Patil, Peter Hase, and Mohit Bansal. 2023. Can sensitive information be deleted from llms? objectives for defending against extraction attacks. arXiv preprint arXiv:2309.17410. find
  41. Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. 2018. Film: Visual reasoning with a general conditioning layer. In AAAI.
  42. Yuhan Pu, Hao Zheng, Ziqian Mo, Hill Zhang, Tianyi Fan, Shuhong Wu, and Jiaheng Wei. 2026. Cameo: A conditional and quality-aware multi-agent image editing orchestrator. arXiv preprint arXiv:2604.03156. — [arXiv:2604.03156](https://arxiv.org/abs/2604.03156) find
  43. Yaxuan Wang, Jiaheng Wei, Chris Yuhao Liu, Jinlong Pang, Quan Liu, Ankit Parag Shah, Yujia Bao, Yang Liu, and Wei Wei. 2024b. Llm unlearning via loss adjustment with only forget data. arXiv preprint arXiv:2410.11143. — [arXiv:2410.11143](https://arxiv.org/abs/2410.11143) find
  44. Mika Westerlund. 2019. The emergence of deepfake technology: A review. Technology innovation management review, 9(11). find
  45. Haoming Xu, Ningyuan Zhao, Liming Yang, Sendong Zhao, Shumin Deng, Mengru Wang, Bryan Hooi, Nay Oo, Huajun Chen, and Ningyu Zhang. Relearn: Unlearning via learning for large language models. Zhaopan Xu, Pengfei Zhou, Weidong Tang, Jiaxin Ai, Wangbo Zhao, Xiaojiang Peng, Kai Wang, Yang You, Wenqi Sha… find
  46. Jin Yao, Eli Chien, Minxin Du, Xinyao Niu, Tianhao Wang, Zezhou Cheng, and Xiang Yue. 2024a. Machine unlearning of pre-trained large language models. arXiv preprint arXiv:2402.15159.
  47. Yuanshun Yao, Xiaojun Xu, and Yang Liu. 2024b. Large language model unlearning. In NeurIPS, pages 105425–105475.
  48. Chenlong Zhang, Zhuoran Jin, Hongbang Yuan, Jiaheng Wei, Tong Zhou, Kang Liu, Jun Zhao, and Yubo Chen. 2025. Rule: Reinforcement unlearning achieves forget-retain pareto optimality. find
  49. Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv preprint arXiv:2404.05868.
  50. Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A Smith, and Chiyuan Zhang. 2024. Muse: Machine unlearning six-way evaluation for language models. arXiv preprint arXiv:2407.06460. find
  51. Kozo Sugiyama, Shojiro Tagawa, and Mitsuhiko Toda. 2007. Methods for visual understanding of hierarchical system structures. IEEE Transactions on Systems, Man, and Cybernetics, 11(2):109–125. find
  52. Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. 2024a. A comprehensive survey of continual learning: Theory, method and application. IEEE transactions on pattern analysis and machine intelligence, 46(8):5362–5383. find
  53. Xi Wang, Songlei Jian, Shasha Li, Xiaopeng Li, Zhaoye Li, Bin Ji, Baosheng Wang, and Jie Yu. 2026. Jpu: Bridging jailbreak defense and unlearning via on-policy path rectification. Preprint, arXiv:2601.03005. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zheng-2026-offside-benchmarking-unlearning-misinformation-multimodal/