---
title: "LOTUS — paper by Wang et al. (2026) — summary and ideas"
description: "They argue the root cause is geometric: Euclidean space represents concept hierarchies poorly. 29 references."
authors: ["Zekun Wang", "Jingjie Zeng", "Yingxu Li", "Hongfei Lin", "Liang Yang"]
year: 2026
venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)"
source: "https://aclanthology.org/2026.acl-long.2195/"
doi: "10.18653/v1/2026.acl-long.2195"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/wang-2026-lotus-evolving-multimodal-unlearning-hyperbolic/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wang-2026-lotus-evolving-multimodal-unlearning-hyperbolic/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/35/) / LOTUS: Evolving Multimodal Unlearning via Hyperbolic Entailment and Lorentz Transport

# LOTUS: Evolving Multimodal Unlearning via Hyperbolic Entailment and Lorentz Transport

**Authors:** [Zekun Wang](https://md.research.elamur.ai/free/en/authors/zekun-wang/), [Jingjie Zeng](https://md.research.elamur.ai/free/en/authors/jingjie-zeng/), Yingxu Li, [Hongfei Lin](https://md.research.elamur.ai/free/en/authors/hongfei-lin/) and [Liang Yang](https://md.research.elamur.ai/free/en/authors/liang-yang/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.acl-long.2195/) · [PDF](https://aclanthology.org/2026.acl-long.2195.pdf) · [DOI 10.18653/v1/2026.acl-long.2195](https://doi.org/10.18653/v1/2026.acl-long.2195) · [OpenAlex](https://openalex.org/W7166819989)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
Multimodal LLMs can memorize sensitive material such as private identities or copyrighted images, and removing that knowledge on request is hard. Existing unlearning techniques either damage general abilities (catastrophic forgetting) or quietly swap the protected entity for a generic one, which the authors consider unsafe rather than real forgetting. They argue the root cause is geometric: Euclidean space represents concept hierarchies poorly. LOTUS addresses this by projecting features onto the Lorentz (hyperbolic) manifold, where general concepts sit near the origin and specific instances lie at the edges. It works in two stages. First, an Inverted Entailment Cone Loss pushes a sensitive instance outside the entailment cone of its parent concept. Second, Lorentz Transport, an Optimal Transport objective in the tangent space, moves the pruned features toward a safety-refusal distribution. On MLLMU-Bench, the unlearned model declines to name the person or reveal private facts but still describes what the image shows. This matters for privacy-compliant deployment, for example under GDPR and the EU AI Act.

## Key points
- Evaluated on MLLMU-Bench. Compared with constraint-based state-of-the-art methods on LLaVA, LOTUS improves unlearning efficacy by more than 9%. The conclusion explicitly names MANU as a method it surpasses.
- The safety refusal prior used in the transport stage is distilled from Qwen3-VL-Plus.
- Ablation results on LLaVA-1.5-7B (forget accuracy ↓ / retain accuracy ↑):

  | Variant | Forget acc. | Retain acc. |
  |---|---|---|
  | Full LOTUS | 36.85 | 43.18 |
  | Without hyperbolic space | 41.24 | 39.65 |
  | Without Lorentz Transport | 39.10 | 42.05 |

- Qualitative comparisons: GA produces degenerate, repetitive output, and NPO invents identities or addresses. LOTUS refuses the sensitive query but still describes visible features such as hair or glasses.
- Stated limitations:
  - LOTUS needs an additional hyperbolic projection module.
  - Its effectiveness depends on the quality of the safety anchor distribution generated by the teacher model.

*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

- [Dalian University of Technology](https://md.research.elamur.ai/free/en/universities/dalian-university-of-technology/) — University, Dalian, China · 12 works · 46 citations

📚 Shelves (2): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works

🧭 Topics (1): [🖼️ Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) · 657 works

## Cite

Cite as: Zekun Wang, Jingjie Zeng, Yingxu Li, Hongfei Lin and Liang Yang. 2026. LOTUS: Evolving Multimodal Unlearning via Hyperbolic Entailment and Lorentz Transport. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{wang2026lotus,
  title = {{LOTUS: Evolving Multimodal Unlearning via Hyperbolic Entailment and Lorentz Transport}},
  author = {Zekun Wang, Jingjie Zeng, Yingxu Li, Hongfei Lin and Liang Yang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.2195},
  url = {https://doi.org/10.18653/v1/2026.acl-long.2195}
}
```

- [← previous Lost in Translation: Repurposing semantic similarity benchmarks for evaluating lexical-semantic consistency in LLM-based machine translation](https://md.research.elamur.ai/free/en/papers/ye-2026-lost-translation-repurposing-semantic-similarity/)
- [next → Low-Rank Compression of Language Models via Differentiable Rank Selection](https://md.research.elamur.ai/free/en/papers/sundrani-2026-low-rank-compression-language-models/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/35/)

## References

29 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Michael C Anderson and Collin Green. 2001. Suppressing unwanted memories by executive control. Nature, 410(6826):366–369. [find](https://scholar.google.com/scholar?q=Michael+C+Anderson+and+Collin+Green%2E+2001%2E+Suppressing+unwanted+memories+by+executive+control%2E+Nature%2C+410%286826%29%3A366%E2%80%93369%2E)
2. Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. 2023. Hyperbolic image-text representations. ArXiv, abs/2304.09172. — [arXiv:2304.09172](https://arxiv.org/abs/2304.09172) [find](https://scholar.google.com/scholar?q=Karan+Desai%2C+Maximilian+Nickel%2C+Tanmay+Rajpurohit%2C+Justin+Johnson%2C+and+Ramakrishna+Vedantam%2E+2023%2E+Hyperbolic+image%2Dtext+representations%2E+ArXiv%2C+abs%2F2304%2E09172%2E+%E2%80%94+%5BarXiv%3A2304%2E09172%5D%28https%3A%2F%2Farxiv%2Eorg%2F)
3. Ronen Eldan and Mark Russinovich. 2023. Who’s harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238. [find](https://scholar.google.com/scholar?q=Ronen+Eldan+and+Mark+Russinovich%2E+2023%2E+Who%E2%80%99s+harry+potter%3F+approximate+unlearning+in+llms%2E+arXiv+preprint+arXiv%3A2310%2E02238%2E)
4. Europe. 2016. Regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (general data protection regulation). [https://eur-lex.europa.eu/eli/](https://eur-lex.europa.eu/eli/) reg/201… [↗](https://eur-lex.europa.eu/eli/)
5. Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. 2018. Hyperbolic entailment cones for learning hierarchical embeddings. ArXiv, abs/1804.01882. — [arXiv:1804.01882](https://arxiv.org/abs/1804.01882) [find](https://scholar.google.com/scholar?q=Octavian%2DEugen+Ganea%2C+Gary+B%C3%A9cigneul%2C+and+Thomas+Hofmann%2E+2018%2E+Hyperbolic+entailment+cones+for+learning+hierarchical+embeddings%2E+ArXiv%2C+abs%2F1804%2E01882%2E+%E2%80%94+%5BarXiv%3A1804%2E01882%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F1804)
6. Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2022. [Editing models with task arithmetic](https://md.research.elamur.ai/free/en/papers/ilharco-2022-editing-models-task-arithmetic/). arXiv preprint arXiv:2212.04089.
7. Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, et al. 2024. The wmdp benchmark: Measuring and reducing malicious use with unlearning. arXiv preprint arXiv:2403.03218. [find](https://scholar.google.com/scholar?q=Nathaniel+Li%2C+Alexander+Pan%2C+Anjali+Gopal%2C+Summer+Yue%2C+Daniel+Berrios%2C+Alice+Gatti%2C+Justin+D+Li%2C+Ann%2DKathrin+Dombrowski%2C+Shashwat+Goel%2C+Long+Phan%2C+et+al%2E+2024%2E+The+wmdp+benchmark%3A+Measuring+and+reduci)
8. Zongxia Li, Xiyang Wu, Hongyang Du, Huy Nghiem, and Guangyao Shi. 2025. Benchmark evaluations, applications, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189. — [arXiv:2501.02189](https://arxiv.org/abs/2501.02189) [find](https://scholar.google.com/scholar?q=Zongxia+Li%2C+Xiyang+Wu%2C+Hongyang+Du%2C+Huy+Nghiem%2C+and+Guangyao+Shi%2E+2025%2E+Benchmark+evaluations%2C+applications%2C+and+challenges+of+large+vision+language+models%3A+A+survey%2E+arXiv+preprint+arXiv%3A2501%2E02189%2E+)
9. Bo Liu, Qiang Liu, and Peter Stone. 2022. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243–254. PMLR. [find](https://scholar.google.com/scholar?q=Bo+Liu%2C+Qiang+Liu%2C+and+Peter+Stone%2E+2022%2E+Continual+learning+and+private+unlearning%2E+In+Conference+on+Lifelong+Learning+Agents%2C+pages+243%E2%80%93254%2E+PMLR%2E)
10. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024a. [Visual instruction tuning](https://md.research.elamur.ai/free/en/papers/liu-2023-visual-instruction-tuning/). Advances in neural information processing systems, 36.
11. 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](https://scholar.google.com/scholar?q=Yujian+Liu%2C+Yang+Zhang%2C+Tommi+Jaakkola%2C+and+Shiyu+Chang%2E+2024b%2E+Revisiting+who%E2%80%99s+harry+potter%3A+Towards+targeted+unlearning+from+a+causal+intervention+perspective%2E+arXiv+preprint+arXiv%3A2407%2E16997%2E)
12. 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](https://md.research.elamur.ai/free/en/papers/liu-2025-protecting-privacy-multimodal-large-language/). arXiv preprint arXiv:2410.22108.
13. Lingzhi Wang, Tong Chen, Wei Yuan, Xingshan Zeng, Kam-Fai Wong, and Hongzhi Yin. 2023. Kga: A general machine unlearning framework based on knowledge gap alignment. arXiv preprint arXiv:2305.06535. [find](https://scholar.google.com/scholar?q=Lingzhi+Wang%2C+Tong+Chen%2C+Wei+Yuan%2C+Xingshan+Zeng%2C+Kam%2DFai+Wong%2C+and+Hongzhi+Yin%2E+2023%2E+Kga%3A+A+general+machine+unlearning+framework+based+on+knowledge+gap+alignment%2E+arXiv+preprint+arXiv%3A2305%2E06535%2E)
14. Zheyuan Liu, Guangyao Dou, Xiangchi Yuan, Chunhui Zhang, Zhaoxuan Tan, and Meng Jiang. 2025. [Modality-aware neuron pruning for unlearning in multimodal large language models](https://md.research.elamur.ai/free/en/papers/liu-2025-modality-aware-neuron-pruning-unlearning/). In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 591…
15. Xinwei Wu, Junzhuo Li, Minghui Xu, Weilong Dong, Shuangzhi Wu, Chao Bian, and Deyi Xiong. 2023. Depn: Detecting and editing privacy neurons in pretrained language models. arXiv preprint arXiv:2310.20138. [find](https://scholar.google.com/scholar?q=Xinwei+Wu%2C+Junzhuo+Li%2C+Minghui+Xu%2C+Weilong+Dong%2C+Shuangzhi+Wu%2C+Chao+Bian%2C+and+Deyi+Xiong%2E+2023%2E+Depn%3A+Detecting+and+editing+privacy+neurons+in+pretrained+language+models%2E+arXiv+preprint+arXiv%3A2310%2E201)
16. Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary C Lipton, and J Zico Kolter. 2024. [Tofu: A task of fictitious unlearning for llms](https://md.research.elamur.ai/free/en/papers/maini-2024-tofu-task-fictitious-unlearning-llms/). arXiv preprint arXiv:2401.06121.
17. Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. 2020. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025–16036. [find](https://scholar.google.com/scholar?q=Quoc+Phong+Nguyen%2C+Bryan+Kian+Hsiang+Low%2C+and+Patrick+Jaillet%2E+2020%2E+Variational+bayesian+unlearning%2E+Advances+in+Neural+Information+Processing+Systems%2C+33%3A16025%E2%80%9316036%2E)
18. Maximilian Nickel and Douwe Kiela. 2017. [Poincaré embeddings for learning hierarchical representations](https://md.research.elamur.ai/free/en/papers/nickel-2017-poincare-embeddings-learning-hierarchical-representations/). ArXiv, abs/1705.08039.
19. Avik Pal, Max van Spengler, Guido Maria D’Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. 2025. Compositional entailment learning for hyperbolic vision-language models. In The Thirteenth International Conference on Learning Representations. [find](https://scholar.google.com/scholar?q=Avik+Pal%2C+Max+van+Spengler%2C+Guido+Maria+D%E2%80%99Amely+di+Melendugno%2C+Alessandro+Flaborea%2C+Fabio+Galasso%2C+and+Pascal+Mettes%2E+2025%2E+Compositional+entailment+learning+for+hyperbolic+vision%2Dlanguage+models%2E+In+)
20. Karalyn E Patterson, Peter J. Nestor, and Timothy T. Rogers. 2007. Where do you know what you know? the representation of semantic knowledge in the human brain. Nature Reviews Neuroscience, 8:976– 987. [find](https://scholar.google.com/scholar?q=Karalyn+E+Patterson%2C+Peter+J%2E+Nestor%2C+and+Timothy+T%2E+Rogers%2E+2007%2E+Where+do+you+know+what+you+know%3F+the+representation+of+semantic+knowledge+in+the+human+brain%2E+Nature+Reviews+Neuroscience%2C+8%3A976%E2%80%93+987)
21. Gabriel Peyré and Marco Cuturi. 2018. Computational optimal transport. Found. Trends Mach. Learn., 11:355–607. [find](https://scholar.google.com/scholar?q=Gabriel+Peyr%C3%A9+and+Marco+Cuturi%2E+2018%2E+Computational+optimal+transport%2E+Found%2E+Trends+Mach%2E+Learn%2E%2C+11%3A355%E2%80%93607%2E)
22. Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. 2024. [Mllm-protector: Ensuring mllm’s safety without hurting performance](https://md.research.elamur.ai/free/en/papers/pi-2024-mllm-protector-ensuring-mllm-s/). arXiv preprint arXiv:2401.02906.
23. Nianwen Si, Hao Zhang, Heyu Chang, Wenlin Zhang, Dan Qu, and Weiqiang Zhang. 2023. Knowledge unlearning for llms: Tasks, methods, and challenges. arXiv preprint arXiv:2311.15766. — [arXiv:2311.15766](https://arxiv.org/abs/2311.15766) [find](https://scholar.google.com/scholar?q=Nianwen+Si%2C+Hao+Zhang%2C+Heyu+Chang%2C+Wenlin+Zhang%2C+Dan+Qu%2C+and+Weiqiang+Zhang%2E+2023%2E+Knowledge+unlearning+for+llms%3A+Tasks%2C+methods%2C+and+challenges%2E+arXiv+preprint+arXiv%3A2311%2E15766%2E+%E2%80%94+%5BarXiv%3A2311%2E15766%5D%28)
24. Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. 2022. Unrolling sgd: Understanding factors influencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303–319. IEEE. [find](https://scholar.google.com/scholar?q=Anvith+Thudi%2C+Gabriel+Deza%2C+Varun+Chandrasekaran%2C+and+Nicolas+Papernot%2E+2022%2E+Unrolling+sgd%3A+Understanding+factors+influencing+machine+unlearning%2E+In+2022+IEEE+7th+European+Symposium+on+Security+and+P)
25. Cédric Villani et al. 2008. Optimal transport: old and new, volume 338. Springer. [find](https://scholar.google.com/scholar?q=C%C3%A9dric+Villani+et+al%2E+2008%2E+Optimal+transport%3A+old+and+new%2C+volume+338%2E+Springer%2E)
26. Yibo Yan, Jiamin Su, Jianxiang He, Fangteng Fu, Xu Zheng, Yuanhuiyi Lyu, Kun Wang, Shen Wang, Qingsong Wen, and Xuming Hu. 2024. A survey of mathematical reasoning in the era of multimodal large language model: Benchmark, method & challenges. arXiv preprint arXiv:2412.11936. [find](https://scholar.google.com/scholar?q=Yibo+Yan%2C+Jiamin+Su%2C+Jianxiang+He%2C+Fangteng+Fu%2C+Xu+Zheng%2C+Yuanhuiyi+Lyu%2C+Kun+Wang%2C+Shen+Wang%2C+Qingsong+Wen%2C+and+Xuming+Hu%2E+2024%2E+A+survey+of+mathematical+reasoning+in+the+era+of+multimodal+large+langu)
27. Yibo Yan, Shen Wang, Jiahao Huo, Jingheng Ye, Zhendong Chu, Xuming Hu, Philip S Yu, Carla Gomes, Bart Selman, and Qingsong Wen. 2025. Position: Multimodal large language models can significantly advance scientific reasoning. arXiv preprint arXiv:2502.02871. [find](https://scholar.google.com/scholar?q=Yibo+Yan%2C+Shen+Wang%2C+Jiahao+Huo%2C+Jingheng+Ye%2C+Zhendong+Chu%2C+Xuming+Hu%2C+Philip+S+Yu%2C+Carla+Gomes%2C+Bart+Selman%2C+and+Qingsong+Wen%2E+2025%2E+Position%3A+Multimodal+large+language+models+can+significantly+advan)
28. Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. [Negative preference optimization: From catastrophic collapse to effective unlearning](https://md.research.elamur.ai/free/en/papers/zhang-2024-negative-preference-optimization-catastrophic-collapse/). arXiv preprint arXiv:2404.05868.
29. Xingchen Zou, Yibo Yan, Xixuan Hao, Yuehong Hu, Haomin Wen, Erdong Liu, Junbo Zhang, Yong Li, Tianrui Li, Yu Zheng, et al. 2025. Deep learning for cross-domain data fusion in urban computing: Taxonomy, advances, and outlook. Information Fusion, 113:102606. [find](https://scholar.google.com/scholar?q=Xingchen+Zou%2C+Yibo+Yan%2C+Xixuan+Hao%2C+Yuehong+Hu%2C+Haomin+Wen%2C+Erdong+Liu%2C+Junbo+Zhang%2C+Yong+Li%2C+Tianrui+Li%2C+Yu+Zheng%2C+et+al%2E+2025%2E+Deep+learning+for+cross%2Ddomain+data+fusion+in+urban+computing%3A+Taxonomy)

*Updated 1 November 2026*