Elamur Research⌕
☰
home / papers / 2026 / Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problem

Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problem

Paper · 2026 · Zeguan Xiao, Siqing Li, Yong Wang et al.

Summary

LLM unlearning aims to erase targeted knowledge, such as hazardous biosecurity or cybersecurity facts or personal data, while keeping the model generally useful. In practice, the forgetting and retention updates tend to pull parameters in opposite directions. The authors argue this is not an ordinary balanced multi-task setup. Preserving capability is the main goal, and forgetting is a secondary task that must avoid causing damage.

They propose a modular training loop that computes separate forget and retain gradients and then merges them with a conflict-aware rule. They study two merging rules:

  • a module-wise adaptation of PCGrad, which removes conflicting forget components by projection;
  • SAGO, a new method that keeps only gradient elements whose signs match the retain gradient.

They prove that neither rule ever produces an update opposing the retain direction, and that SAGO aligns more tightly with it. The framework is plugged into GA+GD, NPO+GD and SimNPO+GD and evaluated on WMDP and RWKU. SAGO preserves far more general ability at similar forgetting strength. This suggests that reshaping gradient geometry matters more than reweighting losses.

Key points

  • On WMDP Bio with SimNPO+GD, recovery of the original model's MMLU score rises from 44.6% (naive) to 94.0% with PCGrad and 96.0% with SAGO, with comparable forgetting.
  • Across WMDP settings, SAGO improves MMLU over the naive objectives by 17.8–30.7 points on Bio and 4.1–11.7 points on Cyber. It adds a further 0.4–1.2 points beyond PCGrad, and similar gains are reported on RWKU.
  • PCGrad is applied per module rather than over one flattened parameter vector, unlike the original PCGrad and the GRU unlearning method.
  • A gradient-geometry analysis on WMDP Cyber (cosine similarities averaged over 100 training steps) shows that SAGO has the weakest forget–retain conflict and the highest alignment of the combined update with the retain gradient.
  • Stated limitations are that only two benchmarks are covered and that computing separate gradients adds overhead. Code is available at github.com/sustech-nlp/SAGO.

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

🏢Alibaba GroupHangzhou · China168 works · 15,386 citations🎓Beihang UniversityBeijing · China65 works · 8,452 citations🎓Southern University of Science and TechnologyShenzhen · China19 works · 1,337 citations🎓Shanghai University of Finance and EconomicsShanghai · China13 works · 825 citations

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Zeguan Xiao, Siqing Li, Yong Wang, Xuetao Wei, Jian Yang, Yun Chen and Guanhua Chen. 2026. Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problem. 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. DOI 10.18653/v1/2026.acl-long.890

@inproceedings{xiao2026modeling,
  title = {{Modeling LLM Unlearning as an Asymmetric Two-Task Learning Problem}},
  author = {Zeguan Xiao, Siqing Li, Yong Wang, Xuetao Wei, Jian Yang, Yun Chen and Guanhua Chen},
  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.890},
  url = {https://doi.org/10.18653/v1/2026.acl-long.890}
}

References

30 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yinzhi Cao and Junfeng Yang. 2015. Towards making systems forget with machine unlearning. In 2015 IEEE symposium on security and privacy, pages 463– 480. IEEE. find
  2. Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. 2020. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. Advances in Neural Information Processing Systems, 33:2039–2050. find
  3. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv–2407.
  4. Chongyu Fan, Jiancheng Liu, Licong Lin, Jinghan Jia, Ruiqi Zhang, Song Mei, and Sijia Liu. 2024. Simplicity prevails: Rethinking negative preference optimization for llm unlearning. arXiv preprint arXiv:2410.07163. — [arXiv:2410.07163](https://arxiv.org/abs/2410.07163) find
  5. Antonio Andrea Gargiulo, Donato Crisostomi, Maria Sofia Bucarelli, Simone Scardapane, Fabrizio Silvestri, and Emanuele Rodola. 2025. Task singular vectors: Reducing task interference in model merging. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 18695–18705. find
  6. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5484–5495. find
  7. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR).
  8. Jiabao Ji, Yujian Liu, Yang Zhang, Gaowen Liu, Ramana R Kompella, Sijia Liu, and Shiyu Chang. 2024. Reversing the forget-retain objectives: An efficient llm unlearning framework from logit difference. Advances in Neural Information Processing Systems, 37:12581–12611. find
  9. Zhuoran Jin, Pengfei Cao, Chenhao Wang, Zhitao He, Hongbang Yuan, Jiachun Li, Yubo Chen, Kang Liu, and Jun Zhao. 2024. Rwku: Benchmarking real-world knowledge unlearning for large language models. Preprint, arXiv:2406.10890. — [arXiv:2406.10890](https://arxiv.org/abs/2406.10890) find
  10. Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, and 1 others. 2024. The wmdp benchmark: Measuring and reducing malicious use with unlearning. arXiv preprint arXiv:2403.03218. find
  11. Bo Liu, Qiang Liu, and Peter Stone. 2022. Continual learning and private unlearning. In Conference on Lifelong Learning Agents, pages 243–254. PMLR. find
  12. Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. 2021. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems, 34:18878–18890. find
  13. Xiaoqian Liu, Yangfan Du, Jianjin Wang, Yuan Ge, Chen Xu, Tong Xiao, Guocheng Chen, and Jingbo Zhu. 2025. A modular-based strategy for mitigating gradient conflicts in simultaneous speech translation. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASS… find
  14. 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.
  15. Daniel Marczak, Simone Magistri, Sebastian Cygert, Bartłomiej Twardowski, Andrew D Bagdanov, and Joost van de Weijer. 2025. No task left behind: Isotropic model merging with common and task-specific subspaces. arXiv preprint arXiv:2502.04959. — [arXiv:2502.04959](https://arxiv.org/abs/2502.04959) find
  16. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in neural information processing systems, 35:17359–17372. find
  17. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. Preprint, arXiv:1609.07843.
  18. Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. 2022. Multi-task learning as a bargaining game. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16428–16446.… find
  19. Martin Pawelczyk, Seth Neel, and Himabindu Lakkaraju. 2024. In-context unlearning: language models as few-shot unlearners. In Proceedings of the 41st International Conference on Machine Learning, pages 40034–40050. find
  20. Hadi Reisizadeh, Jinghan Jia, Zhiqi Bu, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, Volkan Cevher, Sijia Liu, and Mingyi Hong. 2025. Blur: A bi-level optimization approach for llm unlearning. arXiv preprint arXiv:2506.08164.
  21. 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:2407.06460. arXiv preprint Filip Sondej, Yushi Yang, MikoĹ Kniejski, Marcel… find
  22. Wenju Sun, Qingyong Li, Yangli-ao Geng, and Boyang Li. 2025. Cat merging: A training-free approach for resolving conflicts in model merging. arXiv preprint arXiv:2505.06977. — [arXiv:2505.06977](https://arxiv.org/abs/2505.06977) find
  23. Vinith M Suriyakumar, Ayush Sekhari, and Ashia Wilson. 2025. Ucd: Unlearning in llms via contrastive decoding. arXiv preprint arXiv:2506.12097. — [arXiv:2506.12097](https://arxiv.org/abs/2506.12097) find
  24. Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Clémentine Fourrier, Nathan Habib, and 1 others. 2023. Zephyr: Direct distillation of lm alignment. arXiv preprint arXiv:2310.16944. find
  25. Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, and Bo Han. 2025. GRU: Mitigating the trade-off between unlearning and retention for LLMs. In Forty-second International Conference on Machine Learning. find
  26. Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Forty-first International Conference on Machine Learning.
  27. Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824– 5836.
  28. Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. 2023. Gpt-4 is too smart to be safe: Stealthy chat with llms via cipher. Preprint, arXiv:2308.06463.
  29. Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. Negative preference optimization: From catastrophic collapse to effective unlearning. arXiv preprint arXiv:2404.05868.
  30. Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. 2023. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/xiao-2026-modeling-llm-unlearning-as-asymmetric/