Elamur Research⌕
☰
home / papers / 2025 / Model Unlearning via Sparse Autoencoder Subspace Guided Projections

Model Unlearning via Sparse Autoencoder Subspace Guided Projections

Paper · 2025 · Xu Wang, Zihao Li, Benyou Wang et al.

Summary

The paper addresses selective knowledge removal from LLMs, where hazardous or private information must be erased without hurting general ability. Gradient-based unlearning changes weights but offers little insight into what changed internally. Steering with sparse autoencoders (SAEs) is interpretable, but it only alters activations at inference time, which can reduce utility and leave the model open to jailbreaks. The authors propose SSPU (SAE–Guided Subspace Projection Unlearning), which uses SAE features to guide real weight updates. First, a data-driven step selects an SAE layer and ranks latent features by how strongly they activate on forget data versus retain data. Next, decoder vectors of the most and least topic-related features are orthonormalized with QR decomposition into two subspaces. Finally, fine-tuning pushes forget-set activations toward a target inside the topic-unrelated subspace, keeps retain activations close to a frozen copy, and adds a regularizer on the parameter shift. On WMDP–Cyber, SSPU forgets more than the strongest baseline, keeps more utility, and resists jailbreak prompts better. This shows that interpretable features can make unlearning both controllable and robust.

Key points

  • Evaluation: WMDP–Cyber serves as the forget set, and MMLU, TruthfulQA and GSM8K measure retained utility. Baselines are Gradient Ascent, NPO, RMU and SAE steering via conditional clamping (e.g., clamp level −300).
  • Compared with RMU, the best baseline, SSPU lowers WMDP–Cyber accuracy by a further 3.22% and raises average utility by 2.88%.
  • Robustness was tested on four jailbreak-prompt unlearning tasks built around WMDP–Cyber. SSPU cuts malicious accuracy by 13.59% relative to SAE-based unlearning and by 2.83% relative to RMU.
  • The training objective sums three terms: an unlearning loss, a retention loss, and a λ-weighted regularization term tied to the SAE-derived subspaces. The SAE uses JumpReLU activations.
  • Stated limitations: SSPU needs a well-trained SAE plus separable forget and retain corpora. Subspace size (K and rank) requires tuning, and it gives no formal guarantee that unrelated capabilities stay untouched.

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

🎓Chinese University of Hong KongHong Kong · China126 works · 14,073 citations🎓University of Hong KongPok Fu Lam · China89 works · 11,774 citations🎓Chinese University of Hong Kong, ShenzhenShenzhen · China47 works · 2,596 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

Cite

Cite as Xu Wang, Zihao Li, Benyou Wang, Yan Hu and Difan Zou. 2025. Model Unlearning via Sparse Autoencoder Subspace Guided Projections. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.emnlp-main.1348

@inproceedings{wang2025model,
  title = {{Model Unlearning via Sparse Autoencoder Subspace Guided Projections}},
  author = {Xu Wang, Zihao Li, Benyou Wang, Yan Hu and Difan Zou},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  doi = {10.18653/v1/2025.emnlp-main.1348},
  url = {https://doi.org/10.18653/v1/2025.emnlp-main.1348}
}

References

49 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Fazl Barez, Tingchen Fu, Ameya Prabhu, Stephen Casper, Amartya Sanyal, Adel Bibi, Aidan O’Gara, Robert Kirk, Ben Bucknall, Tim Fist, Luke Ong, Philip Torr, Kwok-Yan Lam, Robert Trager, David Krueger, Sören Mindermann, José Hernandez-Orallo, Mor Geva, and Yarin Gal. 2025. Open problems in machine unl… — [arXiv:2501.04952](https://arxiv.org/abs/2501.04952) find
  2. Karuna Bhaila, Minh-Hao Van, and Xintao Wu. 2024. Soft prompting for unlearning in large language models. Preprint, arXiv:2406.12038.
  3. Chein-I Chang. 2005. Orthogonal subspace projection (osp) revisited: A comprehensive study and analysis. IEEE transactions on geoscience and remote sensing, 43(3):502–518. 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. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. Preprint, arXiv:2110.14168.
  6. Enjun Du, Xunkai Li, Tian Jin, Zhihan Zhang, Rong-Hua Li, and Guoren Wang. 2025a. Graphmas-ter: Automated graph synthesis via llm agents in data-limited environments. ArXiv preprint arXiv:2504.00711v2. — [arXiv:2504.00711](https://arxiv.org/abs/2504.00711) find
  7. Enjun Du, Siyi Liu, and Yongqi Zhang. 2025b. Graphor-acle: A foundation model for knowledge graph reasoning. ArXiv preprint arXiv:2505.11125. — [arXiv:2505.11125](https://arxiv.org/abs/2505.11125) find
  8. Enjun Du, Siyi Liu, and Yongqi Zhang. 2025c. Mixture of length and pruning experts for knowledge graphs reasoning. ArXiv preprint arXiv:2507.20498. find
  9. Eoin Farrell, Yeu-Tong Lau, and Arthur Conmy. 2024. Applying sparse autoencoders to unlearn knowledge in language models. In Neurips Safe Generative AI Workshop 2024. find
  10. Walter Gander. 1980. Algorithms for the qr decomposition. Res. Rep, 80(02):1251–1268. find
  11. Leo Gao, Tom Dupre la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2025. Scaling and evaluating sparse autoencoders. In The Thirteenth International Conference on Learning Representations. find
  12. Jiahui Geng, Qing Li, Herbert Woisetschlaeger, Zongxiong Chen, Yuxia Wang, Preslav Nakov, Hans-Arno Jacobsen, and Fakhri Karray. 2025. A comprehensive survey of machine unlearning techniques for large language models. Preprint, arXiv:2503.01854. — [arXiv:2503.01854](https://arxiv.org/abs/2503.01854) find
  13. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  14. Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024. Parameter-efficient fine-tuning for large models: A comprehensive survey. Transactions on Machine Learning Research.
  15. Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang, and Xipeng Qiu. 2024. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. Preprint, arXiv:2410.20526. find
  16. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations.
  17. Joel Jang, Dongkeun Yoon, Sohee Yang, Sungmin Cha, Moontae Lee, Lajanugen Logeswaran, and Minjoon Seo. 2023. Knowledge unlearning for mitigating privacy risks in language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages…
  18. Jinghan Jia, Yihua Zhang, Yimeng Zhang, Jiancheng Liu, Bharat Runwal, James Diffenderfer, Bhavya Kailkhura, and Sijia Liu. 2024. Soul: Unlocking the power of second-order optimization for llm unlearning. CoRR, abs/2404.18239.
  19. Dahyun Jung, Jaehyung Seo, Jaewook Lee, Chanjun Park, and Heuiseok Lim. 2025. CoME: An unlearning-based approach to conflict-free model editing. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies… find
  20. Matthew Khoriaty, Andrii Shportko, Gustavo Mercier, and Zach Wood-Doughty. 2025. Don’t forget it! conditional sparse autoencoder clamping works for unlearning. Preprint, arXiv:2503.11127. — [arXiv:2503.11127](https://arxiv.org/abs/2503.11127) find
  21. Edward Kim. 2024. Nevermind: Instruction override and moderation in large language models. Preprint, arXiv:2402.03303. — [arXiv:2402.03303](https://arxiv.org/abs/2402.03303) find
  22. Hyoseo Kim, Dongyoon Han, and Junsuk Choe. 2024. Negmerge: Consensual weight negation for strong machine unlearning. In Adaptive Foundation Models: Evolving AI for Personalized and Efficient Learning. find
  23. Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, Xin Zhou, Enzhi Wang, and Xiaohang Dong. 2024. Better zero-shot reasoning with role-play prompting. In NAACL-HLT, pages 4099– 4113. find
  24. Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew Bo Liu, Michael Chen, Isabelle Barrass, Oliver Zhang, Xiaoyuan Zhu, Rishub Tamirisa, an… find
  25. Zihao Li, Xu Wang, Yuzhe Yang, Ziyu Yao, Haoyi Xiong, and Mengnan Du. 2025. Feature extraction and steering for enhanced chain-of-thought reasoning in language models. arXiv preprint arXiv:2505.15634.
  26. Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, Janos Kramar, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Inte… find
  27. Johnny Lin. 2023. Neuronpedia: Interactive reference and tooling for analyzing neural networks. Software available from neuronpedia.org. find
  28. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214–3252, Dublin, Ireland. Association for Computational Linguistics…
  29. Aofan Liu, Lulu Tang, Ting Pan, Yuguo Yin, Bin Wang, and Ao Yang. 2025a. Pico: Jailbreaking multimodal large language models via Pictorial Code contextualization. Preprint, arXiv:2504.01444. — [arXiv:2504.01444](https://arxiv.org/abs/2504.01444) find
  30. Chris Yuhao Liu, Yaxuan Wang, Jeffrey Flanigan, and Yang Liu. 2024. Large language model unlearning via embedding-corrupted prompts. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. find
  31. Zheyuan Liu, Guangyao Dou, Xiangchi Yuan, Chunhui Zhang, Zhaoxuan Tan, and Meng Jiang. 2025b. Modality-aware neuron pruning for unlearning in multimodal large language models. Preprint, arXiv:2502.15910.
  32. Christopher J Lynch, Erik J Jensen, Virginia Zamponi, Kevin O’Brien, Erika Frydenlund, and Ross Gore. 2023. A structured narrative prompt for prompting narratives from large language models: sentiment assessment of chatgpt-generated narratives and real tweets. Future Internet, 15(12):375. find
  33. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. Preprint, arXiv:1609.07843.
  34. Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacche… find
  35. Aashiq Muhamed, Jacopo Bonato, Mona Diab, and Virginia Smith. 2025. Saes Can improve unlearning: Dynamic sparse autoencoder guardrails for precision unlearning in llms. Preprint, arXiv:2504.08192. — [arXiv:2504.08192](https://arxiv.org/abs/2504.08192) find
  36. David Pape, Sina Mavali, Thorsten Eisenhofer, and Lea Schönherr. 2025. Prompt obfuscation for large language models. Preprint, arXiv:2409.11026. — [arXiv:2409.11026](https://arxiv.org/abs/2409.11026) find
  37. Martin Pawelczyk, Seth Neel, and Himabindu Lakkaraju. 2024. In-context unlearning: Language models as few-shot unlearners. In Forty-first International Conference on Machine Learning. find
  38. Nicholas Pochinkov and Nandi Schoots. 2024. Dissecting language models: Machine unlearning via selective pruning. CoRR, abs/2403.01267. — [arXiv:2403.01267](https://arxiv.org/abs/2403.01267) find
  39. Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda. 2024. Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders. Preprint, arXiv:2407.14435. — [arXiv:2407.14435](https://arxiv.org/abs/2407.14435) find
  40. Mark Russinovich and Ahmed Salem. 2025. Obliviate: Efficient unmemorization for protecting intellectual property in large language models. Preprint, arXiv:2502.15010. — [arXiv:2502.15010](https://arxiv.org/abs/2502.15010) find
  41. Xu Shen, Yixin Liu, Yiwei Dai, Yili Wang, Rui Miao, Yue Tan, Shirui Pan, and Xin Wang. 2025. Understanding the information propagation effects of communication topologies in llm-based multi-agent systems. arXiv preprint arXiv:2505.23352.
  42. Nianwen Si, Hao Zhang, Heyu Chang, Wenlin Zhang, Dan Qu, and Weiqiang Zhang. 2023. Knowledge unlearning for llms: Tasks, methods, and challenges. Preprint, arXiv:2311.15766. — [arXiv:2311.15766](https://arxiv.org/abs/2311.15766) find
  43. Xu Wang, Yan Hu, Wenyu Du, Reynold Cheng, Benyou Wang, and Difan Zou. 2025a. Towards understanding fine-tuning mechanisms of LLMs via circuit analysis. In ICLR 2025 Workshop on Building Trust in Language Models and Applications. find
  44. Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, and Bo Han. 2025b. Gru: Mitigating the trade-off between unlearning and retention for large language models. Preprint, arXiv:2503.09117. — [arXiv:2503.09117](https://arxiv.org/abs/2503.09117) find
  45. Haoming Xu, Ningyuan Zhao, Liming Yang, Sendong Zhao, Shumin Deng, Mengru Wang, Bryan Hooi, Nay Oo, Huajun Chen, and Ningyu Zhang. 2025. Relearn: Unlearning via learning for large language models. Preprint, arXiv:2502.11190. find
  46. Yuanshun Yao, Xiaojun Xu, and Yang Liu. 2024. Large language model unlearning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems.
  47. Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. 2024. Negative preference optimization: From catastrophic collapse to effective unlearning. In First Conference on Language Modeling.
  48. Xuandong Zhao, Will Cai, Tianneng Shi, David Huang, Licong Lin, Song Mei, and Dawn Song. 2025. Improving llm safety alignment with dual-objective optimization. Preprint, arXiv:2503.03710. — [arXiv:2503.03710](https://arxiv.org/abs/2503.03710) find
  49. Xiaofan Zheng, Zinan Zeng, Heng Wang, Yuyang Bai, Yuhan Liu, and Minnan Luo. 2025. From predictions to analyses: Rationale-augmented fake news detection with large vision-language models. In Proceedings of the ACM on Web Conference 2025, WWW ’25, page 5364–5375, New York, NY, USA. Association for Co… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2025-model-unlearning-sparse-autoencoder-subspace/