Elamur Research⌕
☰
home / papers / 2026 / Controllable LLM Reasoning via Sparse Autoencoder-Based Steering

Controllable LLM Reasoning via Sparse Autoencoder-Based Steering

Paper · 2026 · Yi Fang, Wenjie Wang, Mingfeng Xue et al.

Summary

Large Reasoning Models (LRMs) choose their own reasoning moves, such as re-checking or backtracking, and these choices can lead to wasted effort or wrong answers. The authors aim to let users steer which strategy the model applies next. Prompt instructions are often ignored. Contrastive activation vectors tend to mix several strategies together. To get cleaner directions, the authors train Sparse Autoencoders (SAEs) on a model layer's residual stream. Each SAE feature tends to encode one concept, and a chosen feature is added repeatedly to the hidden state while the model generates subsequent tokens. The SAE has tens of thousands of features, so the proposed SAE-Steering pipeline selects useful ones in two steps. First, a cheap filter keeps only features that raise the logits of strategy keywords. Second, the surviving features are ranked by how well they actually steer generation on a small validation set. The selected features control strategies more than 15% better than existing methods. They can also push a model off a faulty path, which improves accuracy. This matters for making LRM reasoning more reliable and interpretable.

Key points

  • Five strategies are targeted: Problem Understanding, Procedural Planning, Backtracking, Multi-Perspective Verification and Hypothesis Reasoning. The authors describe the method as general beyond these five.
  • The SAEs use Top-K sparsity. Steering strength α is lowered step by step from a high value until repetitive output disappears.
  • Stage 1 discards over 99% of features, shrinking roughly 30k candidates to fewer than 100. Stage 2 ranks the remainder by measured control effect.
  • Reported results:
    • over 15% higher control effectiveness than baselines;
    • 7% absolute accuracy gain when redirecting erroneous reasoning;
    • about 5% average improvement in correction rate.
  • Limitations: the pipeline relies on manually chosen strategy keywords and covers only five strategies. Negative steering was explored only briefly.
  • Code and data are at github.com/Peter-Fy/SAE-Steering. The paper appears at ACL 2026.

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

🎓Tsinghua UniversityBeijing · China333 works · 87,019 citations🎓University of Science and Technology of ChinaHefei · China108 works · 112,698 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works

Cite

Cite as Yi Fang, Wenjie Wang, Mingfeng Xue, Boyi Deng, Fengli Xu, Dayiheng Liu and Fuli Feng. 2026. Controllable LLM Reasoning via Sparse Autoencoder-Based Steering. 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.974

@inproceedings{fang2026controllable,
  title = {{Controllable LLM Reasoning via Sparse Autoencoder-Based Steering}},
  author = {Yi Fang, Wenjie Wang, Mingfeng Xue, Boyi Deng, Fengli Xu, Dayiheng Liu and Fuli Feng},
  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.974},
  url = {https://doi.org/10.18653/v1/2026.acl-long.974}
}

References

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

  1. AIME. 2025. Aime problems and solutions. find
  2. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2023. To… find
  3. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. Do NOT think that much for 2+3=? on the overthinking of long reasoning models. In ICML. OpenReview.net. find
  4. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arX…
  5. Boyi Deng, Yu Wan, Baosong Yang, Yidan Zhang, and Fuli Feng. 2025. Unveiling language-specific features in large language models via sparse autoencoders. In ACL (1), pages 4563–4608. Association for Computational Linguistics.
  6. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger B. Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. Toy models of superposition. arXiv pr… find
  7. Zihao Fu, Wai Lam, Anthony Man-Cho So, and Bei Shi. 2021. A theoretical analysis of the repetition problem in text generation. In AAAI. find
  8. Jia LI, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, and Stanislas Polu. 2024. Numinamath. find
  9. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s verify step by step. arXiv preprint arXiv:2305.20050.
  10. Zhengkai Lin, Zhihang Fu, Ze Chen, Chao Chen, Liang Xie, Wenxiao Wang, Deng Cai, Zheng Wang, and Jieping Ye. 2025. Controlling thinking speed in reasoning models. NeurIPS. find
  11. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. find
  12. Andrey V. Galichin, Alexey Dontsov, Polina Druzhinina, Anton Razzhigaev, Oleg Y. Rogov, Elena Tutubalina, and Ivan V. Oseledets. 2025. I have covered all the bases here: Interpreting reasoning features in large language models via sparse autoencoders. arXiv preprint arXiv:2503.18878. — [arXiv:2503.18878](https://arxiv.org/abs/2503.18878) find
  13. Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han Lù, Nicholas Meade, Dongchan Shin, Amirhossein Kazemnejad, Gaurav Kamath, Marius Mosbach, Karolina Stańczak, and Siva Reddy. 2025. Deeps… find
  14. Kanishk Gandhi, Ayush Chakravarthy, Anikait Singh, Nathan Lile, and Noah D. Goodman. 2025. Cognitive behaviors that enable self-improving reasoners, or, four habits of highly effective stars. Second Conference on Language Modeling.
  15. Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel J. Candès, and Tatsunori Hashimoto. 2025. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393.
  16. Sandeep Reddy Gantla. 2025. Exploring mechanistic interpretability in large language models: Challenges, approaches, and insights. In 2025 International Conference on Data Science, Agents & Artificial Intelligence (ICDSAAI). nostalgebraist. 2020. Interpreting gpt: The logit lens. Less-Wrong (blog). find
  17. Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2025. Scaling and evaluating sparse autoencoders. In ICLR. find
  18. OpenAI. 2025. Openai o1 system card. Accessed: 2025-02-21.
  19. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  20. OpenAI. 2024. Gpt-4o system card. arXiv:2410.21276.
  21. Preprint, Qianjun Pan, Wenkai Ji, Yuyang Ding, Junsong Li, Shil-ian Chen, Junyi Wang, Jie Zhou, Qin Chen, Min Zhang, Yulan Wu, and Liang He. 2025. A survey of slow thinking-based reasoning llms using reinforced learning and inference-time scaling law. arXiv preprint arXiv:2505.02665. find
  22. Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoencoders find highly interpretable features in language models. In ICLR. find
  23. Yunjia Qi, Hao Peng, Xiaozhi Wang, Amy Xin, Youfeng Liu, Bin Xu, Lei Hou, and Juanzi Li. 2025. AGEN-TIF: benchmarking instruction following of large language models in agentic scenarios. CoRR, abs/2505.16944. — [arXiv:2505.16944](https://arxiv.org/abs/2505.16944) find
  24. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In EMNLP (1). Association for Computational Linguistics.
  25. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. 2023. GPQA: A graduate-level google-proof q&a benchmark. CoRR, abs/2311.12022.
  26. Leheng Sheng, An Zhang, Zijian Wu, Weixiang Zhao, Changshuo Shen, Yi Zhang, Xiang Wang, and Tat-Seng Chua. 2025. On reasoning strength planning in large reasoning models. NeurIPS. find
  27. Wei Shi, Sihang Li, Tao Liang, Mingyang Wan, Gojun Ma, Xiang Wang, and Xiangnan He. 2025. Route sparse autoencoder to interpret large language models. CoRR, abs/2503.08200.
  28. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learning. In NeurIPS.
  29. Chandan Singh, Jeevana Priya Inala, Michel Galley, Rich Caruana, and Jianfeng Gao. 2024. Rethinking interpretability in the era of large language models. arXiv preprint arXiv:2402.01761.
  30. Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In ICLR. OpenReview.net. find
  31. Xinyu Tang, Xiaolei Wang, Zhihao Lv, Yingqian Min, Xin Zhao, Binbin Hu, Ziqi Liu, and Zhiqiang Zhang. 2025. Unlocking general long chain-of-thought reasoning capabilities of large language models via representation engineering. In ACL (1), pages 6832– 6849. Association for Computational Linguistics.
  32. Openthoughts Team. 2025a. Openthoughts: Data recipes for reasoning models.
  33. Qwen Team. 2025b. Qwen3 technical report. find
  34. Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.
  35. Zeyu Yun, Yubei Chen, Bruno A. Olshausen, and Yann LeCun. 2021. Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors. In DeeLIO@NAACL-HLT, pages 1–10. Association for Computational Linguistics. find
  36. Wojciech Zaremba, Evgenia Nitishinskaya, Boaz Barak, Stephanie Lin, Sam Toyer, Yaodong Yu, Rachel Dias, Eric Wallace, Kai Xiao, Johannes Heidecke, and 1 others. 2025. Trading inference-time compute for adversarial robustness. arXiv preprint arXiv:2501.18841. find
  37. Xingsheng Zhang, Luxi Xing, Chen Zhang, Yanbing Liu, Yifan Deng, Yunpeng Li, Yue Hu, and Chenxu Niu. 2025. Can we steer reasoning direction by thinking intervention? In Findings of EMNLP 2025. find
  38. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric P. Xing, Joseph E. Gonzalez, Ion Stoica, and Hao Zhang. 2024. Lmsys-chat-1m: A large-scale real-world LLM conversation dataset. In ICLR. OpenReview.net.
  39. Qihuang Zhong, Kang Wang, Ziyang Xu, Juhua Liu, Liang Ding, and Bo Du. 2024. Achieving\> 97% on gsm8k: Deeply understanding the problems makes llms better solvers for math word problems. Frontiers of Computer Science. find
  40. Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V. Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, and Huaixiu Steven Zheng. 2024. SELF-DISCOVER: large language models self-compose reasoning structures. In NeurIPS. find
  41. Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, and Neel Nanda. 2025. Understanding reasoning in thinking language models via steering vectors. CoRR, abs/2506.18167. — [arXiv:2506.18167](https://arxiv.org/abs/2506.18167) find
  42. Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, Rui Wang, Zhaopeng Tu, Haitao Mi, and Dong Yu. 2025. Thoughts are all over the place: On the underthinking of o1-like llms. arXiv preprint arXiv:2501.18585. find
  43. Tong Wu, Chong Xiang, Jiachen T. Wang, and Prateek Mittal. 2025. Effectively controlling reasoning models through thinking intervention. arXiv preprint arXiv:2503.24370. — [arXiv:2503.24370](https://arxiv.org/abs/2503.24370) find
  44. Chenxu Yang, Qingyi Si, Mz Dai, Dingyu Yao, Mingyu Zheng, Minghui Chen, Zheng Lin, and Weiping Wang. 2025a. Test-time prompt intervention. find
  45. Jingyuan Yang, Rongjun Li, Weixuan Wang, Ziyu Zhou, Zhiyong Feng, and Wei Peng. 2025b. Lf-steering: Latent feature activation steering for enhancing semantic consistency in large language models. arXiv preprint arXiv:2501.11036. — [arXiv:2501.11036](https://arxiv.org/abs/2501.11036) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/fang-2026-controllable-llm-reasoning-sparse-autoencoder/