Elamur Research⌕
☰
home / papers / 2026 / Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering

Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering

Paper · 2026 · Haiyan Zhao, Xuansheng Wu, Fan Yang et al.

Summary

Steering large language models by adding linear concept vectors to their hidden states works poorly for abstract concepts such as sycophancy or corrigibility. The authors argue that this is because contrastive training data carries irrelevant features that contaminate the learned directions. They propose SDCV (Sparse Autoencoder-Denoised Concept Vectors) to address this. A pretrained SAE encodes hidden states from positive and negative examples. Each SAE latent is scored by how well it separates the two classes, using a mean difference normalized by class variance. The top-k latents are multiplied by a scale factor, and the representation is then decoded, with the original reconstruction error added back. The denoised states are used to build concept vectors via linear probing or difference-in-mean, and these vectors are injected into the residual stream at inference time. Tests on Llama-3.1-8B and Gemma-2-2B across six concepts show steering success gains of 4–16% with comparable topic relevance. Counterfactual noise-injection experiments support the idea that filtering noisy latents drives these gains. The method is a cheap, plug-in improvement to existing steering pipelines.

Key points

  • Concepts tested: three multiple-choice behavioural datasets following Rimsky et al. (sycophancy, corrigibility, myopic reward) and three public corpora (Amazon reviews, political speech, Twitter sentiments). Each concept was evaluated with 50 prompts.
  • Model setup: steering was applied at layer 25 of Llama-3.1-8B (SAE with 131,072 latents) and layer 20 of Gemma-2-2B (GemmaScope JumpReLU SAE with 65,536 latents).
  • Hyperparameters searched: k ranged from 50 to 50,000 for Llama and 50–100 for Gemma; scale factors ranged from 10 to 100. Optimal values were found empirically and showed no simple trend.
  • Selected Llama gains (success rate):
    • Difference-in-mean on political speech improved from 20% to 36%.
    • Linear probing on corrigibility improved from 6% to 20%.
    • SDCV with difference-in-mean was usually the strongest variant.
    • SDCV with linear probing did better on the noisy Twitter data.
  • Limitations: the method requires an existing trained SAE for the model family, and the hyperparameter tuning is computationally costly.

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

🎓University of GeorgiaAthens · United States17 works · 26,331 citations🎓New Jersey Institute of TechnologyNewark · United States15 works · 1,333 citations🎓Wake Forest UniversityWinston-Salem · United States5 works · 928 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 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Haiyan Zhao, Xuansheng Wu, Fan Yang, Bo Shen, Ninghao Liu and Mengnan Du. 2026. Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.40

@inproceedings{zhao2026denoising,
  title = {{Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering}},
  author = {Haiyan Zhao, Xuansheng Wu, Fan Yang, Bo Shen, Ninghao Liu and Mengnan Du},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.40},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.40}
}

References

27 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Guillaume Alain and Yoshua Bengio. 2016. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644. find
  2. Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. Scaling and evaluating sparse autoencoders. arXiv preprint. ArXiv:2406.04093 [cs]. find
  3. Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, and 1 others. 2024. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526. find
  4. Zirui He, Haiyan Zhao, Yiran Qiao, Fan Yang, Ali Payani, Jing Ma, and Mengnan Du. 2025. SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Models. arXiv preprint. ArXiv:2502.11356 [cs]. — [arXiv:2502.11356](https://arxiv.org/abs/2502.11356) find
  5. Ole Jorgensen, Dylan Cope, Nandi Schoots, and Murray Shanahan. 2023. Improving activation steering in language models with mean-centring. arXiv preprint arXiv:2312.03813. — [arXiv:2312.03813](https://arxiv.org/abs/2312.03813) find
  6. Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and 1 others. 2018. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668–2677. PMLR. find
  7. 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
  8. Kai Konen, Sophie Jentzsch, Diaoulé Diallo, Peer Schütt, Oliver Bensch, Roxanne El Baff, Dominik Opitz, and Tobias Hecking. 2024. Style vectors for steering generative large language model. arXiv preprint arXiv:2402.01618. — [arXiv:2402.01618](https://arxiv.org/abs/2402.01618) find
  9. David Chanin, James Wilken-Smith, Tomáš Dulka, Hardik Bhatnagar, and Joseph Bloom. 2024. A is for absorption: Studying feature splitting and absorption in sparse autoencoders. arXiv preprint arXiv:2409.14507. — [arXiv:2409.14507](https://arxiv.org/abs/2409.14507) find
  10. Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. arXiv preprint arXiv:2408.05147. find
  11. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022a. Toy models of superposition. Transforme… find
  12. Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, and 1 others. 2022b. Toy models of superposition. arXiv preprint arXiv:2209.10652. find
  13. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36.
  14. Johnny Lin. 2023. Neuronpedia: Interactive reference and tooling for analyzing neural networks. Software available from neuronpedia.org. find
  15. Sheng Liu, Lei Xing, and James Zou. 2023. In-context vectors: Making in context learning more effective and controllable through latent space steering. arXiv preprint arXiv:2311.06668. find
  16. Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824.
  17. Nedjma Ousidhoum, Xinran Zhao, Tianqing Fang, Yangqiu Song, and Dit-Yan Yeung. 2021. Probing toxic content in large pre-trained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language… find
  18. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2023. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681. find
  19. Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lindsey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, and 1 others. 2025. Open problems in mechanistic interpretability. arXiv preprint arXiv:2501.16496. — [arXiv:2501.16496](https://arxiv.org/abs/2501.16496) find
  20. Dong Shu, Xuansheng Wu, Haiyan Zhao, Mengnan Du, and Ninghao Liu. 2025a. Beyond input activations: Identifying influential latents by gradient sparse autoencoders. arXiv preprint arXiv:2505.08080. find
  21. Dong Shu, Xuansheng Wu, Haiyan Zhao, Daking Rai, Ziyu Yao, Ninghao Liu, and Mengnan Du. 2025b. A survey on sparse autoencoders: Interpreting the internal mechanisms of large language models. arXiv preprint arXiv:2503.05613. find
  22. Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, and Besmira Nushi. 2024. Improving Instruction-Following in Language Models through Activation Steering. arXiv preprint. ArXiv:2410.12877 [cs]. — [arXiv:2410.12877](https://arxiv.org/abs/2410.12877) find
  23. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.
  24. Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D Manning, and Christopher Potts. 2025. Axbench: Steering llms? even simple baselines outperform sparse autoencoders. arXiv preprint arXiv:2501.17148. find
  25. Jason Zhang and Scott Viteri. 2025. Uncovering Latent Chain of Thought Vectors in Language Models. arXiv preprint. ArXiv:2409.14026 [cs]. find
  26. Haiyan Zhao, Heng Zhao, Bo Shen, Ali Payani, Fan Yang, and Mengnan Du. 2025. Beyond single concept vector: Modeling concept subspace in LLMs with gaussian distribution. In The Thirteenth International Conference on Learning Representations. find
  27. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhao-2026-denoising-concept-vectors-sparse-autoencoders/