Elamur Research⌕
☰
home / papers / 2025 / Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders

Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders

Paper · 2025 · Boyi Deng, Yu Wan, Baosong Yang et al.

Summary

The paper asks how large language models represent individual languages internally. Earlier work relied on single neurons, which suffer from superposition, or on logit-lens readouts, which are unreliable outside the final layers. The authors instead use pretrained Sparse Autoencoders (SAEs), which split residual-stream activations into sparse, more interpretable features. Their metric for a feature's monolinguality is its mean activation on one language minus its average activation on the other languages. They compute it on Flores-10, a 10-language parallel subset of Flores-200, for Gemma 2 2B, Gemma 2 9B and Llama-3.1-8B.

They find strongly language-specific features, often a single dominant one per language. These features respond to tokens and also to the surrounding linguistic context, as code-switching tests show. Zeroing them out through directional ablation harms mainly the matching language, and some languages have several features that work together. The features also serve as gates for steering vectors, giving better control over the output language. The work offers an interpretable way to locate and manipulate language-specific behaviour in LLMs.

Key points

  • Data and models:
    • Flores-10 covers en, es, fr, ja, ko, pt, th, vi, zh and ar; monolinguality scores use the first 100 sentences.
    • SAEs come from Gemma Scope (Gemma 2 2B/9B) and Llama Scope (Llama-3.1-8B), applied to the residual stream.
    • Ukrainian and other low-resource languages are not covered.
  • Example feature: feature 13,788 in layer 10 of Gemma 2 2B averages about 16.5 activation on Chinese, with only small values for the other languages.
  • Effect of context: a Spanish prefix raises Spanish-feature activation on French or Korean nouns, more strongly in deeper layers. A prefix in a different language lowers the original language's feature activation.
  • Steering results:
    • Feature-gated steering applied over 3 layers (SAE 3L) was compared with a standard steering vector on 1 layer (SV 1L).
    • SAE 3L generally beat SV 1L on both success rate and CE loss in Cross-Lingual Continuation and Adversarial Language Identification, but not in every case.
  • Code and limitations: code is at github.com/Aatrox103/multilingual-llm-features. The SAEs were not trained on curated multilingual data, which the authors list as a limitation.

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 Science and Technology of ChinaHefei · China108 works · 112,698 citations🎓Sichuan UniversityChengdu · China14 works · 1,293 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 Boyi Deng, Yu Wan, Baosong Yang, Yidan Zhang and Fuli Feng. 2025. Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders. Proceedings of the 63rd 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/v1/2025.acl-long.229

@inproceedings{deng2025unveiling,
  title = {{Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders}},
  author = {Boyi Deng, Yu Wan, Baosong Yang, Yidan Zhang and Fuli Feng},
  booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2025},
  doi = {10.18653/v1/2025.acl-long.229},
  url = {https://doi.org/10.18653/v1/2025.acl-long.229}
}

References

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

  1. Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. In Advances in Neural Information Processing Systems, volume 37, pages 136037– 136083. Curran Associates, Inc.
  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, Brayden McLean, Josiah… find
  3. Transformer Circuits Thread. Https://transformer-circuits.pub/2023/monosemantic-features/index.html. find
  4. Laurie Burchell, Alexandra Birch, Nikolay Bogoychev, and Kenneth Heafield. 2023. An open dataset and model for language identification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pag…
  5. 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
  6. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672. find
  7. Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600. find
  8. 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. 2022. Toy models of superposition. Preprint, a… find
  9. Javier Ferrando, Oscar Obeso, Senthooran Rajamanoharan, and Neel Nanda. 2024. Do i know this entity? knowledge awareness and hallucinations in language models. Preprint, arXiv:2411.14257. — [arXiv:2411.14257](https://arxiv.org/abs/2411.14257) find
  10. Javier Ferrando, Oscar Obeso, Senthooran Rajamanoharan, and Neel Nanda. 2025. Do i know this entity? knowledge awareness and hallucinations in language models. Preprint, arXiv:2411.14257. — [arXiv:2411.14257](https://arxiv.org/abs/2411.14257) find
  11. 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. find
  12. Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Krishnan, Marc’Aurelio Ranzato, Francisco Guzmán, and Angela Fan. 2022. The flores-101 evaluation benchmark for low-resource and multilingual machine translation. Transactions of the Association for Computat… 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, Arun Rao,… — The Llama 3 Herd of Models
  14. Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, et al. 2024. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526. find
  15. Geoffrey E. Hinton and Richard S. Zemel. 1993. Autoencoders, minimum description length and helmholtz free energy. In Advances in Neural Information Processing Systems 6, [7th NIPS Conference, Denver, Colorado, USA, 1993], pages 3–10. Morgan Kaufmann. find
  16. Jie Huang and Kevin Chen-Chuan Chang. 2023. Towards reasoning in large language models: A survey. In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023, pages 1049–1065. Association for Computational Linguistics.
  17. Kaiyu Huang, Fengran Mo, Xinyu Zhang, Hongliang Li, You Li, Yuanchi Zhang, Weijian Yi, Yulong Mao, Jinchen Liu, Yuzhuang Xu, Jinan Xu, Jian-Yun Nie, and Yang Liu. 2025. A survey on large language models with multilingualism: Recent advances and new frontiers. Preprint, arXiv:2405.10936. — [arXiv:2405.10936](https://arxiv.org/abs/2405.10936) find
  18. Takeshi Kojima, Itsuki Okimura, Yusuke Iwasawa, Hitomi Yanaka, and Yutaka Matsuo. 2024. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons. In Proceedings of the 2024 Conference of the North American Chapter of the Association…
  19. Garry Kuwanto, Chaitanya Agarwal, Genta Indra Winata, and Derry Tanti Wijaya. 2024. Linguistics theory meets llm: Code-switched text generation via equivalence constrained large language models. arXiv preprint arXiv:2410.22660. find
  20. 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
  21. Renze Lou, Kai Zhang, and Wenpeng Yin. 2024. Large language model instruction following: A survey of progresses and challenges. Preprint, arXiv:2303.10475.
  22. Harry Mayne, Yushi Yang, and Adam Mahdi. 2024. Can sparse autoencoders be used to decompose and interpret steering vectors? Preprint, arXiv:2411.08790. nostalgebraist. 2020. Interpreting gpt: the logit lens. LessWrong. — [arXiv:2411.08790](https://arxiv.org/abs/2411.08790) find
  23. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Ja… — GPT-4 Technical Report
  24. Gonçalo Paulo, Alex Mallen, Caden Juang, and Nora Belrose. 2024. Automatically interpreting millions of features in large language models. Preprint, arXiv:2410.13928. — [arXiv:2410.13928](https://arxiv.org/abs/2410.13928) find
  25. Libo Qin, Qiguang Chen, Yuhang Zhou, Zhi Chen, Yinghui Li, Lizi Liao, Min Li, Wanxiang Che, and Philip S. Yu. 2024. Multilingual large language model: A survey of resources, taxonomy and frontiers. Preprint, arXiv:2404.04925. find
  26. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng… find
  27. 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
  28. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand,… find
  29. Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu…
  30. Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, Soroosh Mariooryad, Yifan Ding, Xinyang Geng, Fred Alcober, Roy Frostig, Mark Omernick, Lexi Walker, Cosmin Paduraru, Christina Sorokin, Andrea Tacchetti, Colin… — Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
  31. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. 2024b. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. find
  32. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller,… — Llama 2: Open Foundation and Fine-Tuned Chat Models
  33. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering language models with activation engineering. Preprint, arXiv:2308.10248.
  34. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Juny… find
  35. Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. 2024a. Instruction tuning for large language models: A survey. Preprint, arXiv:2308.10792.
  36. Zhihao Zhang, Jun Zhao, Qi Zhang, Tao Gui, and Xuanjing Huang. 2024b. Unveiling linguistic regions in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 6228…
  37. Yiran Zhao, Wenxuan Zhang, Guizhen Chen, Kenji Kawaguchi, and Lidong Bing. 2024. How do large language models handle multilingualism? In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, Decembe…
  38. Chengzhi Zhong, Fei Cheng, Qianying Liu, Junfeng Jiang, Zhen Wan, Chenhui Chu, Yugo Murawaki, and Sadao Kurohashi. 2024. Beyond english-centric llms: What language do multilingual language models think in? Preprint, arXiv:2408.10811.
  39. Chris Wendler, Veniamin Veselovsky, Giovanni Monea, and Robert West. 2024. Do llamas work in english? on the latent language of multilingual transformers. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand,…
  40. Genta Winata, Alham Fikri Aji, Zheng Xin Yong, and Thamar Solorio. 2023. The decades progress on code-switching research in NLP: A systematic survey on trends and challenges. In Findings of the Association for Computational Linguistics: ACL 2023, pages 2936–2978, Toronto, Canada. Association for Com… find
  41. Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, and Tat-Seng Chua. 2025. Personalized generation in large model era: A survey. arXiv preprint arXiv:2503.02614.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/deng-2025-unveiling-language-specific-features-large/