Elamur Research
☰
home / papers / 2024 / Do Llamas Work in English? On the Latent Language of Multilingual Transformers

Do Llamas Work in English? On the Latent Language of Multilingual Transformers

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/2024.acl-long.820.pdf

Summary

The paper asks whether multilingual LLMs trained mostly on English text silently route non-English processing through English. Using the open Llama-2 models (7B, 13B, 70B), the authors build non-English prompts for three tasks: translation, cloze and word repetition. Each prompt has exactly one correct single-token answer. They decode the last token's hidden state at every layer with the logit lens. They avoid the tuned lens on purpose, because its training would erase the signal they want to detect. They also analyse the latents geometrically as points in embedding space.

The forward pass shows three stages:

  1. Early layers decode nothing meaningful.
  2. Middle layers already reflect the correct meaning but prefer the English word over the target-language one.
  3. Late layers move into the region of the input language.

The authors read this as a path from an input space, through a "concept space" that leans toward English, to an output space. They argue this is a semantic bias, not an explicit translation step. This matters because such a bias could shape model behaviour in other languages and disadvantage less-resourced ones.

Key points

  • Llama-2 was trained on about two trillion tokens, of which 89.70% are English. Small shares still amount to billions of tokens, for example roughly 3.4B German and 2.6B Chinese tokens.
  • The three model sizes have 32/40/80 layers and a shared 32,000-token vocabulary. All were run with 8-bit quantization.
  • Logit-lens decoding shows the English version of the answer peaking in middle layers before the correct target-language token takes over. One example is French "fleur" → Chinese "花", where "flower" appears first.
  • The English detour is weaker on the repetition task and where words map to single tokens. The shares of single-token words were 100% for Chinese, 55% for French, 43% for German and 13% for Russian, which suggests tokenization plays a role.
  • The study is limited to Llama-2, with initial evidence that Mistral-7B behaves the same way. Code and data are released at github.com/epfl-dlab/llm-latent-language.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

120 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 119 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.acl-long.820

@inproceedings{wendler2024do,
  title = {{Do Llamas Work in English? On the Latent Language of Multilingual Transformers}},
  author = {Wendler, Chris and Veselovsky, Veniamin and Monea, Giovanni and West, Robert},
  booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2024},
  doi = {10.18653/v1/2024.acl-long.820},
  url = {https://doi.org/10.18653/v1/2024.acl-long.820}
}

References

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

  1. Kabir Ahuja, Harshita Diddee, Rishav Hada, Millicent Ochieng, Krithika Ramesh, Prachi Jain, Akshay Nambi, Tanuja Ganu, Sameer Segal, Maxamed Axmed, Kalika Bali, and Sunayana Sitaram. 2023. Mega: Multilingual evaluation of generative ai. find our card
  2. Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics. find our card
  3. Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023. find our card
  4. Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, 4 https://github.com/nrimsky/LM-exp/blob/main/ intermediate\_decoding/intermediate\_decoding. ipynb and Jacob Steinhardt. 2023. Eliciting latent predictions from transformers with the tuned lens. arXiv… ↗ our card
  5. Mor Geva, Avi Caciularu, Kevin Ro Wang, and Yoav Goldberg. 2022. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. find our card
  6. Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. 2023. Pythia: A suite for analyzing large language models across training and scaling. In International Conferenc… find
  7. Charles Goddard. 2023. Llama-polyglot-13b. https:// huggingface.co/chargoddard/llama-polyglot-13b. Accessed: 2024-01-22. find
  8. Lera Boroditsky, Lauren A. Schmidt, and Webb Phillips. 2003. Sex, syntax, and semantics. In Dedre Gentner and Susan Goldin-Meadow, editors, Language in Mind: Advances in the Study of Language and Thought, pages 61–79. MIT Press, Cambridge, MA. find
  9. Nicola Cancedda. 2024. Spectral filters, dark signals, and attention sinks. arXiv preprint arXiv:2402.09221. find
  10. Steven Cao, Nikita Kitaev, and Dan Klein. 2020. Multilingual alignment of contextual word representations. find
  11. Yihong Chen, Kelly Marchisio, Roberta Raileanu, David Ifeoluwa Adelani, Pontus Stenetorp, Sebastian Riedel, and Mikel Artetxe. 2023. Improving language plasticity via pretraining with active forgetting. find
  12. Rochelle Choenni and Ekaterina Shutova. 2020. What does it mean to be language-agnostic? probing multilingual sentence encoders for typological properties. arXiv preprint arXiv:2009.12862. — [arXiv:2009.12862](https://arxiv.org/abs/2009.12862) find
  13. Arthur Conmy, Augustine N Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. 2023. Towards automated circuit discovery for mechanistic interpretability. arXiv preprint arXiv:2304.14997. find
  14. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. find
  15. Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6022–6034. find our card
  16. Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. LLM.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339. find
  17. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. find
  18. Anna Di Natale, Max Pellert, and David Garcia. 2021. Colexification networks encode affective meaning. Affective Science, 2(2):99–111. find
  19. Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. 2021. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1. find
  20. Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lopez de Lacalle, and Mikel Artetxe. 2023. Do multilingual language models think better in english? find
  21. Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. 2023. Finding neurons in a haystack: Case studies with sparse probing. arXiv preprint arXiv:2305.01610. find
  22. Bofeng Huang. 2023. vigogne-2-13b-instruct. https://huggingface.co/bofenghuang/ vigogne-2-13b-instruct. Accessed: 2024-01-22. ↗
  23. Haoyang Huang, Tianyi Tang, Dongdong Zhang, Wayne Xin Zhao, Ting Song, Yan Xia, and Furu Wei. 2023. Not all languages are created equal in llms: Improving multilingual capability by cross-lingual-thought prompting. find our card
  24. Jaavid Aktar Husain, Raj Dabre, Aswanth Kumar, Ratish Puduppully, and Anoop Kunchukuttan. 2024. Romansetu: Efficiently unlocking multilingual capabilities of large language models models via romanization. find
  25. Daekeun Kim. 2023. Llama-2-ko-dpo-13b. https:// huggingface.co/daekeun-ml/Llama-2-ko-DPO-13B. Accessed: 2024-01-22. find
  26. Kenneth Li, Aspen K Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2022. Emergent world representations: Exploring a sequence model trained on a synthetic task. arXiv preprint arXiv:2210.13382. find
  27. Jindřich Libovický, Rudolf Rosa, and Alexander Fraser. 2020. On the language neutrality of pre-trained multilingual representations. arXiv preprint arXiv:2004.05160. find
  28. Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O’Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyano… find our card
  29. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726–742. find
  30. 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. find our card
  31. 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
  32. Giovanni Monea, Maxime Peyrard, Martin Josifoski, Vishrav Chaudhary, Jason Eisner, Emre Kıcıman, Hamid Palangi, Barun Patra, and Robert West. 2023. A glitch in the matrix? locating and detecting language model grounding with fakepedia. arXiv preprint arXiv:2312.02073. find
  33. Benjamin Muller, Yanai Elazar, Benoît Sagot, and Djamé Seddah. 2021. First align, then predict: Understanding the cross-lingual ability of multilingual bert. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2214–2231. find
  34. Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217. find our card
  35. Nostalgebraist. 2020. Interpreting gpt: The logit lens. Less-Wrong. find
  36. Rafael E. Núñez and Eve Sweetser. 2006. With the future behind them: Convergent evidence from aymara language and gesture in the crosslinguistic comparison of spatial construals of time. Cognitive Science, 30(3):401–450. find
  37. OpenAI. 2023. Gpt-4 technical report. find our card
  38. Isabel Papadimitriou, Kezia Lopez, and Dan Jurafsky. 2022. Multilingual bert has an accent: Evaluating english influences on fluency in multilingual models. find
  39. Rait Piir. 2023. Finland’s chatgpt equivalent begins to think in estonian as well. ERR News. find
  40. Björn Plüster. 2023. LeoLM: Ein Impuls für Deutschsprachige LLM-Forschung. https: Accessed: 2024-//laion.ai/blog-de/leo-lm/. 01-22. find
  41. Lucia Quirke, Lovis Heindrich, Wes Gurnee, and Neel Nanda. 2023. Training dynamics of contextual n-grams in language models. arXiv preprint arXiv:2311.00863. find
  42. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  43. Nina Rimsky. 2023. Decoding intermediate activations in Llama-2-7b. LessWrong. find
  44. Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100. find
  45. Tal Schuster, Ori Ram, Regina Barzilay, and Amir Globerson. 2019. Cross-lingual alignment of contextual word embeddings, with applications to zero-shot dependency parsing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lang… find
  46. Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. 2022. Language models are multilingual chain-of-thought reasoners. find our card
  47. Oleh Shliazhko, Alena Fenogenova, Maria Tikhonova, Vladislav Mikhailov, Anastasia Kozlova, and Tatiana Shavrina. 2022. mgpt: Few-shot learners go multilingual. find our card
  48. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. find
  49. 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. find
  50. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824–24837. find our card
  51. Xiangpeng Wei, Haoran Wei, Huan Lin, Tianhao Li, Pei Zhang, Xingzhang Ren, Mei Li, Yu Wan, Zhiwei Cao, Binbin Xie, Tianxiang Hu, Shangjie Li, Binyuan Hui, Bowen Yu, Dayiheng Liu, Baosong Yang, Fei Huang, and Jun Xie. 2023. Polylm: An open source polyglot large language model. find
  52. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Ling… find
  53. Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transferable are features in deep neural networks? Advances in neural information processing systems, 27. find
  54. Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. 2015. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579. — [arXiv:1506.06579](https://arxiv.org/abs/1506.06579) find
  55. Xiang Zhang, Senyu Li, Bradley Hauer, Ning Shi, and Grzegorz Kondrak. 2023. Don’t trust ChatGPT when your question is not in English: A study of multilingual abilities and types of LLMs. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7915–7927, Singa… find
  56. Wenhao Zhu, Shujian Huang, Fei Yuan, Shuaijie She, Jiajun Chen, and Alexandra Birch. 2024. Question translation training for better multilingual reasoning. find
  57. Wenhao Zhu, Yunzhe Lv, Qingxiu Dong, Fei Yuan, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2023. Extrapolating large language models to non-english by aligning languages. find