Llamas (mostly) think in English: On Causal Interventions in the Latent Language of Transformers
Keywords: mechanistic interpretability, large language models, transformers, residual stream
TL;DR: Llama models tend to think in English, you can project out concepts in English to damage non-English predictions.
Abstract: Previous research on the Llama-2 family of Large Language Models (LLMs) suggested a
correlation indicating the use of English as a intermediary language within
these models for tasks in non-English languages. We improve on this by demonstrating a causal relationship. By
intervening on the intermediate layers during a forward pass, we show
that projecting out the activations onto a subspace corresponding to the correct
prediction in English impairs the model's ability to make
correct predictions on non-English translation tasks. Projecting onto an unrelated
English subspace, or a related subspace in a non-English language, has little effect,
demonstrating that
this family of models store concepts that have a high similarity to the corresponding
concept in English in the residual stream.
Supplementary Material: zip
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12992
Loading