Source-linked AI summary
Do Multilingual LLMs Think In English?
Lisa Schut, Yarin Gal, Sebastian Farquhar
TL;DR
The paper investigates whether multilingual LLMs use a language-specific or English-centered representation space. Using logit-lens analysis, activation steering, and cross-lingual factual interventions, it finds English-centric semantic processing, stronger steering from English vectors, and shared but English-biased fact representations.
Problem
Although LLMs operate across languages, the paper asks how their internal representations support multilingual processing and whether semantic decisions are language-specific or English-centric.
Method
The paper analyzes internal representations with a logit lens, tests activation steering across languages, and examines factual representations using cross-lingual causal interventions and interpolation.
Results
LLMs make semantic decisions close to English across input and output languages; English representations often precede target-language lexical outputs, English steering vectors work better, and facts share representations across languages.
Takeaways & Limitations
Multilingual LLM representations are shared across languages but remain English-centric for semantic generation, with non-lexical words retaining more language-specific processing.
Takeaways & Limitations
The analysis does not actively address language confusion because correcting it could alter the natural behavior the study aims to understand.
Abstract
from arXiv · showhide
Large language models (LLMs) have multilingual capabilities and can solve tasks across various languages. However, we show that current LLMs make key decisions in a representation space closest to English, regardless of their input and output languages. Exploring the internal representations with a logit lens for sentences in French, German, Dutch, and Mandarin, we show that the LLM first emits representations close to English for semantically-loaded words before translating them into the target language. We further show that activation steering in these LLMs is more effective when the steering vectors are computed in English rather than in the language of the inputs and outputs. This suggests that multilingual LLMs perform key reasoning steps in a representation that is heavily shaped by English in a way that is not transparent to system users.
1. Introduction
Because LLMs are predominantly trained on English but deployed across many languages, the paper asks how they operate across languages and whether their concept space is English-centric or language-specific.
- LLMs are predominantly trained on English data yet deployed across languages, including languages rarely seen during training.
- The paper frames multilingual operation as a choice between a language-specific concept space and a language-agnostic one.
- One hypothesis is that LLMs operate in a space centered on English, or on their main pretraining language.
2. LLMs ‘operate’ in a language-agnostic space
The paper presents evidence that multilingual LLMs reason in an English-centric representation space, especially for lexical and semantically loaded words, while non-lexical words can remain in the prompt language.
- LLMs reason in an English-centric way during open-ended multi-token generation, contrasting with prior single-token evidence for a language-specific space.
- For lexical words, English-focused representations often appear before being transformed into the target language, whereas non-lexical words do not route through English.
- In a French prompt, lexical words such as “water,” “lake,” and “sun” are selected in English, unlike grammatical elements such as “du” and “le”.
- The English-centric pattern varies across models: Aya is least English-centric, while Gemma is most English-centric.
- Non-English sentences are steered more effectively with English-derived vectors than target-language vectors, although the vectors also encode language-specific components.
- Fact representations are shared between languages, allowing interpolation that preserves the correct answer while changing only the output language.
2. Background
The paper combines mechanistic interpretability methods with multilingual model and dataset descriptions to examine internal representations, factual encoding, and activation steering.
- The study covers four open-source models that differ in architecture and language coverage, and evaluates them across English, French, German, Dutch, and Chinese.
- 2.2.1. LOGIT LENS: The logit lens decodes intermediate LLM representations into tokens by applying the model’s unembedding layer.
- 2.2.1. LOGIT LENS: The logit-lens display places layers on the y-axis and generated-token positions on the x-axis, with middle-layer decodings generally more interpretable.
- Causal tracing compares corrupted and clean hidden states to identify where facts are encoded in the network.
- 2.2.3. STEERING VECTORS: Steering vectors modify model behavior by adding activation vectors during the forward pass, with γ controlling the steering strength.
3. Datasets
The paper uses purpose-built multilingual resources for steering analysis and a translated factual dataset for comparing how facts are encoded across languages.
- LLM-Insight is a released dataset designed specifically to study steering in LLMs.
- Prompts are open-ended enough to permit semantically unrelated completions, while still allowing the target word to appear as the next token.
- Sentence pairs such as “good” and “bad” provide steering directions, while unpaired words use general counter sets.
- The City facts dataset tests capital-city retrieval and is augmented with German, Dutch, and French translations for cross-lingual encoding analysis.
4. Experiments
The experiments examine how multilingual LLMs represent and manipulate information across languages, finding English-centric routing for semantic words, stronger English-based steering, and shared cross-lingual fact representations.
- Inspecting the latent space: Logit-lens analyses show that nouns and pronouns in non-English prompts can be routed through English representations, while coordinating conjunctions may remain outside English routing.For the Dutch prompt, ‘fruit’, ‘kweken’, and ‘they’ are routed through English words, whereas ‘en’ is not.
- Inspecting the latent space: Lexical words, especially nouns and verbs, are often selected in English, whereas adpositions, determiners, and compositional conjugates are infrequently routed through English in Aya-23-35B and Llama-3.1-70B.The evaluation prompts each model to generate 720 sentences and checks whether English equivalents appear in decoded latent representations.
- Inspecting the latent space: English-routing varies by model: less multilingual or smaller models generally route through English more frequently, although pretraining diversity does not explain every model difference.Cross-lingual homographs are included in the measured routing but cannot be disambiguated as English routing.
- Cross-Lingual Steering: English steering vectors generally outperform vectors generated in the desired output language when steering models across topics and prompts.Steering success requires including the target word while avoiding incoherent or stuttering output.
- Cross-Lingual Steering: Steering vectors for the same concepts have relatively high cosine similarity across languages but retain language-specific components.Their similarity is especially high in early middle layers, and the vectors are therefore not fully language-agnostic.
- Investigating the Representation Space: Causal traces place facts in similar layers across languages, and interpolating between bilingual hidden states preserves accuracy while changing only the output language.Models nevertheless show a propensity to answer in English, providing further evidence for an English-centric representation space.
5. Limitations
The paper identifies several constraints on interpreting multilingual LLMs, including tokenization differences, language confidence and confusion, steering variability, and limited alternative methods. These issues motivate caution in cross-lingual comparisons and indicate areas for future work.
- Tokenization: Different tokenization lengths complicate cross-lingual comparisons and remain an important consideration for future multilingual interpretability methods.The paper uses heuristics for per-token analyses such as causal tracing, but tokenization remains unresolved.
- Language confidence and confusion: Language confidence can require higher noise levels in causal tracing, while unaddressed language confusion may affect the analysis.The authors mitigate probability-related issues but avoid altering language confusion because they aim to study natural model behavior.
- Factors affecting interpretability methods: Steering results depend on prompt steerability and can push activations outside the expected data distribution, producing unintended outputs.The authors use a dataset designed for comparable steerability and check generated outputs for stuttering, but call for more robust evaluations.
- Other methods: Training sparse autoencoders for every layer is computationally expensive, while available pretrained alternatives are predominantly English-trained and may introduce bias.This makes SAE-based multilingual analysis difficult within the paper’s computational budget.
6. Related Work
Related work examines multilingual LLMs from internal and external perspectives, asking how representations and operations produce language-specific behavior. Prior studies debate whether concept spaces are universal or shaped by dominant training languages, while documenting effects on performance, fluency, and cultural bias.
- Internal and external perspectives: Internal analyses study latent representations and model operations, whereas external analyses evaluate multilingual outputs and task performance.The paper argues that combining both perspectives helps connect mechanisms with real-world effects.
- A three-phase view: Mechanistic interpretability describes forward passes as detokenization, processing abstract units, and selecting outputs in the target language.These phases cover forming referents or instructions, reasoning over abstractions, and mapping concepts to output tokens.
- Universal representations: A central question is whether multilingual models use a universal concept space, meaning equivalent concepts share representations across languages.The paper illustrates this with equivalent English and Dutch representations of “cat.”
- Competing findings: Prior work reports both language-agnostic concepts aligned with English and language-specific spaces centered on dominant training languages.Studies differ by language, model, task, and whether they examine single-token, translation, or open-generation settings.
- Behavioral effects: Multilingual LLM performance often favors English, with cross-lingual prompting and translation strategies sometimes used to improve other-language tasks.Related work also reports language confusion and unnatural non-English phrasing, including when outputs use the correct language.
- Bias and culture: Related studies report Western or culturally specific biases, including weaker performance on facts, proverbs, and conversational content from non-Western contexts.These findings connect multilingual model behavior with cultural representation and reasoning challenges.
7. Conclusion
The conclusion finds that multilingual LLMs make semantic decisions in representations close to English while processing non-lexical words in the prompt language. Cross-lingual factual representations occupy similar regions, but English-centric latent structure may improve some behavior while contributing to downstream, fluency, and fairness problems.
- Conclusion: Semantic decisions are predominantly made near English, while non-lexical words are processed in the prompt language.The paper also reports that English-derived steering vectors work better for non-English prompts than vectors from the prompt language.
- Conclusion: Factual knowledge across languages is stored in roughly the same model regions, and interpolation preserves predictive accuracy while changing the output language.Interpolated outputs are most frequently English, reinforcing the English-centric bias of the latent space.
- Implications: An English-centric latent space may improve performance when models translate non-English prompts into English, but it may also contribute to Western-centric biases.The conclusion connects this possibility to prior observations about multilingual LLM behavior.
8. Discussion
The findings favor an English-centric, rather than purely universal or input-language-specific, account of multilingual representations, while showing that task design and model size affect observed sharing. These nuances matter for fairness and robustness.
- The work aligns more closely with an English-centric, training-dominant representation space than with fully universal concepts.It also considers language-specific representations determined by the input or output language.
- English steering vectors outperform target-language vectors for interventions, challenging the expectation that a truly universal concept space would treat languages equally.This result is consistent with concurrent work finding English steering comparable to or slightly better than target-language steering.
- The apparent disagreement with prior work may reflect a distinction between concepts encoded in the model and concepts used during generation.The paper also notes that single-token selection and open-ended continuation tasks can produce different conclusions.
- Open-ended generation reveals English-centric latent representations for semantically loaded words, but not for non-lexical words.This setting permits analysis across different parts of speech, unlike single-token tasks.
- Smaller models may emit more shared representations, whereas larger models permit more language-specific representations.The paper presents this as a supported hypothesis rather than a universal rule.
- Representation sharing across languages is nuanced, with task and model size likely influencing behavior and with implications for fairness and robustness.
Impact Statement
Because LLMs are widely deployed across linguistic and cultural contexts, understanding their English-centric behavior is important for equitable and reliable outcomes. The findings also suggest that language-specific safety tuning may matter, especially where safety coverage is limited.
- LLMs’ English-centric semantic decisions may contribute to biased behavior in diverse linguistic and cultural contexts.The paper frames this as relevant to achieving equitable and reliable outcomes.
- Language-specific safety tuning may be needed because non-lexical representations do not show the same English-centric bias.The authors also observed language-dependent vulgar terminology, particularly in languages lacking safety tuning.
- The latent presence of vulgar terms may increase jailbreak vulnerability, although it does not necessarily imply vulgar model outputs.
A.1. Causal Tracing
The paper constructs and analyzes multilingual data and representations using causal tracing, steering, and related interpretability procedures. The supplied passages describe fact localization, interpolation, tokenization, prompt construction, and language-specific latent phenomena.
- Causal Tracing: Causal tracing corrupts information, then restores hidden-state components to identify where a fact is encoded.The example tracks the representation needed to recover the answer “Ottawa.”
- Causal Tracing: Noise is added to the subject-token embedding, and the corrupted embedding is propagated to measure the probability of the correct output.The clean embedding represents the prompt “The capital of Canada is.”
- Dataset Construction: The dataset targets open-ended sentence generation and includes sentences, target-word variants, and prompts admitting semantically distinct continuations.Native speakers review and correct the generated data.
- Dataset Construction: Prompt quality matters because a nearly forced continuation cannot reveal how interventions influence alternative model behaviors.The paper contrasts a constrained “money” prompt with a more open-ended “bus” prompt.
- Dataset Construction: The study varies multilingual word forms by token count, English overlap, and part of speech across the evaluated languages.English words require fewer tokens on average than words in other languages.
- Steering: Topic steering vectors have approximately 0.8 cosine similarity across languages, increasing when language vectors are incorporated.This suggests interaction between shared topic and language-specific representations.
- Cross-Lingual Facts: Causal traces localize facts in similar layers across languages, with mid-layer subject-token and later attribute-recollection traces.
- Cross-Lingual Facts: Interpolating hidden states generally yields accuracy between the two languages’ performances, while models most often answer in English.Mixtral-8x22B is especially likely to answer in English and least likely to answer in Dutch.