PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Chris Wendler, Veniamin Veselovsky, Giovanni Monea, Robert West

arXiv:2402.10588v4cs.CLcs.CY

TL;DR

The paper investigates whether English is an implicit pivot in multilingual LLMs, a question relevant to understanding model function and linguistic bias. Using controlled single-token multilingual prompts and layerwise analyses of Llama-2 latents, it finds an English-biased abstract concept space rather than evidence of literal translation into English and restarted processing. The study is limited mainly by its focus on Llama-2 and its reliance on model parameters and controlled toy-like tasks.

  • Problem

    English-dominated multilingual LLMs perform well in non-English languages, but whether they implicitly use English as a pivot—and how this relates to linguistic bias—lacks empirical evidence.

  • Method

    The study uses carefully designed multilingual prompts with unique single-token continuations, applies the logit lens across Llama-2 layers, and analyzes latent embeddings geometrically.

  • Results

    Llama-2 latents progress from no appropriate decoded token, through English-favored semantic predictions, to input-language-specific output; geometric analysis identifies an English-biased concept space.

  • Takeaways & Limitations

    The paper characterizes Llama-2’s internal lingua franca as concepts biased toward English, suggesting English functions as a semantic rather than purely lexical pivot.

  • Takeaways & Limitations

    Claims are limited primarily to Llama-2 because the method requires access to model parameters, and the controlled tasks are simple and toy-like.

Abstract

from arXiv · show

We ask whether multilingual language models trained on unbalanced, English-dominated corpora use English as an internal pivot language -- a question of key importance for understanding how language models function and the origins of linguistic bias. Focusing on the Llama-2 family of transformer models, our study uses carefully constructed non-English prompts with a unique correct single-token continuation. From layer to layer, transformers gradually map an input embedding of the final prompt token to an output embedding from which next-token probabilities are computed. Tracking intermediate embeddings through their high-dimensional space reveals three distinct phases, whereby intermediate embeddings (1) start far away from output token embeddings; (2) already allow for decoding a semantically correct next token in the middle layers, but give higher probability to its version in English than in the input language; (3) finally move into an input-language-specific region of the embedding space. We cast these results into a conceptual model where the three phases operate in "input space", "concept space", and "output space", respectively. Crucially, our evidence suggests that the abstract "concept space" lies closer to English than to other languages, which may have important consequences regarding the biases held by multilingual language models.

1 Introduction

The paper asks whether English functions as an implicit pivot in multilingual LLMs and develops an empirical approach using Llama-2 latent representations. Its evidence instead supports an English-biased concept space between input processing and language-specific output.

  • Motivation: English-dominated training raises the question of how LLMs achieve strong non-English performance and whether they implicitly pivot through English.The authors connect this question to possible Anglocentric linguistic and behavioral biases.
  • Research gap: Whether LLMs use an internal pivot language had not previously been addressed empirically because intermediate states are high-dimensional vectors rather than discrete tokens.The study frames mapping these vectors to languages as an open problem.
  • Approach: The authors use logit lens to prematurely apply the language-modeling head to intermediate Llama-2 embeddings, exposing next-token distributions during the forward pass.Carefully designed prompts make the decoded token both semantically identifiable and attributable to a language.
  • Core finding: Intermediate layers first lack contextually appropriate decoded tokens, then favor English “flower,” and finally favor the Chinese answer “花.”This layerwise pattern motivates the paper’s analysis of an English-biased internal representation.
  • Interpretation: Direct geometric analysis suggests middle layers operate in an abstract concept space partly orthogonal to language-specific token space reached near the output.The authors interpret English proximity as concept-space bias, not literal translation into English followed by a restarted computation.
  • Implications: The paper argues that studying latent biases and their effects is important for developing trustworthy AI.It presents this discussion as a direction for implications and future research.

2 Related work

Prior work studies multilingual transfer, English-mediated methods, cross-language latent representations, and mechanistic interpretability. This paper builds on logit-lens analysis while avoiding tuned-lens training because it could erase the English-intermediate signal under investigation.

  • Multilingual language models: Multilingual language models support multiple input languages, while frontier models often perform better in English despite broader multilingual competence.The cited literature includes mBERT, mBART, XLM-R, mT5, XGLM, mGPT, BLOOM, PolyLM, GPT-4, PaLM, and Llama-2.
  • Multilingual language models: Researchers transfer capabilities across languages by aligning contextual embeddings, relearning embedding matrices, or repeatedly adapting them during pretraining.These approaches target more efficient multilingual capability transfer.
  • English as a pivot: Several methods explicitly leverage English representations, translation data, or English-only instruction to improve multilingual instruction following and reasoning.Related work also combines romanized and English data for low-resource languages.
  • Latent representations: Encoder-only multilingual models provide evidence for language-specific early layers followed by a more language-agnostic later space.This motivates comparing the internal organization of multilingual representations in autoregressive models.
  • Mechanistic interpretability: Mechanistic interpretability reverse-engineers neural networks using techniques including circuit discovery, task-specific training, causal tracing, and intermediate-state decoding.The paper situates logit lens, tuned lens, and direct logit attribution within this toolkit.
  • Methodological choice: The authors rely on logit lens rather than tuned lens because tuned-lens optimization could map English-like intermediate states directly to final non-English predictions and remove the signal of interest.Logit lens applies the final language-modeling head prematurely without additional training.

3 Materials and methods

The study analyzes Llama-2’s layerwise latent states using controlled multilingual completion prompts and logit-lens decoding. It compares probabilities for correct Chinese tokens and English analogs across tasks, languages, layers, and model sizes.

  • Models: Llama-2 models contain 7B, 13B, or 70B parameters, use 32, 40, or 80 layers respectively, and share a 32,000-token vocabulary.Experiments cover all three model sizes using 8-bit quantization.
  • Model architecture: Llama-2 is an autoregressive decoder-only residual transformer whose latent vectors retain their dimension while being updated layer by layer.Transformer blocks combine masked self-attention, feed-forward processing, residual connections, and RMS normalization.
  • Logit lens: Final latents are multiplied by an unembedding matrix to produce vocabulary logits, which softmax converts into next-token probabilities.The logit lens applies this normally final-layer operation to intermediate latents as well.
  • Prompt design: The prompts require an obvious, unambiguously language-attributable single-token continuation, enabling comparison between the correct input-language token and its English analog.The study constructs translation, repetition, and cloze tasks, with Chinese as the principal target and German, French, and Russian for robustness.
  • Evaluation: Figure 2 tracks Chinese and English probabilities over layers for translation, repetition, and cloze tasks across 7B, 13B, and 70B models.The figure reports 95% Gaussian confidence intervals over 353 translation inputs and 139 repetition and cloze inputs.
  • Word selection: The word-selection procedure uses Chinese words with single-token English translations and filters other languages for ambiguous shared token prefixes.The dataset contains 139 Chinese, 104 German, 56 French, and 115 Russian words.
  • Measurement: Language probability is defined as the logit-lens probability of the selected correct token in Chinese or English for each layer and prompt.These probabilities need not sum to one because other tokens and tokenizations remain possible.

4 Results

Across tasks and model sizes, intermediate predictions move from diffuse uncertainty to English-favored semantic predictions, then to target-language outputs. Geometric analyses show trajectories passing through English-associated regions before reaching Chinese-specific token space.

  • 4.1 Probabilistic view: Logit lens: Translation and cloze prompts show no notable probability for correct Chinese or English tokens early, followed by an English rise and late Chinese spike.On repetition, Chinese rises alongside English; other languages show English rising first.
  • 4.1 Probabilistic view: Logit lens: High early entropy accompanies near-zero Chinese and English probabilities, then drops when English probability rises and remains low during the shift to Chinese.Early entropy is about 14 bits, close to the roughly 15-bit uniform-vocabulary value.
  • 4.1 Probabilistic view: Logit lens: Latent trajectories first pass through an English token cluster and only later reach a Chinese cluster.The joint 2D visualization uses negative log-likelihood distances and multidimensional scaling to embed tokens and latents.
  • 4.2 Geometric view: An 8192D space Odyssey: The logit lens maps intermediate latent states to token probabilities, while geometric analysis studies those states directly as Euclidean points.The forward computation incrementally modifies each latent by adding residuals across layers.
  • 4.2 Geometric view: An 8192D space Odyssey: Latents lie on a hypersphere of radius d ≈90.1, while output token embeddings approximately lie on a sphere with radius 1.52.The output-embedding radius has mean norm 1.52 and standard deviation 0.23.
  • 4.2 Geometric view: An 8192D space Odyssey: Token energy stays around 20% before layer 70, then spikes as next-token predictions switch from English to Chinese.This pattern supports three layer-wise phases in the forward pass.
  • 4.2 Geometric view: An 8192D space Odyssey: Phase 1 spans layers 1–40 with 14-bit entropy and no dominant language; Phase 2 spans layers 41–70 with 1–2-bit entropy and English dominance.Phase 3 spans layers 71–80, with Chinese dominance and token energy rising from 20% to 30%.

5 Conceptual model

The conceptual model divides the forward pass into input, concept, and output spaces. Latents first build representations, then encode concepts with an English proximity bias, and finally select concrete target-language tokens.

  • Phase 1: Input space: Phase 1 builds a feature representation of the current token by resolving tokenization and integrating words into larger semantic units.This phase is not yet directly concerned with predicting the next token.
  • Phase 2: Concept space: In Phase 2, latents enter an abstract concept space that overlaps output-token space and supports low-entropy predictions across languages and synonyms.English variants lie closer to concept embeddings than non-English variants because of overwhelming English training exposure.
  • Phase 3: Output space: Phase 3 maps abstract concepts to concrete target-language tokens while discarding information irrelevant to next-token prediction.This information loss produces the observed spike in token energy.
  • Geometric interpretation: The toy geometric sketch represents language along one axis, concepts along another, and context-related information along a third.The actual model operates in 8192 dimensions, while the sketch uses 3D space and a hyperspherical trajectory.

6 Discussion

Llama-2 shows an English-biased intermediate representation, but the evidence supports a semantic concept-space pivot rather than literal translation into English. This pattern varies with task and tokenization, while motivating further study of downstream bias.

  • Latent embeddings are farther from the correct input-language token than from its English analogue, producing overwhelmingly English logit-lens representations.Much latent energy remains orthogonal to output-token embeddings, so English-like decoding does not imply a literal English restart.
  • The paper characterizes the internal lingua franca as concepts biased toward English, making English a semantic rather than purely lexical pivot.
  • Across translation, cloze, and repetition tasks, most languages exhibit an initial English phase before target-language output.The repetition task is interpreted as requiring semantic understanding before copying can occur.
  • The English-first pattern is less pronounced in repetition, with the input language rising earlier and sometimes simultaneously with or faster than English.For Chinese, all selected repetition words were single tokens, compared with 13% Russian, 43% German, and 55% French.
  • The authors propose studying whether English-biased latent concepts affect downstream behavior and how language mix or less Anglocentric tokenizers change that bias.

Limitations

The study’s evidence is constrained by its model family, simple single-token tasks, incomplete understanding of concept-space geometry, and the partial visibility of the logit lens.

  • The analysis focuses on Llama-2, limiting claims about other English-dominated models and excluding closed-source models whose parameters are unavailable.The authors report initial evidence that Mistral-7B behaves identically, but broader validation remains future work.
  • The tasks are highly controlled but toy-like, so the findings should be extended to culturally sensitive problems, popular use cases, and analyses beyond single tokens.
  • The structure of the proposed concept space remains poorly understood in its original high-dimensional form.Mapping it more fully is presented as a future direction for strengthening the conceptual model.
  • The logit lens approximates output beliefs but hides other intermediate information, which enters this analysis only as noise.

A.1 Word translation

The appendix describes how translation, repetition, and cloze datasets were constructed and how language probabilities were computed from possible word-start tokens.

  • Dataset construction: Translations were produced with DeepL using both individual words and cloze sentences to disambiguate homonyms.Translations were filtered when English and target words shared a prefix token or when cloze and individual-word translations disagreed.
  • Dataset construction: The appendix reports dataset sizes for aggregated translation, repetition, and cloze tasks.
  • Language probabilities: Language probability sums next-token probabilities over vocabulary tokens that can begin the correct word in each language.Starting tokens include prefixes with and without leading spaces, plus UTF-8 character tokenizations for Chinese and Russian.
  • Language probabilities: Translation statistics include total examples and single-token translations for language pairs.

B Additional results

Additional analyses report language-probability and energy patterns across languages, tasks, model sizes, and individual translation pairs.

  • Scope: Results are provided for Chinese, English, French, German, and Russian.
  • Language probabilities: Figures 5, 7, and 9 show aggregated translation, repetition, and cloze language probabilities, respectively.Individual translation-pair results appear in Figures 11, 13, 15, 17, and 19.
  • Language probabilities: Across almost all languages and model sizes, layers progress from early noise to English probability and finally the target language.The Chinese repetition task is the stated exception.
  • Energy: Energy plots cover the aggregated translation, repetition, and cloze tasks and the individual translation pairs.They are presented in Figures 6, 8, 10, 12, 14, 16, 18, and 20.
  • Energy: The energy plots are consistent with the theory outlined in Section 5.

B.1 Low-resource language Estonian

On low-resource Estonian, Llama-2-7B shows mixed task difficulty but retains the intermediate English-to-Estonian transition observed elsewhere.

  • Resource level: Only one of 99 Estonian words in the analysis is represented with a single token.
  • Copy task: In the copy task, Estonian probability exceeds English probability in intermediate layers, resembling Chinese.
  • Translation task: In translation, intermediate distributions concentrate on correct English tokens before transitioning to Estonian in the final layers.Final-layer translation success is significantly smaller than for the main-paper languages.
  • Cloze task: The Estonian cloze task has 0% final-layer success, with slightly greater-than-zero success in intermediate layers decoded to English.The authors suggest synonyms or human-created cloze examples might increase success probability.

B.2 Other models: Mistral

Additional figures extend the analysis across Mistral-7B and multiple Llama-2 translation, repetition, and cloze settings, using layerwise probability and energy plots.

  • Other models: Mistral-7B results are consistent with Llama-2 results, pointing at the universality of the findings.
  • Translation: Translation probability figures plot layer number against correct-token probability, contrasting English and non-English target-word probabilities across model sizes.Figures cover non-English and English input settings, including aggregated and individual translation analyses.
  • Translation: Translation energy figures plot layer number against energy for the corresponding model sizes and input-language settings.
  • Repetition and cloze: Repetition and cloze analyses provide analogous layerwise plots, with repetition copying a non-English word and cloze using German fill-in-the-blank examples.
  • Cross-language analyses: Estonian and Chinese analyses plot language probability and token energy across layers for copy, translation, and cloze tasks.
Loading 2402.10588v4…