Multilingual Language Models

Overview

Multilingual language models study how systems can understand, reason, translate, and generate text across languages, including those with limited training data or distinct scripts. Common approaches include multilingual pretraining and data balancing, continued pretraining and instruction tuning, language-aware tokenization, and transfer methods such as shared representations, adapters, and mixture-of-experts models. Researchers evaluate performance with translation, comprehension, reasoning, and safety benchmarks, while probing internal representations and using inference-time steering to improve language control. A central tension is how to expand language coverage and transfer shared knowledge without degrading performance, reinforcing English or high-resource-language biases, or overlooking differences between languages and scripts.

Latest papers

278 papers19 papers in the last four weeks, down 44% on the four weeks before. 0.1% of all new papers.

Nov 16, 2026

Beyond Bilingual: Studying Quadrilingual Code-Switching with Conversational AI

Multilingual speakers choose languages by topic and conversation partner. Large language models (LLMs) support these languages unequally. Research on code-switching with conversational AI has mostly examined bilingual pairs of well-resourced languages. In Curaçao and Aruba, many people use Dutch, English, Spanish, and the low-resource Creole Papiamentu/o. In a preregistered study, 42 university students held two five-minute chatbot conversations, one about university life and one about local food and culture. Surveys assessed language background, AI use, and switching comfort. 20 international students in Curaçao formed a comparison sample. Participants switched between turns and rarely mixed languages within a message. English dominated both conversations. Participants switched more often in the cultural conversation, where Papiamentu/o and Spanish rose from 15% to 24% of messages. Self-reports showed no lower comfort switching with the chatbot than with a person. Conversational agents must follow turn-by-turn language changes and handle local low-resource languages well.
Nov 12, 2026

Lara Mini Base: Language Diversity as a Resource, Not a Trade-off

General-purpose language models are trained mostly on English: over open models we review, at most, non-English pretraining data makes up only a 12% share. Models built for European languages give them far more, but score well below general-purpose models of the same size. A common concern is that, in small models, languages compete for capacity. We present Lara Mini base, a model trained from scratch on 6T tokens in English and 44 other languages, which together cover all 24 official languages of the European Union. As training progresses, the 44 other languages together come to outweigh English. To give these languages enough capacity at low cost, Lara Mini base uses a mixture of experts architecture with 7.2B total and 0.6B active parameters. Among open base models with up to 2B active parameters, it achieves the highest average score across standard English, multilingual, math, and code benchmarks, ahead of models trained on 4 to 6 times as many tokens. Its margin is largest on multilingual benchmarks in the languages it supports and on math benchmarks. It supports contexts of up to 128k tokens. We find that our MoE architecture reduces capacity competition between languages. For a single user on one GPU, without quantization, it generates text at least 1.8 times faster than any model with a higher average score. We release the weights of Lara Mini base under the CC BY-NC-SA 4.0 license. In addition, we will release T-HellaSwag, a higher-quality machine translation of the commonsense reasoning benchmark HellaSwag into 110 language variants.
Nov 10, 2026

When does text encoding matter for language modeling?

UTF-8's role as the de facto encoding standard in language models introduces cross-lingual inequities due to its variable-length design. However, the effect that this encoding choice has on model performance remains largely unexamined. In this work, we investigate how text encoding impacts the performance of byte-level language models, comparing UTF-8 with a wide range of other encoding strategies, including other Unicode-based approaches, language-specific encodings, and our own customized encodings. To facilitate our analysis, we propose Aether: a hybrid encoding scheme that uses script-specific encodings for our target setting while maintaining full Unicode coverage through UTF-8 fallbacks, using this framework to model four orthographically diverse languages. Surprisingly, we find that while languages with information-dense characters (Chinese, Korean) significantly benefit from compressive encodings, languages whose characters carry less information individually (Georgian, Tamil) are largely insensitive to encoding choice. In light of these findings, we argue that text encoding should be a language-specific design decision, rather than a universal default.
Nov 8, 2026

MoTo: Mixture of Tokenizers Towards Fair Multilingual Language Modeling

Large language models rely on a single tokenizer that is chosen at training time and fixed thereafter. Most tokenizers are trained on English-dominant corpora, making them under-serve morphologically rich and non-Latin-script languages, where they can produce fragmented representations, longer sequence lengths, and reduced information density during training. We propose Mixture of Tokenizers MoTo, a modular tokenization framework that trains dedicated per-language BPE tokenizers and composes them into a unified superset vocabulary, where tokens shared across languages are deduplicated into a common ID space. Each tokenizer maintains its own normalization policy, pre-tokenization rules, and vocabulary budget. Despite allocating significantly fewer entries to English, MoTo retains English performance while delivering consistent gains across 18 out of 21 typologically diverse languages spanning six scripts, with improvements most pronounced on morphologically complex and non-Latin-script languages that are systematically underserved by English-dominant tokenizer design. Our results suggest that modular subword tokenization is a practical and extensible alternative to monolithic tokenization for multilingual parity.
Nov 6, 2026

I play of the cello: Evaluating Constructional Bleeding in Multilingual LLMs with CALQUE

Languages use different morphosyntactic strategies to express the same meaning. Speakers of multiple languages are able to apply strategies from one language to another, either by mistake or intentionally, to produce a literal translation, or calque. Considering that large language models (LLMs) have been shown to generalize crosslingually, despite being predominantly trained on English, we investigate whether they overgeneralize constructional strategies from English to other languages. To test this, we create a dataset of 5583 sentence pairs across seven languages, matching literal translations of strategies from English into other languages, and literal translations of the other languages' strategies into English. We then quantify constructional bleeding by measuring how often strategies from English are assigned a higher probability than native strategies. We find that models often fail to assign higher probabilities to native strategies and tend to prefer English-derived strategies even when evaluating sentences in other languages. Our findings highlight the limitations of the current English-dominant training paradigm.
Nov 5, 2026

Language Defaults Before Content: A Causal Layer Map of Multilingual Generation

Multilingual generation requires large language models (LLMs) to determine both what to say and in what language to say it, yet these two commitments are difficult to disentangle from model outputs alone. We use translation as a controlled probe: for multi-parallel segments that a model translates well in one language direction but poorly in another, content is held fixed while language varies. Using activation patching, we write residual-stream states from successful runs into failing ones and compare content-matched states with content-independent controls, allowing us to trace content-specific and language-related signals across layers. Across three LLMs of different scales and architectures, we find a consistent three-stage pattern of multilingual generation. Intermediate layers first express a default English signal; content information becomes causally available one layer later; and the final target language is resolved only at later layers. The ordering of the English default before content also appears in a reading-comprehension task. Our sparse autoencoder analyses reveal a further dissociation: the intermediate language signal is only weakly represented by sparse features, whereas a later language distinction is clearly sparse, but resists intervention at the patched position. These results provide a mechanistic account of how language selection and content information emerge and interact during multilingual generation.
Nov 5, 2026

CardioLM: A multilingual suite of small language models for the cardiology domain

Background: Cardiovascular care generates large volumes of clinical free-text documentation in local languages. However, clinical natural language processing (NLP) resources remain concentrated in English, and no existing benchmark is cardiology-specific, multilingual, and broad in entity coverage. Methods: We present two contributions. The first is CardioCCC, a corpus of 508 cardiology case reports originally published in Spanish and extended to Czech, Dutch, English, Italian, Romanian, and Swedish through machine translation, annotation projection, and clinician review, with annotations covering diseases, symptoms, procedures, and medications. The second is CardioLM, comprising CardioBERTa, a set of language-specific encoders obtained by continued pretraining on cardiology and biomedical corpora, and CardioNER, a set of named entity recognition (NER) models fine-tuned on CardioCCC. Results: Cardiology-specific pretraining reduced pseudo-perplexity by up to a factor of eight, yet changed strict span-level F1 by at most 0.04, with benefits concentrated in languages whose backbones lacked biomedical coverage. The jointly trained multilingual XLM-RoBERTa-large models achieved the best strict F1 in nearly every language and entity type, exceeding the strongest monolingual system by up to 0.13 F1. A large share of residual errors concerned boundary detection: relaxed matching recovered 0.10-0.23 F1 for disease, procedure, and symptom, but only 0.01-0.07 for medication. Clinicians rated CardioNER well above a rule- and dictionary-based baseline, with the mean rating across entity types rising from 26% to 73%. Conclusions: Specialty-specific pretraining is not the decisive factor for downstream entity recognition, whereas multilingual joint training yields the strongest results in nearly all settings and a simpler deployment path for multi-country clinical information extraction.
Nov 4, 2026

NaijaFacts: A Framework for Disentangling Language from Knowledge in Facts about Nigeria

Multilingual benchmarks for African languages mostly translate generic questions into local languages, measuring linguistic competence but not whether models know facts about Africa. We introduce NaijaFacts, a benchmark of factual and statistical questions about Nigeria, authored by native speakers and grounded in authoritative official statistics, with every question available in English, Nigerian Pidgin, Yorùbá, Igbo, and Hausa. Answering similar questions for the UK and US yields a country by language matrix that disentangles a language effect (the model knows the fact but degrades in a low-resource language) from a knowledge effect (the fact is absent in every language). Across eight frontier and Nigeria-adapted models, the knowledge effect dominates: English-query accuracy is 49-50% for the US/UK but 20% for Nigeria. The language effect is smaller and tracks web representation, degrading from English to Yorùbá (9 points on Nigeria, up to 24 on US/UK questions). We find scale does not help: the strongest models reach 66-79% on the UK and the US but none exceed 31% on Nigerian statistics. Extending to all 54 African countries with our best performing frontier model shows the same profile continent-wide, 96% on culture and 98% on politics against 26% on statistics, so the gap is not specific to Nigeria.
Nov 4, 2026

Familiar letters, not related languages, let unseen languages into a shared space of meaning

Hundreds of languages have written text but almost no translations, and whether a model learning one from untranslated text builds on a related language or familiar letters is unclear: they usually coincide. Here we show that familiar letters, not a related language, let an unseen language into a shared space of meaning. GlyphMind, a tokenizer-free vision transformer that reads text as pixels and lets any sentence be shown in any script, was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, a language and script it had never seen, it retrieved English translations seven times as often as the strongest simple baseline. Rewritten in unfamiliar letters, Lao did not enter although its close relative Thai was in pretraining; Santali, which had none, entered in Devanagari, familiar from Hindi, but not in its own script. Transliteration into familiar scripts may thus open multilingual models to languages without translations.
Oct 28, 2026

Leaving the Cave by Breaking the Symmetry: Identifiable Structure Can Enable Cross-Lingual Generalization

Why do multilingual language models generalize across languages even without shared vocabulary? The Platonic Representation Hypothesis posits that models trained on different data, objectives, and modalities converge toward a shared statistical model of reality. Extending this intuition to representations inside a single multilingual model, we propose the Relational Identifiability Hypothesis: relations in a shared world can promote entity alignment by ruling out alternative cross-lingual correspondences, making corresponding entities more likely to become directly readable without an additional transformation. We test this hypothesis in a controlled bilingual Transformer whose two token-disjoint languages describe the same modular-arithmetic world. We find that the languages can develop a shared answer subspace without aligning corresponding numbers. Cross-lingual alignment emerges selectively when the available relations constrain number correspondences: subtraction makes zero task-level identifiable, inducing selective alignment, while an order-sensitive operation produces broad alignment across numbers. These patterns predict success in zero-shot cross-lingual transfer. Together, our results show that relational structure in a shared world can promote entity alignment and enable transfer.
Oct 26, 2026

Beyond IceBERT: Open-Data ModernBERT Encoders for Icelandic from 12M to 1B Parameters

IceBERT, the reference encoder for Icelandic, found that a cleaned web crawl is enough to train a strong model and that starting from a multilingual model can help on some tasks. We revisit both findings with the ModernBERT architecture. We train Icelandic encoders from 12M to 1B parameters on 9B tokens of open data, and rerun IceBERT's benchmarks and five newer tasks for 30 encoders under one protocol with confidence intervals. Our 1B model trained from scratch, to our knowledge the first monolingual Icelandic encoder of that size, is significantly better than IceBERT on all eleven tasks; warm-started from the English Ettin encoder of the same shape, it has the best score of all encoders on every task we rank. Because our model shapes match the English Ettin encoders and the multilingual mmBERT, we can also test the warm start under controlled conditions, at six model sizes and four token budgets. With our vocabulary transplanted, a warm start was never significantly worse than scratch training on the same Icelandic text, and from base size up it was not significantly worse on four tasks than scratch training on 4 times the text (8 times at base and 1B). As models grow, most of the transfer comes from the source model's transformer layers, so even an English source helps. Keeping the source vocabulary instead cost 5 F1 on question answering and nearly three times the compute.
Oct 26, 2026

Quality over Quantity: Rethinking Data and Model Assumptions for Bengali Coreference Resolution

Multilingual transformer models are often assumed to generalise across languages, but their effectiveness for coreference resolution in low-resource settings remains unverified. We investigate this gap in Bengali, a language with over 250 million speakers but limited annotated coreference resources. We evaluate transformer-based pre-trained encoders under multiple settings, including zero-shot, fine-tuning, data augmentation, and cross-lingual transfer. Our results challenge several common assumptions about model generalisation and data efficiency in low-resource NLP. We demonstrate that a small amount of high-quality annotated data significantly outperforms larger mixed-quality datasets (p < 0.01). We find that data augmentation, particularly back-translation, does not improve performance and can degrade results in coreference resolution. Cross-lingual transfer is also inconsistent: transfer from a distant language like English reduces performance, and transfer from the related language Hindi provides no consistent benefit across models. Several widely used multilingual models exhibit collapsed embedding spaces, rendering them unsuitable for similarity-based clustering tasks. These findings highlight the importance of annotation quality over data quantity and provide empirical guidance for developing coreference resolution systems in low-resource languages.
Oct 24, 2026

When the Weakest Language Keeps Moving: ARMOR-DP for Multilingual BabyLM

Multilingual BabyLM systems usually fix language sampling proportions even though the weakest language can change during training. We present ARMOR-DP, a modular framework that diagnoses imbalance arising from representation, exposure, capacity, and objective pressure, and uses per-language validation loss to update sampling. Controlled experiments show that equal character budgets create a severe representation artifact for Chinese; byte-level encoding removes much of it, while static Chinese protection can subsequently make English the weakest language. The dynamic controller tracks these transitions and achieves a better score--balance tradeoff than fixed protection. On official evaluation, curriculum-dynamic checkpoints outperform a matched static Chinese-protective baseline while using fewer Chinese epochs; scaling capacity yields the strongest internal ARMOR-DP system. These results support a simple conclusion: multilingual weakness is a training-time state rather than a fixed property of a language. ARMOR-DP therefore separates diagnosis from intervention, allowing protection to follow the current bottleneck instead of a predetermined language label.
Oct 23, 2026

"Not" is Not Lost in Translation: A Causal Analysis of Negation Representations in Multilingual LLMs

Negation occurs in every language, yet its surface realization differs drastically. Despite this, multilingual large language models (LLMs) translate negation robustly, raising the question of whether they share an underlying representation of negation across languages. We study this question in six open models across English, French, Vietnamese, and Chinese, using causal interventions on a synthetic contrastive dataset and a negation translation dataset derived from FLORES-200. Across models, we find strong evidence that negation representations are causally controllable and largely shared across languages. In particular, we show that a contrastive activation direction extracted from one language can induce or remove negation in the others, and directions extracted from different languages are substantially aligned. We further find that negation is represented in sparse features: a single sparse autoencoder (SAE) feature per model suffices to control polarity across languages, while responding selectively to negated inputs. Finally, both dense and sparse interventions generalize to four additional languages and FLORES-derived sentences. We release data and code to support future work: https://anonymous.4open.science/r/negation-translation/.
Oct 22, 2026

CrisiLangs: A Multilingual Resource for Crisis Warning Generation

Communicating with the public during an emergency entails issuing warning messages promptly. In Europe, it often requires producing them in more than one language. Recent work has shown that language models can be trained to generate warning messages as a crisis unfolds, but the available data covers only English. To fill this gap, we present CrisiLangs, a multilingual extension of CrisiText (an English dataset of crisis events paired with warning messages written according to expert communication guidelines) to Italian, Bulgarian, German, and Greek. Expert native speakers translated the warning messages of 1,000 events drawn from 411 scenarios, for a total of 5,000 event-message pairs across five languages. We test the resource to fine-tune a multilingual model on all five languages jointly, with and without the emergency guidelines in the prompt, and compare it with a zero-shot baseline. The fine-tuned models outperform the baseline on ROUGE-L, BLEU and BERTScore in every language and configuration. In a human preference study, low decoding temperatures receive the most preferences in all five languages and in both configurations, while the effect of the guidelines varies by language.
Oct 21, 2026

Scaling Multilingual Pretraining of Pixel Language Models

Pixel language models process text as rendered images instead of a finite vocabulary, which gives them strong cross-lingual transfer abilities. However, not much is known about how these abilities are influenced by larger-scale pretraining. In this paper, we train a family of MPIXEL models along three axes of variation: more languages, more scripts, and larger model capacity under fixed per-language compute. In evaluations on five tasks covering 83 languages, we find that scaling acts locally: each axis improves the added languages and scripts, while increased capacity mitigates the curse of multilinguality from pretraining on more languages. Transfer to unseen scripts improves with multilingual pretraining, but does not further improve with scaling. On typologically distant low-resource languages, our models outperform subword and character-level baselines across many writing systems. We also show that larger pixel models reserve more layers for visual processing, and that cross-lingual alignment weakens as more languages are added. Overall, our findings characterize how multilingual scaling improves pixel language models.
Oct 21, 2026

Same Answer, Different Reasons: Trace-Level Divergence in Multilingual Large Language Models

Multilingual large language models are increasingly deployed in settings where visible reasoning traces support verification, education, and safety auditing, yet standard evaluations typically measure only final-answer agreement or holistic response similarity. As a result, they establish outcome consistency without determining whether users across languages receive comparable supporting evidence. We introduce a trace-level divergence framework that compares visible reasoning traces as structured sequences, decomposing agreement into structural similarity, order consistency, and holistic similarity, with pair-specific calibration to enable comparison across languages. Across three reasoning benchmarks, seven languages, and six frontier LLMs, the natural baseline---embedding each whole response and comparing the two by cosine---systematically overstates agreement: 52.1%52.1\%--67.3%67.3\% of examples in the top quartile of holistic similarity are not in the top quartile of structural similarity. This dissociation is stable across sentence encoders and step segmentations, and blind bilingual annotation studies in two English-anchored language pairs show that human judgments align with structural similarity rather than with the holistic baseline. A same-language control shows that the level of divergence is language-specific: two English traces for paraphrases of the same question agree more closely than an English trace and its counterparts in other languages. Our results show that answer-level and holistic evaluations can substantially overestimate the consistency of multilingual explanatory artifacts, motivating trace-level evaluation for reliable multilingual deployment.
Oct 20, 2026

How to Use Parallel Data for Multilingual Pretraining

High-quality human-translated parallel corpora can improve cross-lingual transfer, but are typically too small for large-scale pretraining. Machine-translated web corpora help overcome this limitation, making parallel data available at pretraining scale. Standard multilingual pretraining, however, samples source documents and their translations independently, without explicitly using their parallel structure. We investigate how to use the parallel structure of the data to improve multilingual pretraining by concatenating each document with its translation in the same training sequence. We compare this parallel pretraining (PPT) approach with standard pretraining (SPT), training models from scratch at matched token budgets. Surprisingly, despite the benefits of parallel data reported in prior cross-lingual work, PPT alone does not consistently improve language understanding over SPT. Our further analysis suggests that although PPT builds a strong cross-lingual foundation, as shown by its better use of shared information across languages, the mismatch between parallel training contexts and single-language task inputs limits its gains in general language understanding. This finding motivates a two-stage approach following the pretraining-fine-tuning paradigm: we first use PPT to build a cross-lingual foundation, then continue with SPT to adapt it to standard language modeling. In both high-resource and low-resource settings, the two-stage approach consistently improves general language understanding and cross-lingual task performance over standard pretraining alone at matched training budgets.
Oct 19, 2026

Cross-Lingual Reusability of LLM Pipelines: Financial Extraction, Multilingual Reasoning, and Semantic Screening

Financial institutions increasingly deploy large language models (LLMs) in multilingual markets, raising a practical question: can pipelines developed and validated in English be reused across languages, or must they be rebuilt per locale? We test cross-lingual reusability along three dimensions--accuracy, reasoning-effort transfer, and behavioural stability--through three complementary tasks. Task 1 (cheque extraction) and Task 2 (Mintaka QA with pinned evidence across nine languages and nine model-effort configurations) test whether accuracy and effort settings transfer; Task 3 (FinanceBench pilot, 50 items over SEC 10-K filings in four languages) validates whether these patterns generalise to financial content. Accuracy reusability is broadly confirmed: Gemini 3.5 Flash achieves 95.2-97.6% extraction accuracy across prompt languages, and FinanceBench per-language means span 5.8pp (77.1-82.9%), though the language-difficulty ranking reverses between domains (Hindi is worst on Mintaka, best on FinanceBench at 82.9%). Effort reusability is domain-dependent: higher reasoning effort yields marginal gains on general-knowledge QA but up to 8pp on financial content, and effort configurations that help in English help proportionally in other languages. Behavioural reusability holds: citation-set overlap with English averages 77-79% Jaccard, with no single language consistently worse. Together, these results support qualified reuse of English-built pipelines, but per-domain, per-language validation remains necessary--neither semantic similarity nor self-reported confidence reliably predicts cross-lingual accuracy.
Oct 19, 2026

DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia

While language models have demonstrated remarkable capabilities, their performance remains disproportionately skewed towards high-resource languages. This disparity is particularly acute for the 700 million inhabitants of Southeast Asia, where vast linguistic diversity remains under-represented in existing benchmarks, often suffering from limited language coverage or narrow task scopes. To address this, we introduce DeepSEA, a comprehensive benchmark constructed from established datasets through professional human translation into 12 Southeast Asian languages, spanning both high-resource (e.g., Indonesian, Thai, Vietnamese) and low-resource (e.g., Burmese, Khmer, Lao) tiers across five distinct tasks: mathematical reasoning, scientific reasoning, question answering, abstractive summarization, and machine translation. Extensive evaluation of frontier proprietary and open-weights models reveals that while frontier models like Gemini-3.7-Flash maintain robustness across resource tiers, smaller models face severe performance degradation on low-resource languages. Crucially, models perform better in reasoning and summarization when pivoted through English rather than target languages, exposing a critical dependency on English representations, especially for low-resource languages. Beyond these observations, we provide a cross-lingual error analysis and a cross-script tokenizer fertility analysis. We hope these findings shed light on the development of next-generation inclusive multilingual models.
Oct 18, 2026

Where Are Language-Specific Features Represented in Multilingual LLMs? Everywhere

Previous work has argued that multilingual LLMs encode language-specific features most strongly in their initial and final layers, implying that intermediate layers support language-agnostic representations. This paper reconsiders this finding by performing four different experiments: three that undertake a fine-grained analysis of representations of Portuguese and Spanish, two closely related languages, and one large-scale language identification experiment that covers over 200 languages. The results of these experiments challenge the strict separation between language-specific and language-agnostic layers in multilingual LLMs, using Llama 3.2 1B as a test case. By analyzing and manipulating contextual sentence embeddings, we demonstrate that language identity remains strongly recoverable throughout the model rather than being restricted to initial and final layers. Moreover, we find that the model tends to classify ambiguous strings as closer to Spanish than to Portuguese, revealing a skew in model behavior that differs from human judgments.
Oct 18, 2026

One Head Away from Language Confusion: Single Attention Heads Across Model Families

Large language models sometimes reply in an unexpected language, a failure known as language confusion. We ablate individual attention heads in 12 model checkpoints spanning seven model series, using FLORES-200 continuations in five languages and Language Confusion Benchmark (LCB) chat replies. In ten models, zeroing one head causes substantially more departures from the expected language than zeroing any other head in its layer; its LCB pass-rate drop persists, though reduced, under mean ablation. Seven heads move 8.0 to 51.5 percent of all FLORES continuations out of the prompt's language, versus at most 1.4 percent for other same-layer heads; the other three mainly affect explicitly requested languages. These seven heads matter less in base models given plain text. In two models, replacing the values a head reads at the requested-language name with another language's mean makes replies fail and shifts some into that language, unlike a random change of equal size. Replacing a head's output with another language's mean redirects 84 to 99 percent of scored non-English monolingual LCB replies into that language in five models. Yet the requested language's mean improves crosslingual replies beyond every control head only in the two Gemma models, whose baselines are low. Code and results are available in the public repository at https://github.com/christian-hoang-04/anonymous-release.
Oct 16, 2026

Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal

While deriving text embeddings directly from frozen LLMs has emerged as a promising direction, existing methods focus almost exclusively on English, leaving multilingual settings largely unexplored. In this work, we present the first systematic study of multilingual training-free text embeddings in LLMs. Our analysis reveals a language-conditioned embedding geometry in which semantically equivalent sentences remain misaligned across languages. We hypothesize that language-specific variation can be disentangled from semantic content to improve embedding quality. To this end, we propose Paired-Residual Subspace Removal (PRSR), a training-free inference-time intervention that steers LLM representations toward a more language-invariant semantic space. PRSR first constructs paired residuals from multilingual parallel sentences to capture language-specific deviations. It then estimates layer-wise low-rank subspaces from these residuals and suppresses the corresponding components in the last-token hidden states. We conduct more than 200 monolingual evaluations across 18 languages using MTEB and MMTEB, along with cross-lingual bitext mining between English and over 200 other languages. Empirical results demonstrate that PRSR consistently improves the performance of existing training-free baselines in both monolingual and cross-lingual settings across different LLMs.
Oct 15, 2026

Unfragmenting Nawatl NMT: Single-Variant Specialization for Masewalkopa

In minoritized (low-resource) language technologies, a common response to data sparsity is to aggregate text across multiple varieties and treat the combined data as a single corpus. However, mixed corpora of closely related indigenous variants can produce an artificial register that is accountable to no single speaker community, a structural failure we call variant fragmentation. This issue is acute for minoritized languages native to Mexico, such as Nawatl. Building on a 70,000-pair multi-variant foundation, we specialize an mT5-large checkpoint to Masewalkopa (nhn), the Nawatl variety spoken in Santa Ana Tlacotenco. We curate 2,665 parallel pairs, reserve all sentence-level data for evaluation, and incorporate 192 elder-transcribed target sentences via asymmetric backtranslation. A custom finite-state analyzer reduces unanalyzed surface forms from 82.0% to 19.9% and supplies functional syntax tags as plain-text prefixes, allowing the model to reuse existing subword embeddings rather than relying on frozen special tokens. The resulting LoRA-v9c-b model achieves 58.46 dev chrF++ and 58.20/57.90 directional test chrF++. Finally, CTranslate2 int8 quantization reduces CPU latency from 137.5 to 1.8-2.6 seconds per request, enabling cost-bounded Cloud Run deployment.
Oct 15, 2026

Multi-SFT: Permissively-Licensed Multilingual Instruction-Following

Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose terms of service forbid training competing systems. We introduce Multi-SFT, a commercially permissive collection of 502,818 instruction-response pairs across five European languages, built via translation and regeneration pipelines for tasks where translation is insufficient. Human evaluation yields inter-annotator agreement of at least 0.87 across all datasets and languages. Fine-tuning nine models from three different families (EuroLLM, OLMo 2, Qwen3; 1B-32B) and evaluating them on 13 benchmarks shows that Multi-SFT alone is competitive with larger augmented mixes and official instruction-tuned releases, with decisive advantages on safety tasks.
Oct 15, 2026

The Right Byte First: How UTF-16 Helps Multilingual Boundary Prediction

Many hierarchical byte-level language models group bytes into patches using a learned boundary predictor whose view of the next codepoint depends on encoding and lookahead. In UTF-8, codepoints of many scripts take three bytes: 8 of 24 bits mark length and continuation, and the first byte carries only 4 high codepoint bits. After the same left context, a mark extending a grapheme cluster and a letter starting the next can thus share their first byte. Universal Byte-Level Encoding (UBE) keeps one- and two-byte UTF-8 encodings but writes these codepoints as two UTF-16 bytes, low byte first, so its first symbol carries 8 low codepoint bits that can distinguish such continuations. We train small causal predictors on matched FineWeb-2 text and evaluate on FLORES-200 under UTF-8, UBE, and UBE's same-length big-endian variant. At one-symbol lookahead, balanced grapheme-boundary error at codepoint boundaries across 17 languages in 13 Brahmic scripts is 8.76% with UTF-8, 8.62% with the big-endian variant, and 0.06% with UBE. When the target instead retains a parent tokenizer's subword cuts at grapheme boundaries, UBE lowers the mean error relative to UTF-8 at the same lookahead by 72.91% in Korean, 35.29% in Chinese, and 67.32-99.55% in eight more scripts. In the seven of them whose main Unicode block fits on one 256-codepoint page, the big-endian variant stays within 3.0% of UTF-8, so the gain follows byte order, not length. In Korean and Chinese, UBE's mean error falls below UTF-8's final mean error after only 25% and 37.5% of training. Which byte comes first thus affects how much lookahead a byte-level boundary predictor needs.
Oct 15, 2026

Translation Circuit Discovery and Targeted Optimization for Low-Resource Multilingual Translation

Recent multilingual large language models achieve strong translation performance, yet the internal mechanisms responsible for low-resource multilingual translation remain insufficiently understood. In this paper, we study translation circuits in multilingual decoder-only LLMs for Hindi-to-Urdu and Hindi-to-Sindhi (Devanagari). We propose TITAN, a mechanistic framework that combines causal intervention, sparse feature decomposition, and targeted preference optimization to identify and optimize crucial translation components in the transformer. Unlike conventional full-parameter fine-tuning, TITAN restricts optimization to sparse, causally important circuits, enabling scalable and parameter-efficient adaptation while updating fewer than 5% of model parameters. Experimental results show that TITAN achieves BLEU improvements of 33.75% on Hindi-to-Urdu and 45.86% on Hindi-to-Sindhi over the strongest multilingual baselines. Human expert evaluation further confirms substantial gains in adequacy and fluency across both language pairs. Additional analyses reveal that multilingual translation is implemented through sparse, structured circuits operating over a shared semantic subspace rather than through an English-centric pivot representation. Overall, our findings demonstrate that mechanistic circuit discovery can directly improve low-resource multilingual translation while providing interpretable insights into multilingual transformer behavior.
Oct 15, 2026

Calibration Language Shapes Block-Influence Pruning in Multilingual Language Models

Block Influence (BI) ranks Transformer blocks for depth pruning from their input-output cosine change, but the ranking depends on the calibration distribution. We study whether English calibration transfers to Indian languages in Qwen2.5-1.5B-Instruct, Llama-3.2-1B, and Gemma-2-2B. English, Hindi, Marathi, and Tamil use disjoint Wikipedia articles for calibration and evaluation; every selector receives the same 150-chunk budget, and Tamil is excluded from multilingual aggregation to test cross-script transfer. Across all three models, late-block BI is 20.7-55.5% higher for the three Indian languages than for English, and every predeclared late-minus-middle contrast has a 95% bootstrap interval above zero. This mismatch changes the low-BI blocks selected for removal despite high global rank correlations. Averaging BI over equal English, Hindi, and Marathi budgets mitigates the resulting perplexity damage for Qwen at six removed blocks and for Gemma at all tested depths, including held-out Tamil. The mitigation is not universal: on Llama, balancing improves Hindi and Marathi at two removed blocks but severely harms Tamil, while larger pruning budgets select the same blocks and collapse perplexity regardless of calibration. Rank-mean and minimax-rank aggregators do not dominate the raw mean. These results support a diagnostic claim rather than a universal pruning rule: English-only BI systematically misses late-layer activity for Indian-language text, but multilingual calibration must still be validated per architecture and pruning budget.
Oct 14, 2026

Wolof-Adapted Industry-Ready Open-Weight Language Models

Wolof is spoken by 23M people, mostly in Senegal, but remains weakly supported by open-weight language models, while API models raise cost, governance, and sovereignty constraints for local deployments. In this paper, we present 0.6B-14B open-weight models adapted to Wolof through instruction tuning, preference alignment, and model merging. After presenting the pipeline and data, extensive experiments on 63~public and in-house tasks show good performance on Wolof understanding and generation while remaining effective on other languages and cross-lingual skills, with the 8B and 14B models being particularly competitive with substantially larger commercial systems. The models are currently being evaluated in internal customer-care pilots and are publicly available at https://huggingface.co/collections/Orange/wolof-lm .
Oct 14, 2026

Scaling Multilingual Mixture-of-Experts: How Language Coverage Changes Capacity and Compute Allocation

Multilingual models share capacity across languages, so adding languages can hurt performance on existing ones. Mixture-of-experts models expand this capacity with little additional per-token computation. Given this extra capacity, does adding experts reduce what each additional language costs? To answer this question, we propose a repetition-aware joint scaling law relating number of active parameters, training tokens per language, number of experts, and language coverage. Fitted on 26,300 per-language loss observations from 64 pretrained configurations, our law reduces prediction error on held-out configurations by 28% relative to the strongest evaluated baseline. At matched number of active parameters and training tokens per language, 32 experts reduce the loss penalty of expanding from 12 to 200 languages by 83% compared with a dense model. Our law further shows that broader language coverage favors more experts across a wider range of compute budgets, guiding capacity allocation as multilingual models scale.