multilingual chatbot developers

Abstract: LLM-based chat agents are increasingly used for complex, multi-turn, multilingual medical triage in low- and middle-income countries (LMICs), yet evaluation methods focus on single-turn, English-only benchmarks that primarily test the chatbots' knowledge. In practice, users are often confused, overwhelmed, or emotionally vulnerable, giving unclear, misspelled single-word answers or long messages in regional dialects, and the context of a conversation can shift rapidly. We propose SIM-UT (Simulation of User Behaviour for Triage), a multi-turn, simulation-based evaluation method that identifies ecologically valid, contextually relevant failure modes of LLM-based triage chat agents. Using SIM-UT, we evaluated three real-world WhatsApp-based health chatbots: a respiratory screening programme in India, a hearing health support programme in South Africa, and a diabetes and lifestyle coaching agent in Kenya. We compared a medically fine-tuned open model (MedGemma) against general-purpose frontier models (ChatGPT and Claude) across 5,000+ multilingual simulated conversation turns grounded in the personas, constraints, and communication patterns of the real user populations, over 2-4 iterative rounds. A key failure mode emerged in both model families: de-escalation risk. The agents initially identified medical emergencies correctly, but when the user minimised the symptoms without introducing new clinical information, they de-escalated, exhibiting clinically dangerous sycophancy, for example agreeing that vaginal bleeding in pregnancy was "not a problem" or that decreased fetal movements at 32 weeks "can be normal". We recommend user simulation-based behavioural evaluation, grounded in real user context and covering adversarial and edge cases, as standard practice for triage chatbots in production, so that safeguards can be added before deployment or in subsequent iterations.

Abstract: 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: $52.1\%$--$67.3\%$ 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.

Abstract: Negation remains a persistent challenge for language models across NLP tasks and modalities. We show that this weakness can undermine LLM safety guardrails through a structural vulnerability we term polarity blindness: the failure to recognize harmful intent after polarity-altering linguistic transformations. We study two forms of polarity inversion: grammatical negation and antonym replacement, and use synonym replacement as a polarity-preserving control. We further test whether an adversarial textual wrapper amplifies these failures. Using a 260-prompt benchmark across 10 state-of-the-art LLMs and 3 languages (English, Spanish, Czech), stratified by harmfulness potential and evaluated on a validated 0-5 harm scale, we find that polarity blindness affects most evaluated models, with mean harm scores reaching 4.12 for the most vulnerable model. The effect persists across typologically differing negation systems, suggesting that polarity blindness reflects a broader limitation in semantic robustness rather than a language- or grammar-specific failure.

Abstract: Large language models are expected to estimate their own uncertainty, with confidence signals used to guide retrieval, adaptive compute allocation, routing, and selective abstention. We investigate whether confidence signals measured at different temporal stages of problem solving encode distinct forms of epistemic information and remain reliable beyond English-centric settings. Inspired by classical meta-memory theory, we distinguish anticipatory pre-solve confidence (Feeling-of-Knowing, FOK) from retrospective post-solve confidence (Judgement-of-Learning, JOL). We evaluate these signals across two frontier LLMs, multiple reasoning domains, translated Marathi benchmarks, and a newly constructed benchmark of genuinely local Marathi factual questions. Across settings, the cheaper pre-solve signal remains weakly calibrated, while post-solve confidence generally provides stronger epistemic alignment despite its additional inference cost. Confidence behaviour transfers surprisingly well under direct language translation, with calibration trends largely preserved between English and Marathi, but degrades substantially on local Marathi knowledge, suggesting that linguistic transfer alone is insufficient for region-specific information needs. Finally, hallucinated responses maintain persistently elevated confidence after answer generation, highlighting important limitations of current LLM self-assessment mechanisms.

Abstract: Large Language Models (LLMs) increasingly operate across culturally diverse settings, yet existing evaluations largely measure cultural behavior at the output level. We introduce CultureLens, a reference-grounded benchmark and mechanistic study of how LLMs use cultural information. CultureLens contains 9,261 scenarios across 75 countries and systematically varies country, value, true-norm, and contradictory-norm evidence. Across Llama-3.1-8B, Gemma-3-12B, and Qwen-3.5-27B, stating a contradictory norm lowers aligned answers by 32-47 percentage points, even when the model answers correctly from the country alone. Mechanistically, models default to Anglophone norms when no country is given. The failure under a contradictory norm is not one of missing knowledge, as blocking the norm inside the model restores the correct answer in most cases. Models accept a stated norm according to how it is presented and not whether it is true, which leaves them most exposed to errors in retrieved content. Supervised fine-tuning with contradictory-norm examples raises aligned answers under a contradictory norm to 85-93%, including on unseen countries, at a small cost under a true norm.

Abstract: Platforms that host young users screen messages with commercial text-safety APIs, but independent evidence on how often such systems miss harm or flag harmless text is scarce, especially outside English, and the documentation of the one studied here leaves its false-positive rate unquantified. We evaluated one such API, Tuteliq Detect as deployed under the release identifier tuteliq-detect-v5.20261011 (not the separately published v6 model), on six harm categories: authorised push-payment fraud, money-mule recruitment, social engineering, romance scams, gambling harm and bullying. We built SITO (Synthetic Independent Text-safety benchmark with Overflagging tests; sito is Ukrainian for 'sieve'), 2,332 synthetic cases in English, Ukrainian and mixed Ukrainian, with each harmful case matched by hard harmless ones: same-topic messages, minimal-edit contrasts, warnings and retrospective accounts. On English text the system flagged 84.9-94.1% of harmful cases and 0-4.3% of harmless ones per category. It missed third-party reports of ongoing harm far more often than perpetrators' messages (63.6% against 95.1% recall, all languages), when sent without the settings the API offers for reports of harm. Ukrainian and mixed-language input was detected at least as often as English, recall was lower on text from a held-out generator family, so headline recall may be optimistic, and 3.2% of cases received a different decision in repeated runs. A platform should measure false positives on its own traffic and check how bystander reports are handled. The datasets, generation prompts and code are released.

Abstract: Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose ArabPref, which to our knowledge is the first large-scale pan-Arab preference data set for the alignment of cultural behavior conditioned on the country. Our contribution includes two resources in English and Modern Standard Arabic: a culturally grounded preference dataset for training and evaluation, and a multiple-choice benchmark designed to assess culturally aligned behavior across nations. All test data are validated by native speakers to ensure authenticity. To optimize cultural alignment, we experiment with post-training methods across multilingual and Arabic-centric language models, evaluating performance on both generation tasks and our cultural multiple-choice benchmark. By releasing an evaluation benchmark, we provide a foundation for systematically studying and improving culturally aligned behavior in language models.

Abstract: Activation steering has emerged as an effective and precise way to influence the behaviours of language models. However, composing multiple steering directions to control several behaviours simultaneously remains challenging, particularly in multilingual settings and with training-free steering vectors. We address this challenge with contextual steering, jointly training gates that adapt the strength of fixed steering directions to the current hidden state and interactions between interventions. Across a broad range of languages, behavioural attributes, and models, our approach outperforms constant-strength steering and four strong baselines, including prompting. It also better preserves general knowledge and keeps the model's internal computation closer to that of the prompted model. Our analysis further shows how jointly trained gates adapt their contributions during composition, explaining the mechanisms underlying effective compositional steering.

Abstract: Stereotypes are subjective constructs that are mediated by the cultural context in which they are viewed. Successfully modeling them thus also requires cultural competence. When using large language models (LLMs) for stereotype-related tasks, such as their detection, it is therefore crucial that models can access relevant cultural knowledge and adjust responses accordingly. Since LLMs have shown disparities in what cultural knowledge is available to them, their capacity regarding stereotypes may also be disparate across languages and cultural contexts (e.g., for non-English languages or for non-US contexts). In this paper, we systematically investigate the effectiveness of LLMs for culture-specific stereotype detection across three factors: (1) the prompting language, (2) the instruction tuning language, and (3) the fine-tuning dataset size. We conduct experiments for US, Dutch, and Chinese cultural contexts. Our results suggest that LLMs do not reliably account for cultural contexts and overpredict the presence of stereotypes. Non-English prompts seem to improve consideration of cultural context and thus lead to more effective predictions. Moreover, we find that additional instruction tuning reduces false-positive rates and that training data size moderates detection effectiveness non-monotonically.

Abstract: Step-level guards decide, before each tool call, whether an agent's next action is safe given the user's request and the trajectory so far. Users write those requests in many languages, while guard prompts and the benchmarks on which purpose-built guards are trained and evaluated are English. We ask whether a guard's verdict depends on the language of the request. We build LinguaGuardBench from the 1,220 labelled steps of TS-Bench AgentDojo-Traj: we translate only the user instruction into German, French, Turkish, Swahili and Amharic, and keep every tool call, observation and label byte-identical, so any change in a verdict is attributable to language alone. Six English rewordings and a whitespace placebo serve as controls, and managed guards are run three times to measure their noise. Across 19 guards, from frontier APIs to 4B open models and four purpose-built agent guards, most open, small and purpose-built guards change verdicts under translation beyond their own noise floor, and the effect grows with the rarity of the language: among the nine guards below 13B parameters, the mean placebo-adjusted flip rate rises from +1.5 pp on French to +4.4 pp on Amharic. Purpose-built guards are the most sensitive. The direction of the damage tends to follow guard size: open guards of 12-30B parameters block more benign steps (up to +5.0 pp false positives), and every significant loss of recall is in a guard below 10B (up to -7.4 pp). Frontier guards shift by at most about one point; the strongest, Gemini 3.8 Flash, returns identical verdicts with and without the request. The damage is concentrated in a few instructions and, in a post-hoc analysis, falls on steps where the agent's reasoning restates the English request, consistent with guards reacting to inconsistency between channels. English paraphrase moves every guard less than translation does. We release the benchmark and all 489k verdicts.

Abstract: Retrieval-Augmented Generation (RAG) models have become almost a cornerstone of open-domain question answering tasks; in which the quality of answers generated in the process, greatly relies on the effectiveness of retrieval mechanisms. In the realm of LLM, the last two years have seen tremendous progress in RAG architectures, with many models reporting to have improved performance through various design innovations. The cultural question and answering systems of South Asia, in particular, present a specific challenge since the region is home to the vast number of languages and traditions, which are closely interrelated. Along with these deeply intertwined historical narratives increase the need for culturally grounded and context-aware interpretation. Despite major architectural advances in LLMs, claiming significant improvements, challenges remain when answering these region specific questions in native languages. In this study, we aim to extract high-performing factors to guide the development of optimized hybrid RAG systems. We further explore whether the integration of RAG changes performance differently for culturally grounded and non-cultural questions in a South Asian multilingual setup, and whether RAG produces a higher overall percentage of correct answers than contexted answering. We also examine how LLM performance changes when RAG methods are integrated into the question-answering pipeline in comparison with non-contexted and contexted answering and, evaluate whether the introduction of RAG benefits popular or non-popular culturally grounded questions more. Not only does our work evaluate the current methods but also carves a path for future designs of more accurate and efficient question-answering pipelines.

Abstract: Do multilingual language models learn a universal internal geometry for color, and is it perceptually grounded or primarily linguistic? We decompose color representations into two geometric axes: (P1) alignment to perceptual color space, and (P2) cross-lingual symmetry. Analyzing layerwise embeddings across models and multilingual datasets, we find that perceptual grounding is weak: alignment to CIE-LAB peaks in intermediate layers, then collapses toward output layers. Cross-lingual symmetry degrades with depth, suggesting later layers encode language-specific structure. Functionally, we show that geometric properties predict task performance: models achieve 96.7% accuracy on color classification at layers where P1 alignment peaks, with P1 alignment (r=0.70) significantly correlating with downstream capabilities. Our results reveal a dissociation: text-only LMs learn universal color-language geometry, not universal color-perception geometry, providing diagnostics for where perceptual structure is preserved in language models.

Abstract: Persona vectors are linear directions in the residual-stream activations of LLMs that encode behavioral traits and enable fine-grained control over model behavior. Among them, the evil persona vector is particularly relevant to understanding harmful behavior, jailbreak vulnerabilities, and the unlearning of harmful knowledge. Yet whether this vector reflects a genuinely cross-lingual concept, rather than an artifact of its extraction language, remains untested. We conduct a systematic evil-vector transfer study across two high-resource languages (English, Spanish) and two low-resource languages (Swahili, Telugu) on four open-source instruction-tuned LLMs, and find that transfer is directionally asymmetric. Evil vectors are harder to obtain reliably from low-resource languages, and the resulting vectors are largely ineffective at steering, even within their own language. In contrast, vectors extracted from high-resource languages transfer effectively to low-resource inputs while largely preserving general knowledge, as measured by MMLU. Geometric analysis further supports this pattern: evil vectors from the two high-resource languages are closely aligned, whereas low-resource vectors align only weakly with this shared direction. These results suggest that a shared cross-lingual evil direction exists and, once reliably extracted from a high-resource language, can be reused to steer behavior in low-resource languages without additional per-language data collection.

Abstract: Processing multilingual text requires both identifying a broad range of languages and distinguishing the language of each token in inputs containing multiple languages. We introduce MixLID, a token-level language identification model supporting 111 language-script combinations. MixLID follows a two-stage training process: it first learns to distinguish individual languages using the language labels of monolingual examples, then continues training on mixed-language data. The second stage uses concatenated examples from existing code-switching data alongside augmented examples created through word substitution based on correspondences between English source words and loanwords. For monolingual identification, MixLID outperforms all baselines in both accuracy and macro F1 on four benchmarks within the language coverage shared with each baseline. For mixed-language identification, it outperforms the baselines in identifying the language of individual tokens on every evaluated language pair. MixLID achieves high performance across a broad range of languages in both monolingual and mixed-language settings. The model is publicly available at https://huggingface.co/nlpai-lab/MixLID.

Abstract: Pluralistic alignment asks that language models faithfully represent a society's diverse, competing values rather than collapse them into one dominant viewpoint, yet flagship pluralistic-alignment datasets remain heavily skewed toward Western respondents. We introduce NepPlural, a human-anchored benchmark for modelling competing socio-cultural personas in Nepali public discourse on youth migration and brain drain. We curate 2,148 code-mixed (Devanagari, Romanized Nepali, English) YouTube comments, annotate them along a four-axis persona taxonomy (Intent, Primary Driver, Value Orientation, Affect) plus an explicit No-Persona class, and build a 300-comment gold set labelled independently by three native Nepali speakers (mean Fleiss' $κ\approx 0.55$). Using this gold set to audit automatic annotation, we find that LLM annotation reliability is driven by protocol. Labels collected in batches through chat interfaces reach only near-chance inter-model agreement ($κ\approx 0.17$). Annotating one comment at a time raises agreement above the native-speaker level ($κ\approx 0.72$), but consistency is not validity: under a forced-choice schema, models agree on personas for the 39% of comments native speakers judge to express none, keeping macro-F1 against human gold near 0.50. Allowing abstention resolves this: models recover No-Persona comments with F1 of 0.79 to 0.87 and, against leave-one-annotator-out references, match or exceed individual native speakers (macro-F1 0.75 to 0.86 vs. 0.75). Encoders fine-tuned on the batched labels reach only 0.34. Finally, a human-judged steerability probe finds that three frontier models adopt single persona axes with 59 to 77% fidelity but compose full personas only 33% of the time, with no congruence gap detected at this scale between Western-congruent and Nepali-collectivist personas.

Abstract: Semantic relations organise word senses and are fundamental to lexical semantics. In order to evaluate semantic relation knowledge of large language models (LLMs), we propose a challenging task in which an LLM predicts the relation between two glosses (definitions of senses) presented in two different orders. The predictions for the two orders are evaluated for consistency with the symmetry or asymmetry of semantic relations. Our task is designed for both monolingual and cross-lingual evaluation, allowing us to examine whether relation knowledge remains accessible when the same senses are verbalised through glosses in different languages. We construct a dataset of 3,069 sense pairs, with each sense verbalised in English, Greek, Romanian, Turkish, and Japanese. Experiments with 17 LLMs in zero-shot and five-shot settings reveal generally poor performance. In each setting, only eight LLMs achieve a macro-averaged F1 score significantly above a random baseline, and only one (zero-shot) and two (five-shot) significantly outperform a neural baseline. In particular, LLMs often fail to generate bidirectionally consistent predictions for hypernymy and to make correct predictions for holonymy in either direction. Our results suggest limitations in current LLMs' semantic relation knowledge, both in identifying relations and in respecting asymmetry.

Abstract: Stance detection requires identifying an author's attitude toward a given target, sometimes based on conversational context. Jev, a specialized decision model designed for structured decision-making, offers an alternative to general-purpose large language models (LLMs). In this work, we evaluate Jev on two stance detection datasets, VAST (English texts) and ZS-CSD (Chinese conversations), comparing it with four general-purpose LLMs and two fine-tuned models. Results show that Jev achieves competitive performance on VAST, matching GPT-5.6 and outperforming the other general-purpose LLMs. However, it falls behind stronger LLMs on ZS-CSD, particularly in distinguishing favor from against. Further analysis suggests that this limitation may be related to understanding reply relationships and stance direction rather than conversation length alone. These findings highlight both the potential and limitations of Jev for stance detection.

Abstract: This study addresses the problem of structured sentiment analysis, whose goal is to obtain a fine-grained sentiment graph where the nodes represent spans of sentiment holders, targets, and expressions, while the arcs define the relationships among them. Our proposed approach casts the task as dependency graph parsing, but departs from traditional parsing methods by solving it through sequence labeling. To do so, we leverage recent advances in linearized graph encodings that allow each word in the input to be assigned a label, effectively capturing the structure of the dependency graph. We conducted experiments on seven datasets spanning five languages (English, Spanish, Norwegian, Basque, and Catalan), showing performance competitive with leading, more complex single-model approaches.

Abstract: Large language models (LLMs) often default to Modern Standard Arabic (MSA) when generating Arabic, even when prompted with dialectal Arabic. A natural explanation is that their internal representations are dominated by MSA. We test this hypothesis by adapting the language-dominance framework of Shani and Basirat (2025) (https://doi.org/10.18653/v1/2025.blackboxnlp-1.7) to 26 Arabic varieties. Across layers and model families, we find no evidence that MSA acts as a dominant internal representation for Arabic dialects. Instead, dialect representations form a dense and highly overlapping space: normalized mutual information drops sharply compared to patterns reported for more distinct languages. Moreover, the strongest separability effects are not confined to intermediate layers, but can shift toward later layers depending on the architecture. These findings challenge a common interpretation of MSA-biased generation: output preference does not necessarily reveal internal representational dominance. Analyses of multilingual and dialectal LLMs should therefore distinguish generation bias from the geometry of internal representations.

Abstract: During language acquisition, bilingual children are regularly exposed to code-switched input and use it as a cognitive scaffold to accelerate vocabulary growth and cross-linguistic syntactic mapping. In contrast, computational bilingual models are conventionally pretrained on interleaved monolingual corpora. While introducing synthetic code-switching during pretraining has become a promising strategy to enhance cross-lingual alignment and downstream performance, the structural and developmental parameters governing the success remain poorly understood. In this work, we investigate the efficiency of training with synthetic code-switched data across two typologically distinct language pairs by controlling two key variables: the structural location of code-switches and the dynamic switching rate across training stages. Our results show that training with code-switched data improves cross-lingual alignment for typologically close languages.

Abstract: In multilingual scenarios, queries with equivalent semantics but in different languages could guide the model into different reasoning trajectories, leading to performance disparities. To mitigate this gap, previous studies typically apply a one-size-fits-all query rewriting strategy, such as translation, which overlooks the fact that different scenarios require diverse types of semantic transformations. In this paper, we propose mRewriter-R1, an agentic multilingual query rewriting framework with reinforcement learning. Unlike single-turn rewriting, mRewriter-R1 formulates multilingual query rewriting as a multi-turn sequential decision-making process, where the model dynamically performs multi-aspect optimization through adaptive operator selection. Experimental results demonstrate that mRewriter-R1 outperforms all strong multilingual rewriting baselines on different large reasoning backbones. Further analyses show that the learned policy can adaptively decide on rewriting operators according to query characteristics, exhibiting strong generalization ability across diverse reasoning tasks, and plug-and-play compatibility with heterogeneous reasoning language models.

Abstract: Correction-based offline preference pipelines commonly treat model failures only as rejected responses under the original prompt. This supervision is incomplete for boundary failures: responses that violate the given instruction yet coherently satisfy a nearby intent or constraint setting. We introduce Bidirectional Preference Synthesis (BPS), a data-construction method for standard Direct Preference Optimization (DPO) that makes this missing prompt dependence explicit. For each validated boundary failure, BPS keeps the conventional forward pair under the original prompt and adds a reverse pair under a synthesized achieved prompt, so the same response is rejected where it is wrong and chosen where it is right, without changing the DPO objective, training a reward model, or requiring online sampling. On Qwen3-4B-Instruct-2507, BPS preserves original-side pairwise ranking while raising achieved-side ranking accuracy from 6.8% to 62.3% on held-out crossed anchors, with a similar shift under a Kimi-K2.6 cross-teacher probe. A blind human audit supports the intended reverse preference direction, and downstream evaluations show the clearest separation from Forward-DPO in multilingual multi-turn instruction following, with consistent capability-retention patterns on agentic, tool-use, and code checks.

Abstract: Reasoning traces improve large language models (LLMs), but current models are trained to reason mostly in English. It has been shown that forcing a model to reason in another language degrades accuracy, even when the reasoning language matches the language of the prompt -- but only for a setting where the model reasons over a short prompt. Here, we ask whether the same holds for retrieval-augmented generation (RAG), where the model must read and integrate a large amount of retrieved evidence in the target language. To study this, we build a fully monolingual German RAG question-answering testbed over the fictional world of the tabletop role-playing game The Dark Eye, a domain that is richly documented in German but too niche for the model to answer from memory, so that it has to rely on retrieval. Varying the forced reasoning language of an agentic RAG system on this testbed, we find that aligning the reasoning language with the language of the query and the retrieved documents helps. Forced German reasoning outperforms forced French, although the model benchmarks higher in French, so the benefit comes from alignment and not from language proficiency. The advantage grows when the retrieved context is richer and structure-aware. However, forced German only reaches the level of the model's native, unconstrained English reasoning without surpassing it, showing that native multilingual reasoning is needed. We publicly release the testbed and QA benchmark.

Abstract: Large Language Models increasingly serve as interfaces for knowledge-intensive information seeking tasks across languages by synthesizing multilingual evidence. Prior work has shown that they often exhibit query-language preference -- the tendency to favor sources written in the language of the query -- but has largely examined this behavior in settings where equivalent knowledge is available across languages. However, this bias becomes consequential when sources in different languages provide incomplete or inconsistent accounts of the same fact, since the information users receive then depends on the sources a model selects to use. To characterize query-language preference under such cross-lingual knowledge disparities, we introduce Waldo, a multilingual Question-Answering (QA) benchmark constructed from Wikipedia. Waldo contains 12K QA pairs targeting knowledge gaps, where a fact is available in one language but absent in another, and knowledge conflicts, where language editions provide conflicting versions of the same fact. Evaluating eight models across five languages, we find that when one language edition merely lacks the relevant fact, models generally use evidence from the other language regardless of the query language. Under conflicting accounts, however, model responses strongly align with the document in the query language, causing semantically equivalent queries to elicit different accounts depending on the user's language. Finally, we explore two different approaches that could mitigate this preference under knowledge conflicts: a mechanistic intervention that ablates attention heads associated with query-language preference, and LoRA-based training, which reduces the preference gap by up to 61.5%.

Abstract: Misinformation on social media remains a critical problem, and more and more people settle it by asking a language model instead of a fact checker. Whether models judge such claims reliably is debated; whether they judge them equally well in every language people ask in has gone almost unasked. We test eight models from five families, 3B to 70B, on 1,500 encyclopedic factual claims that exist in identical form in eight languages. English is judged better than every other language on every model, and the gap is widest on the smallest ones, where Llama-3B on Arabic is no better than guessing. Existing remedies retrain on more multilingual data or fit an unconstrained map between language representations, and neither asks whether the model already holds the answer and simply fails to say it. It largely does: a linear probe recovers the truth from the very activations the model fails to express. We propose RoSh, a per-language shift and rotation of the residual stream, computed in closed form at three layers, with no training and no weight modified. It improves every model and closes 75% of the gap on average, helping most where the model was worst: Arabic on Llama-3B goes from chance to nearly the English level, and a fifth fewer of the claims answered correctly in English are lost in translation. What remains is no longer a read-out failure: afterwards the head recovers as much of what is encoded outside English as it does in English. An unconstrained map fitted on the same pairs falls below the untouched baseline, so the orthogonality constraint is doing the work, and every model clears a scrambled-correspondence control and ten further controls. On the two benchmarks of the closest inference-time method, latent-space intervention, run with its own data and metric code, RoSh's gains are five to thirteen times larger.

Abstract: A key aspect of culture is a society's norms about everyday behavior. How accurately do large language models (LLMs) represent cultural differences in such norms? To answer this question we used the recent Global Study of Everyday Norms (GSEN), which collected ratings of 150 scenarios in 90 societies, as the human benchmark. We prompted GPT-5 to estimate each society's average rating for every scenario, and later repeated the benchmark in three other LLMs: GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro. Compared to GSEN estimates, all four LLMs misrepresented cultural variation in two ways. First, they greatly underestimated its magnitude, estimating differences between societies to be, on average, less than half their measured size. Second, for many scenarios the LLMs poorly identified the pattern of variation, that is, which societies judged the behavior less acceptable and which societies judged it more acceptable. The pattern of variation was identified better for scenarios that elicit concerns about vulgarity, especially scenarios involving kissing and flirting. We also found that norms in more developed societies tended to be estimated somewhat more accurately, and that prompting in local survey languages rather than English produced only a modest improvement in accuracy. Local-language prompting also reduced, but did not remove, the underestimation of between-society differences. Cultural differences in everyday norms are only weakly and unevenly represented by LLMs.

Abstract: The scarcity of non-English language data in specialized domains significantly limits the development of effective Natural Language Processing (NLP) tools. We present TransBERT, a novel framework for pre-training language models using exclusively synthetically translated text, and introduce TransCorpus, a scalable translation toolkit. Focusing on the life sciences domain in French, our approach demonstrates that state-of-the-art performance on various downstream tasks can be achieved solely by leveraging synthetically translated data. We release the TransCorpus toolkit, the TransCorpus-bio-fr corpus (36.4GB of French life sciences text), TransBERT-bio-fr, its associated pre-trained language model and reproducible code for both pre-training and fine-tuning. Our results highlight the viability of synthetic translation in a high-resource translation direction for building high-quality NLP resources in low-resource language/domain pairs.

Abstract: Chinese social media has generated a vast and continually evolving lexicon of internet buzzwords whose meanings are often non-literal and deeply rooted in local cultural and pragmatic contexts. Existing research has primarily focused on interpreting these buzzwords within Chinese, leaving largely unexplored whether LLMs can transfer such culturally grounded knowledge across languages and accurately convey the intended meanings in English. This cross-lingual capability is also critical for safety, as harmful expressions may obscure their offensive content through culture-specific homophony, euphemism, irony, or coded language. In this paper, we investigate the ability of advanced LLMs to understand Chinese internet buzzwords across languages. To this end, we introduce CIBuzzBench, the first benchmark for cross-lingual Chinese-to-English understanding of Chinese internet buzzwords. CIBuzzBench comprises 3,001 Chinese internet buzzwords annotated with English meaning explanations, English equivalents, category labels, and harmfulness labels. Based on these annotations, we design three evaluation tasks: Meaning Explanation, Cross-lingual Equivalent Matching, and Culturally Grounded Harmfulness Detection. We evaluate representative state-of-the-art proprietary and Chinese LLMs under both English- and Chinese-prompting settings. Our results show that LLMs continue to struggle with the cross-lingual understanding of Chinese internet buzzwords, particularly in fine-grained non-literal interpretation, robust equivalent matching under option perturbations, and calibrated harmfulness detection. These findings highlight the persistent challenges posed by culturally grounded language phenomena for multilingual LLMs and safety-oriented evaluation. The dataset and code are available at https://github.com/SuperYFan/CIBuzzBench.

Abstract: Large language models (LLMs) suffer from a long-tail deficit: culturally specific facts, particularly those concerning underrepresented regions such as Latin America, appear too rarely in pretraining corpora to be reliably memorized. Retrieval-Augmented Generation (RAG) addresses this by grounding generation in external text, but structured alternatives such as Knowledge Graphs (KGs) offer tighter control over what enters the context, along with potential gains in explainability and updatability. We benchmark Graph-RAG against standard RAG on LatamQA, a culturally grounded multiple-choice dataset spanning eight thematic categories. The graphs are built end-to-end from Wikipedia articles with KGGen, a recent open-domain extractor, without manual curation in our main setting. G-Retriever is competitive with RAG and reduces the error of the base LLM by 72\% with a standard KG and 78\% with a benchmark-aware variant, the gap to RAG narrowing further as the graph is oriented toward task-relevant content. The trained projection transfers zero-shot to Portuguese without target-language fine-tuning, indicating multilingual reach.

Abstract: Multilingual Large Language Models (LLMs) traditionally rely on a single vocabulary shared by all supported languages, which can lead to uneven compression across them. Moreover, their large embedding and output matrices increase memory usage and slow inference, notably for small-scale models. It is also wasteful as models are often used for only a subset of languages. To address these issues, we introduce a modular framework for multilingual model training. First, we propose methods to learn large modular BPE and Unigram tokenizers that enable extraction of subtokenizers tailored to any language subset. These subtokenizers achieve compression on par with monolingual tokenizers and improve cross-lingual fairness. Second, we design a pretraining strategy that samples subtokenizers to form batches, restricting predictions to the relevant vocabulary subset and allowing efficient training despite a large vocabulary. This supports efficient inference with any combination of language-specific vocabularies. Therefore, it reduces memory usage and speeds up inference in models without sacrificing performance.

Abstract: Multi-agent LLM architectures, such as LangChain and AutoGen, largely assume English as the lingua franca for internal inter-agent communication, even when the end-user task is non-English. We fill this gap by evaluating a two-agent extraction-answer core, with an additional back-translation agent in the English-forced condition, across four typologically diverse languages (Hindi, Chinese, Spanish, Arabic; n = 300 per language) using the Aya-23-8B model. We compare a native-language pipeline to an English-forced one (which incorporates a final back-translation step from English to the user's language). We discover a statistically significant English-Forcing Tax (surviving a strict Bonferroni correction) that isolates the cost of English routing from general multi-agent orchestration overhead. Forcing inter-agent communication through English reduces Exact Match accuracy by 13.0 percentage points (Spanish) up to 30.6 percentage points (Hindi) compared to native-language multi-agent execution. Using chrF scores as a diagnostic measure of English-reference lexical overlap, we find that lower overlap is strongly associated with pipeline failure, consistent with translation loss being an important contributor to the observed performance drop. These findings suggest a compelling case for native-language routing in agent frameworks when the source and target languages are typologically distant, reducing a compounding translation tax.

Abstract: This study highlights the role of domain-specific pretraining profile (DSPP) in Transformer performance for modeling digital pragmatics in Arabic-English code-switched discourse. It evaluates MARBERT and XLM-R(oBERTa), with BERT serving as a general-purpose baseline. The models were evaluated on their ability to classify context-sensitive pragmatic functions in code-switched social-media discourse. 11695 unique X posts were collected via Python and utilized for the study. The study employs a quantitative and qualitative NLP approach, following a supervised pipeline. Findings unveil that MARBERT consistently surpasses XLM-R with validation Macro F1 increasing from 0.39 to 0.84 and validation loss decreasing from 0.55 to 0.19. On an independent test set, it achieved 0.96 accuracy, 0.83 macro precision, 0.87 macro recall, and 0.85 Macro F1, while XLM-R achieved 0.92 test accuracy but a substantially lower Macro F1 of 0.52. This was also supported by class-level performance where MARBERT outperforms XLM-R considerably with F1 improvements ranging from +0.33 to +0.60, demonstrating a clear advantage in modeling Arabic digital pragmatics. The study concludes that Transformer performance depends more on DSPP than multilingual coverage alone, as the latter does not guarantee optimal performance on a highly specialized pragmatic classification task.

Abstract: Multi-agent LLM systems combine multiple inference calls, but prior work often confounds how calls are connected with how they are diversified. We study these factors independently: inference topology and source of inter-agent diversity. In a controlled $2 \times 3$ matrix, we cross parallel aggregation and sequential refinement with stochastic sampling, role prompting, and learned QLoRA specialization, under a fixed three-call budget and output protocol within each backbone. Using Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct, we evaluate all six configurations on multilingual low-resource emotion detection across nine languages. Parallel learned specialization is strongest on Qwen at 52.83 Macro-F1 and reaches 52.94 on Llama. On Qwen it also exceeds same-backbone zero-shot, few-shot, CoT, and seven-call self-consistency baselines. The preferred topology depends on diversity source: sequential refinement helps stochastic and prompted settings, while the learned Width advantage shrinks from 2.83 points on Qwen to 0.17 on Llama. Depth-wise analysis suggests that later learned specialists can overwrite correct early predictions, although the aggregate effect is backbone-dependent. Overall, how agents are differentiated produces larger performance shifts than topology, which should be evaluated jointly with specialization.

Abstract: Reasoning language models have made substantial advances on a variety of complex tasks, yet their capabilities remain overwhelmingly English-centric: models primarily reason in English regardless of the language they are prompted in. This is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge more readily expressed in the target language. In this work, we advance L2 reasoning, the ability of a model to reason consistently in the language of the user's prompt, thus building an in-language bridge between the prompt and the answer. We approach this problem from a data-centric angle, investigating how to optimize data composition and scheduling in SFT for reasoning generalization. Building Tiny Aya L2-Thinker at 3.35B scale, we achieve an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. We show the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. These findings indicate that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing and without requiring reasoning supervision in every target language. We release our model weights and multilingual reasoning data to support further research on accessible, in-language reasoning.

Abstract: Cross-lingual sentence encoders enable scalable transfer across hundreds of languages, powering applications such as translation mining and zero-shot learning in low-resource settings. Although trained for sentence-level alignment, they are increasingly also applied to token-level tasks such as hallucination detection and sequence tagging, exposing a mismatch between training and usage. We propose SALT, a lightweight post-training method that improves token representations by injecting span-level supervision into existing sentence encoders. Across five multilingual token-level benchmarks, SALT achieves the best overall results on four of them, outperforming alternative fine-tuning strategies and competitive encoders. It also improves sentence-level performance on cross-lingual retrieval and classification tasks. These results demonstrate that span-level supervision is an effective signal for improving both token and sentence representations.

Abstract: Language models often receive a question together with a claim about what another source answered. We audit whether such claims destabilize answers in multiple-choice question answering. For each item, we hold one wrong option fixed across misleading conditions and vary the cue template attached to it. We introduce \emph{neutral-conditioned misleading cue adoption rate} (NC-MCAR), which measures switches to that option only on valid cued trials where the same model first selected the gold answer under a neutral prompt. This is a measure of answer instability, not proof that the model knew the answer or that all deference is irrational. We evaluate four instruction-following models on MMLU-Pro and IndicMMLU-Pro in English, Hindi, Bengali, Tamil, and Telugu. Across 220{,}000 outputs, the expert template yields 41.1\% aggregate NC-MCAR, compared with 12.5\% for the majority template. These two conditions use the same wrong option and final instruction. Filler accuracy remains well above expert-wrong accuracy, while correct-cue prompts have high valid-response accuracy. The audit documents answer instability relevant to grounding under the tested forced-choice prompts: a bare, unverified source claim can outweigh an answer that was previously consistent with the task evidence.