A revised arXiv paper titled Mimir argues that current language modeling is organized around tokens, where corpora are split into tokens and models are trained on token-level objectives such as next-token prediction. The authors propose an alternative that works with concepts at a large multilingual scale. The submission is a replacement version of a cross-listed paper.
Researchers present a parallel corpus pairing Arabic and Russian scientific writing, addressing a gap in resources for these two major research languages. The work also introduces a benchmark for evaluating large language models on the corpus, aimed at supporting cross-community knowledge exchange on sustainability topics. It is a revised arXiv submission in the computation and language category.
Researchers ran a controlled LoRA fine-tuning experiment to see how gains from training on one task or language carry over to others. The work finds that transfer between tasks and languages is uneven rather than symmetric, meaning improvements in one setting do not reliably help elsewhere. The findings point to limits in assuming that fine-tuning benefits generalize broadly across multilingual, multi-task models.
A new arXiv paper proposes a reference-free way to measure whether text-to-speech systems render lexical tone correctly in languages where pitch changes word meaning. The authors note that standard character error rate scoring misses such errors, using Yorùbá words like "ọkọ" (husband), "ọkọ̀" (vehicle) and "ọkọ́" (hoe), which differ only by tone, as an illustration. The work aims to give a budget-friendly evaluation option for massively multilingual speech systems.
A new arXiv paper argues that multilingual LLMs suffer from using one shared vocabulary across all supported languages, which produces uneven compression rates between languages. The authors also note that large embedding and output matrices raise memory demands and slow processing. Their proposed modular tokenizer design assigns separate tokenization components per language to address both issues.
Researchers extend prior work on retrieval heads — attention heads that pull information out of context — from English to multilingual models. They identify retrieval heads and a separate class of retrieval-transition heads, and report that behavior differs across languages. The work is a revised arXiv preprint in computation and language.
A new arXiv paper introduces Turkish MMLU Pro, a benchmark built from 12,000 Turkish-language questions spanning 58 sections. Each item keeps its original question stem and five answer choices, allowing researchers to test whether adding more options actually improves measurement quality. The authors argue that extra options can reduce scores without making the assessment more valid.
A new arXiv paper proposes a multilingual agent system that delivers wildfire evacuation guidance to people who may not speak the dominant local language. The authors note that wildfire seasons now run roughly 84 days longer than in the 1970s, increasing the stakes for timely emergency messaging. The work frames language access as a core part of inclusive disaster response.
A new arXiv paper examines whether a multilingual or English-only encoder is the better starting point when adapting streaming speech recognition models to another language. The authors report that the deciding factor is the amount of training data rather than latency considerations, challenging the assumption that multilingual encoders are always the stronger warm start.
A new arXiv paper proposes an online adaptive sampling strategy for realigning multilingual language models, aiming to improve cross-lingual transfer to extremely low-resource languages. The authors note that existing realignment approaches typically use uniform or random sampling, which may underuse informative language pairs. Their method adjusts sampling dynamically as training proceeds within a distributed setup.
A new arXiv preprint examines why a language model may refuse a harmful prompt in English but comply when the same request is translated into another language. The authors argue that output-level testing alone cannot reliably capture this behavior, and propose a framework based on semantic fibers and cross-gram interference to describe how safety properties drift in overcomplete internal representations. The work is listed under cs.LG and cs.AI.
A new arXiv paper looks at how a model's pretraining domain profile affects Transformer performance on digital pragmatics in Arabic-English code-switched text. It compares MARBERT and XLM-R against a general-purpose BERT baseline. The work focuses on whether domain-targeted pretraining yields better results for this kind of mixed-language discourse.
Researchers added Portuguese-source evaluation data for Xichangana, Mozambican Nyanja, and Sena to the FLORES+ multilingual benchmark. The work examines how conflating closely related language varieties affects machine translation evaluation, comparing Xichangana against Tsonga and Mozambican Nyanja against Chichewa. It argues that merging distinct varieties into a single reference can distort measured translation quality.
Researchers present Enemray, a language model designed for general-purpose interaction in Hassaniya, a low-resource Arabic variety. Training follows a stability-plasticity objective intended to build strong linguistic and cultural competence while limiting degradation of the model's existing capabilities. The work is published as an arXiv preprint.
A new arXiv paper systematically examines how much energy large language models consume when serving users in different languages. The authors frame the gap between language coverage and energy use as a "language-energy divide," suggesting that inference cost varies by language and is not well documented. The work aims to give a measurement basis for evaluating the efficiency of multilingual deployments.
A research team reports on a multi-month engineering effort to build Sophea, a bilingual Greek-English automatic speech recognition system intended for production use. The system was assessed against nine production gates, including word error rates for both languages and language identification performance. The work is presented as a case study in the engineering work required to move speech recognition from research to deployment.
A new arXiv study systematically examines token merging as a way to cut the computational cost of large multilingual speech recognition models such as Whisper. The technique dynamically combines token representations during inference, and the authors test how its effectiveness varies with model size and fine-tuning. The work targets deployment efficiency for transcribing low-resource languages without language-specific training.
Researchers released CVSS-X, a large synthetic speech-to-speech translation corpus that reverses the direction of the earlier CVSS dataset. While CVSS translated 21 languages into English, CVSS-X supports translation out of English into 28 target languages. The corpus is described as a resource for training and evaluating multilingual speech translation systems.
A new arXiv paper examines multi-agent LLM pipelines by treating two design choices as independent variables: how agent calls are wired together and where the differences between agents come from. The authors evaluate this on multilingual, low-resource emotion detection, where labeled data is scarce. The goal is to clarify which gains come from the structure of the agent network versus from the diversity introduced between agents.
Researchers have introduced SEA-SpeechBench, a large-scale multitask benchmark for evaluating how well speech and multimodal language models handle Southeast Asian languages. The paper argues that existing evaluation frameworks for audio-capable models focus heavily on English, leaving the region's languages poorly covered. The benchmark aims to enable broader assessment of speech understanding performance across the area's many languages.
Researchers have introduced SEA-LION-Embedding, a set of text embedding models built for Southeast Asian languages and released with open, documented training resources. The work addresses a persistent gap in the field, where leading embedding models cannot be independently reproduced because their training corpora remain private. The release aims to support reliable performance on downstream tasks across the region's many languages.
Researchers present SWORD, a benchmark that systematically distorts facts from Wikidata and tests whether large language models notice the resulting errors in different languages. Their experiments reveal that models frequently fail to reject distorted statements consistently across languages, even when they perform well on standard multilingual question-answering benchmarks. The findings suggest existing evaluations can overstate a model's genuine factual understanding outside English.
A new arXiv paper introduces SALT, a technique for strengthening how individual tokens are represented within multilingual sentence encoders. These encoders are optimized to align whole sentences across many languages, supporting applications like translation mining and zero-shot learning for low-resource languages. The paper targets the weaker token-level alignment that results from this sentence-focused training.
Researchers have introduced MultiSynt/MT, an open synthetic parallel dataset totaling roughly 4.8 trillion target-language tokens spanning 36 languages. Because web-scale training data is heavily skewed toward English, the corpus is intended to give model builders far more multilingual material for pretraining large language models. The work is described in a paper posted on arXiv.
A new research paper introduces Orukeet, a speech recognition model that replaces half of an adapted Parakeet encoder's temporal filtering layers with 12,288 fitted Gabor kernels, which are then kept frozen. The remaining parameters are trained on multilingual and multi-accent speech data, followed by a final adaptation and checkpoint selection stage. The approach aims to extend strong transcription models to more languages and accents.
A new arXiv paper tackles the tendency of reasoning language models to think in English even when prompted in other languages, which limits access for non-English speakers. The authors show that how training data is mixed across languages is key to getting models to generalize and carry out reasoning in the user's language itself. The work offers a path toward making advanced reasoning capabilities usable beyond English.
A new arXiv paper introduces 5-Dialects-BN, a resource covering five Bangla dialects used to examine how transliteration choices influence large language model performance. The work targets the sharp accuracy drops LLMs show on low-resource, dialectally diverse languages, with Bangla being the world's sixth most spoken language. The authors position script and transliteration practices as an underexamined factor in this degradation.
A new arXiv paper introduces BuzzASR, a collection of more than one hundred monolingual Whisper models fine-tuned for automatic speech recognition in 102 languages. The language-specialized models are designed to cover languages that large general-purpose multilingual ASR systems often serve poorly. The release is presented as a resource for practitioners and researchers working on speech technology across many languages.
A new arXiv paper examines how large language models detect hate speech in Urdu, a language with roughly 246 million speakers that the authors say has been largely overlooked in mainstream AI safety evaluation, including nine years of the Workshop on Online Abuse and Harms. The researchers propose 'Missed-in-Urdu' scores to quantify inconsistencies in how safety systems treat equivalent content across scripts and languages. The preprint appears in both the AI and computational linguistics categories on arXiv.
A new arXiv paper investigates how well multilingual large language models handle open-ended text generation in Urdu, a low-resource language. The authors argue that models marketed as multilingual often fall short in cultural and linguistic correctness outside high-resource languages. The work contributes to broader questions about the reliability of these systems for non-English users.
A new arXiv paper examines whether clinical annotation projection between languages can be treated as a document-level generative task that keeps the original text intact. The work spans six languages and aims to output character-level annotations that can be verified automatically. The stated goal is to support the construction of multilingual clinical corpora.
Researchers post-trained a 27B Qwen model to adopt a Korean response style covering verbosity, list and markdown formatting, discourse structure and register. They then measured two behaviors the training objective never targeted, including abstention on ambiguous social questions in the KoBBQ benchmark. The study reports that style-focused alignment produced measurable side effects on these untrained behaviors.
A new arXiv preprint argues that word error rate alone hides important failures when speech recognition systems and audio language models handle code-switched speech. The authors propose an evaluation method that accounts for language switches, and apply it to English-Yoruba audio, a low-resource pair with diacritics. They find that strong monolingual benchmark scores do not carry over to this setting.
NeoMME is a newly introduced encoder model designed to natively process multiple input modalities while supporting a broad range of languages. Its architecture emphasizes efficiency, aiming to deliver competitive performance at lower computational cost for workloads such as embeddings and retrieval. The release is detailed on the Hugging Face blog, making the model accessible to developers building multimodal, multilingual applications.
WHY IT MATTERS ↘For teams running embedding and retrieval pipelines, a single efficient encoder covering many languages and modalities could reduce inference costs and simplify architectures that otherwise chain separate per-modality or per-language models. It also signals growing competition in a segment long dominated by English-centric encoders, potentially lowering the barrier to shipping multilingual, multimodal search applications.