We present the findings of the WMT26 Multilingual Instruction Shared Task (MIST), evaluating small (=10B parameter) open-weight language models across 24 languages on three sub-tasks: context-based QA, cross-lingual summarization, and open-ended generation. Submissions from eight teams explore fine-tuning, knowledge distillation, and prompt engineering. Our findings demonstrate that small models are not yet polyglots: while leading systems perform well on high-resource languages, performance drops sharply on low-resource varieties, and comprehension degrades significantly when transferring across non-English language pairs. In open-ended generation, human evaluation and automatic rule-based verification exhibit strong overall rank correlation (0.79) but diverge locally at the top: human evaluators favor distilled pipelines with superior fluency and cross-lingual equity, whereas automatic verifiers reward concise prompt-engineered systems. Furthermore, multi-stage pipelines suffer from prompt-language inertia on cross-lingual instructions, and summarization exhibits severe English leakage unless penalized by language gating. Finally, while baseline fluency is largely achieved by modern small models, multi-constraint instruction following and unanswerable question refusal remain the primary performance bottlenecks. We release all test sets, system outputs, human judgments, and evaluation code under a permissive license.

Translation cascades for reasoning translate the query from another language to English, reason in English, and translate the answer back to the original language. This is a competitive approach to multilingual reasoning, but structurally lossy, since each stage discards information later stages may need, including cues for cultural grounding, register, and disambiguation. We examine the benefits of a simple and training-free intervention: a context-aware translation cascade, which additionally provides the original question, the English-translated question, and the reasoning trace to the final cascade stage for target-language generation. We evaluate gains across ten multilingual benchmarks including various task types, three backbone models, and 285 high-, mid-, and low-resource languages, and demonstrate strong gains for open-ended generation across models and resource regimes. We show that the original language question carries most of the beneficial context. Our study emphasizes the need to better design information flow in multilingual reasoning cascades for mitigating error propagation, and provides a simple and actionable default strategy: preserve the original user question until the end of the pipeline.

Large language models (LLMs) show state-of-the-art performance in machine translation, but are also known to be sensitive to errors in user prompts. Given these models are largely trained on and respond best to prompts in standard English, this may affect the quality of LLM outputs for second language English speakers as well as real-world lay users, with potentially disproportionate effects on the former. We explore this effect by modeling a range of error types exhibited by such users, motivated by studies of L2 English, and quantifying their impact on LLM performance. We work with two related tasks: machine translation and machine translation evaluation. We find that LLMs-as-MT are brittle to natural spelling errors but not to errors at the phrasal level. However, the variance in quality caused by these errors is lower than the variance over the initial prompt choice, suggesting that “perfect English” for a given prompt is less important than choosing a good prompt. Since lay users and L2 speakers may use non-optimal prompts as well as display imperfect language skills, our work calls for increasing the resilience of model performance to both these phenomena to best serve a diverse user base, both from a robustness and fairness perspective.

While there is growing interest in in-context language learning (ICLL) for unseen languages with large language models, such languages usually suffer from the lack of NLP tools, data resources, and researcher expertise. This means that progress is difficult to assess, the field does not allow for easy large-scale experimentation, and findings on ICLL are often limited to very few languages and tasks. In light of such limitations, we introduce a framework (Rashid) for studying ICLL wherein we reversibly cipher high-resource languages (HRLs) to construct truly unseen languages with access to a wide range of resources available for HRLs, unlocking previously impossible exploration of ICLL phenomena. We use our framework to assess current ICLL methods for machine translation with SOTA evaluation tools and manual analysis, explore new directions leveraging otherwise expensive resources for improving ICLL, and test ICLL strategies for rich downstream tasks beyond machine translation. These lines of exploration showcase the possibilities enabled by our framework, as well as providing actionable insights regarding current performance and future directions in ICLL.

We introduce Omnilingual MT, a family of multilingual translation (MT) systems covering more than 1,600 languages—the broadest benchmarked coverage of any MT system to date. We make contributions in several key research questions in the field: strategies for the use of monolingual data, curating seed training data, tokenization and suitable architectures, the competitiveness of small specialized MT models as opposed to general purpose large models, and evaluation protocols. Our data strategy combines large-scale multilingual corpora with backtranslation for 2,000+ languages and MeDLEy, a grammatically-diverse seed dataset for 109 low-resource languages. We explore two complementary architectures: OMT-LLAMA, a decoder-only model (1B/3B/8B) built on LLAMA3 with multilingual continual pretraining, and OMT-NLLB, a 3B encoder-decoder model with a novel three-stage training strategy exploiting non-parallel data via autoencoding. For evaluation, we extend the BOUQuET benchmark to 275 languages, develop Met-BOUQuET with human annotations of machine quality in 161 directions, and BLASER 3, a reference-free quality estimation metric. Our specialized 1B–8B models match or exceed 70B general-purpose LLMs, demonstrating that specialization is a more reliable path to multilingual translation than scale alone. Omnilingual MT achieves “passable” translation quality as judged by human annotators for over 440 languages, doubling it over the previous state of the field at ∼200 languages, and translates in zero-shot mode into ∼1,200 languages versus 300–400 for the baselines. With Omnilingual MT, we showcase a path encompassing the full pipeline of data and system development towards designing models that are multilingual from the onset.

Existing benchmarks for large language models (LLMs) are largely restricted to high- or mid-resource languages, and often evaluate performance on higher-order tasks in reasoning and generation. However, plenty of evidence points to the fact that LLMs lack basic linguistic competence in the vast majority of the world’s 3800+ written languages. We introduce ChiKhaPo, consisting of eight subtasks of varying difficulty designed to evaluate the lexical comprehension and generation abilities of generative models. ChiKhaPo draws on existing lexicons, monolingual data, and bitext, and provides coverage for 2700+ languages for two word-translation-based subtasks, surpassing any existing benchmark in terms of language coverage. We further show that six SOTA models struggle on our benchmark, and discuss the factors contributing to performance scores, including language family, language resourcedness, task, and comprehension versus generation directions. With ChiKhaPo, we hope to enable and encourage the massively multilingual benchmarking of LLMs.

We present the Johns Hopkins University’s submission to the 2025 IWSLT Low-Resource Task. We competed on all 10 language pairs. Our approach centers around ensembling methods – specifically Minimum Bayes Risk Decoding. We find that such ensembling often improves performance only slightly over the best performing stand-alone model, and that in some cases it can even hurt performance slightly.

Multilingual generation with large language models (LLMs) is often of poor quality for mid- to low-resource languages, but the causes for this are not well-understood. We first demonstrate the existence of an implicit task-solving→translation pipeline for generation, whereby the model first solves the required task in a largely target-language-agnostic manner, and subsequently translates answer concepts into the intended target language. We hypothesize that the failure of the translation stage, despite task-solving success, is an important culprit for the observed low quality of final outputs, and formalize this as the translation barrier hypothesis. We quantify the extent to which either stage in the pipeline is responsible for final failure for a word translation task across 108 language pairs, and find that the translation barrier explains a dominant portion of error for a majority of language pairs, and is especially severe for low-resource target languages. Our results highlight an important bottleneck for end-to-end multilingual generation, relevant for future work seeking to improve multilinguality in LLMs.

Most of the world’s languages and dialects are low-resource, and lack support in mainstream machine translation (MT) models. However, many of them have a closely-related high-resource language (HRL) neighbor, and differ in linguistically regular ways from it. This underscores the importance of model robustness to dialectal variation and cross-lingual generalization to the HRL dialect continuum. We present DialUp, consisting of a training-time technique for adapting a pretrained model to dialectal data (M–>D), and an inference-time intervention adapting dialectal data to the model expertise (D–>M). M–>D induces model robustness to potentially unseen and unknown dialects by exposure to synthetic data exemplifying linguistic mechanisms of dialectal variation, whereas D–>M treats dialectal divergence for known target dialects. These methods show considerable performance gains for several dialects from four language families, and modest gains for two other language families. We also conduct feature and error analyses, which show that language varieties with low baseline MT performance are more likely to benefit from these approaches.

Most existing approaches for unsupervised bilingual lexicon induction (BLI) depend on good quality static or contextual embeddings requiring large monolingual corpora for both languages. However, unsupervised BLI is most likely to be useful for low-resource languages (LRLs), where large datasets are not available. Often we are interested in building bilingual resources for LRLs against related high-resource languages (HRLs), resulting in severely imbalanced data settings for BLI. We first show that state-of-the-art BLI methods in the literature exhibit near-zero performance for severely data-imbalanced language pairs, indicating that these settings require more robust techniques. We then present a new method for unsupervised BLI between a related LRL and HRL that only requires inference on a masked language model of the HRL, and demonstrate its effectiveness on truly low-resource languages Bhojpuri and Magahi (with <5M monolingual tokens each), against Hindi. We further present experiments on (mid-resource) Marathi and Nepali to compare approach performances by resource range, and release our resulting lexicons for five low-resource Indic languages: Bhojpuri, Magahi, Awadhi, Braj, and Maithili, against Hindi.

Johns Hopkins University (JHU) submitted systems for all eight language pairs in the 2024 Low-Resource Language Track. The main effort of this work revolves around fine-tuning large and publicly available models in three proposed systems: i) end-to-end speech translation (ST) fine-tuning of Seamless4MT v2; ii) ST fine-tuning of Whisper; iii) a cascaded system involving automatic speech recognition with fine-tuned Whisper and machine translation with NLLB. On top of systems above, we conduct a comparative analysis on different training paradigms, such as intra-distillation for NLLB as well as joint training and curriculum learning for SeamlessM4T v2. Our results show that the best-performing approach differs by language pairs, but that i) fine-tuned SeamlessM4T v2 tends to perform best for source languages on which it was pre-trained, ii) multi-task training helps Whisper fine-tuning, iii) cascaded systems with Whisper and NLLB tend to outperform Whisper alone, and iv) intra-distillation helps NLLB fine-tuning.

While Transformer-based neural machine translation (NMT) is very effective in high-resource settings, many languages lack the necessary large parallel corpora to benefit from it. In the context of low-resource (LR) MT between two closely-related languages, a natural intuition is to seek benefits from structural “shortcuts”, such as copying subwords from the source to the target, given that such language pairs often share a considerable number of identical words, cognates, and borrowings. We test Pointer-Generator Networks for this purpose for six language pairs over a variety of resource ranges, and find weak improvements for most settings. However, analysis shows that the model does not show greater improvements for closely-related vs. more distant language pairs, or for lower resource ranges, and that the models do not exhibit the expected usage of the mechanism for shared subwords. Our discussion of the reasons for this behaviour highlights several general challenges for LR NMT, such as modern tokenization strategies, noisy real-world conditions, and linguistic complexities. We call for better scrutiny of linguistically motivated improvements to NMT given the blackbox nature of Transformer models, as well as for a focus on the above problems in the field.

While large language models exhibit certain cross-lingual generalization capabilities, they suffer from performance degradation (PD) on unseen closely-related languages (CRLs) and dialects relative to their high-resource language neighbour (HRLN). However, we currently lack a fundamental understanding of what kinds of linguistic distances contribute to PD, and to what extent. Furthermore, studies of cross-lingual generalization are confounded by unknown quantities of CRL language traces in the training data, and by the frequent lack of availability of evaluation data in lower-resource related languages and dialects. To address these issues, we model phonological, morphological, and lexical distance as Bayesian noise processes to synthesize artificial languages that are controllably distant from the HRLN. We analyse PD as a function of underlying noise parameters, offering insights on model robustness to isolated and composed linguistic phenomena, and the impact of task and HRL characteristics on PD. We calculate parameter posteriors on real CRL-HRLN pair data and show that they follow computed trends of artificial languages, demonstrating the viability of our noisers. Our framework offers a cheap solution for estimating task performance on an unseen CRL given HRLN performance using its posteriors, as well as for diagnosing observed PD on a CRL in terms of its linguistic distances from its HRLN, and opens doors to principled methods of mitigating performance degradation.

Neural language models play an increasingly central role for language processing, given their success for a range of NLP tasks. In this study, we compare some canonical strategies in language modeling for low-resource scenarios, evaluating all models by their (finetuned) performance on a POS-tagging downstream task. We work with five (extremely) low-resource dialects from the Indic dialect continuum (Braj, Awadhi, Bhojpuri, Magahi, Maithili), which are closely related to each other and the standard mid-resource dialect, Hindi. The strategies we evaluate broadly include from-scratch pretraining, and cross-lingual transfer between the dialects as well as from different kinds of off-the- shelf multilingual models; we find that a model pretrained on other mid-resource Indic dialects and languages, with extended pretraining on target dialect data, consistently outperforms other models. We interpret our results in terms of dataset sizes, phylogenetic relationships, and corpus statistics, as well as particularities of this linguistic system.

Word embeddings are growing to be a crucial resource in the field of NLP for any language. This work introduces a novel technique for static subword embeddings transfer for Indic languages from a relatively higher resource language to a genealogically related low resource language. We primarily work with HindiMarathi, simulating a low-resource scenario for Marathi, and confirm observed trends on Nepali. We demonstrate the consistent benefits of unsupervised morphemic segmentation on both source and target sides over the treatment performed by fastText. Our best-performing approach uses an EM-style approach to learning bilingual subword embeddings; we also show, for the first time, that a trivial “copyand-paste” embeddings transfer based on even perfect bilingual lexicons is inadequate in capturing language-specific relationships. We find that our approach substantially outperforms the fastText baselines for both Marathi and Nepali on the Word Similarity task as well as WordNetBased Synonymy Tests; on the former task, its performance for Marathi is close to that of pretrained fastText embeddings that use three orders of magnitude more Marathi data.

Our work aims at developing a multilingual data resource for morphological segmentation. We present a survey of 17 existing data resources relevant for segmentation in 32 languages, and analyze diversity of how individual linguistic phenomena are captured across them. Inspired by the success of Universal Dependencies, we propose a harmonized scheme for segmentation representation, and convert the data from the studied resources into this common scheme. Harmonized versions of resources available under free licenses are published as a collection called UniSegments 1.0.

We present a novel method for unsupervised cognate/borrowing identification from monolingual corpora designed for low and extremely low resource scenarios, based on combining noisy semantic signals from joint bilingual spaces with orthographic cues modelling sound change. We apply our method to the North Indian dialect continuum, containing several dozens of dialects and languages spoken by more than 100 million people. Many of these languages are zero-resource and therefore natural language processing for them is non-existent. We first collect monolingual data for 26 Indic languages, 16 of which were previously zero-resource, and perform exploratory character, lexical and subword cross-lingual alignment experiments for the first time at this scale on this dialect continuum. We create bilingual evaluation lexicons against Hindi for 20 of the languages. We then apply our cognate identification method on the data, and show that our method outperforms both traditional orthography baselines as well as EM-style learnt edit distance matrices. To the best of our knowledge, this is the first work to combine traditional orthographic cues with noisy bilingual embeddings to tackle unsupervised cognate detection in a (truly) low-resource setup, showing that even noisy bilingual embeddings can act as good guides for this task. We release our multilingual dialect corpus, called HinDialect, as well as our scripts for evaluation data collection and cognate induction.

Technical terms may require special handling when the target audience is bilingual, depending on the cultural and educational norms of the society in question. In particular, certain translation scenarios may require “term retention” i.e. preserving of the source language technical terms in the target language output to produce a fluent and comprehensible code-switched sentence. We show that a standard transformer-based machine translation model can be adapted easily to perform this task with little or no damage to the general quality of its output. We present an English-to-Hindi model that is trained to obey a “retain” signal, i.e. it can perform the required code-mixing on a list of terms, possibly unseen, provided at runtime. We perform automatic evaluation using BLEU as well as F1 metrics on the list of retained terms; we also collect manual judgments on the quality of the output sentences.

Verbal prediction has been shown to be critical during online comprehension of Subject-Object-Verb (SOV) languages. In this work we present three computational models to predict clause final verbs in Hindi given its prior arguments. The models differ in their use of prior context during the prediction process – the context is either noisy or noise-free. Model predictions are compared with the sentence completion data obtained from Hindi native speakers. Results show that models that assume noisy context outperform the noise-free model. In particular, a lossy context model that assumes prior context to be affected by predictability and recency captures the distribution of the predicted verb class and error sources best. The success of the predictability-recency lossy context model is consistent with the noisy channel hypothesis for sentence comprehension and supports the idea that the reconstruction of the context during prediction is driven by prior linguistic exposure. These results also shed light on the nature of the noise that affects the reconstruction process. Overall the results pose a challenge to the adaptability hypothesis that assumes use of noise-free preverbal context for robust verbal prediction.

English-Hindi machine translation systems have difficulty interpreting verb phrase ellipsis (VPE) in English, and commit errors in translating sentences with VPE. We present a solution and theoretical backing for the treatment of English VPE, with the specific scope of enabling English-Hindi MT, based on an understanding of the syntactical phenomenon of verb-stranding verb phrase ellipsis in Hindi (VVPE). We implement a rule-based system to perform the following sub-tasks: 1) Verb ellipsis identification in the English source sentence, 2) Elided verb phrase head identification 3) Identification of verb segment which needs to be induced at the site of ellipsis 4) Modify input sentence; i.e. resolving VPE and inducing the required verb segment. This system obtains 94.83 percent precision and 83.04 percent recall on subtask (1), tested on 3900 sentences from the BNC corpus. This is competitive with state-of-the-art results. We measure accuracy of subtasks (2) and (3) together, and obtain a 91 percent accuracy on 200 sentences taken from the WSJ corpus. Finally, in order to indicate the relevance of ellipsis handling to MT, we carried out a manual analysis of the English-Hindi MT outputs of 100 sentences after passing it through our system. We set up a basic metric (1-5) for this evaluation, where 5 indicates drastic improvement, and obtained an average of 3.55. As far as we know, this is the first attempt to target ellipsis resolution in the context of improving English-Hindi machine translation.