We describe Grammatist, our submission to the unconstrained track of the WMT 2026 General Machine Translation Task, developed in direct collaboration with community members of two regional languages of Italy, Ladin and Ligurian. Ladin and Ligurian represent challenging scenarios for machine translation (MT), as they have limited resources for conventional data-driven approaches. However, this scarcity of large-scale parallel data does not reflect a lack of linguistic documentation: both languages are extensively documented in dictionaries and grammar books, which contain valuable knowledge but are not readily exploitable by conventional MT approaches. Grammatist addresses this gap by making this information available to a large language model (LLM) at inference time. The framework is both agentic and model-agnostic: it dynamically consults dictionaries and grammar books during translation to retrieve the information most relevant to the input. Any LLM can serve as the underlying base model, enabling Grammatist to benefit from advances in increasingly capable models. For the WMT 2026 task, we apply Grammatist to structured document-level translation from English into Ligurian and Ladin.

We describe LiLa, our submission to the constrained track of the WMT 2026 General Machine Translation Task. Motivated by recent advances in large language models (LLMs) and their increasing availability as open models, we investigate the adaptation of LLMs for Ligurian and Ladin, two low-resource languages included in the shared task. We evaluate several adaptation strategies for Gemma-4, including Instruction Tuning (IT-only), continuous pre-training followed by Instruction Tuning (CPT+IT), and Low-Rank Adaptation (LoRA), and compare them with a dedicated encoder-decoder pipeline based on fine-tuned No-Language Left Behind (NLLB) models. Our experiments show that, under the evaluated conditions, fine-tuned NLLB models achieve higher automatic translation scores than the investigated LLM adaptation strategies. Based on these results, our final submission adopts a fine-tuned NLLB-based translation pipeline. We discuss the main observations from the different approaches explored during development. We release the code and models produced in these experiments through a publicly accessible GitHub repository https://github.com/schtailmuel/wmt26-lila/

We present the results of the WMT 2026 shared task of the Open Language Data Initiative (OLDI). Participants were invited to contribute to the massively multilingual open datasets supported by OLDI, FLORES+ and OLDI Seed, or to create new resources in line with OLDI’s ethos. We accepted ten submissions: five extending FLORES+, and five extending other open and massively-parallel datasets (BOUQuET, OLDI Seed, SMOL, and WMT24++). These contributions advance the coverage and quality of multilingual datasets, especially for under-served language varieties. All contributions are released under permissive open-source licenses.

This paper presents the results of the General Machine Translation Task organized under the 2026 Conference on Machine Translation (WMT). Participants build systems for any of the 23 language pairs spanning four to five domains. This year we brought major changes to the human evaluation: (1) new annotation platform Pearmut for better reproducibility, (2) new human protocol "contrastive Error Span Annotation" for higher quality and side-by-side evaluation, and (3) Dynamic human evaluation that allows assessing all submitted models while evaluating higher-performing models more frequently. Beyond human evaluation, we (4) extended difficulty sampling with a human-driven stage, (5) added two new domains, (6) made the test sets fully document-level without requiring segment-level alignment and relying on HTML or JSON structure, (7) prepared some human references by post-editing open-weight model outputs, and (8) introduced contextual instructions governing formality and structural style. We evaluated 46 systems in total: 30 submitted by participants and 16 consisting of translations from large language models (LLMs) and industry translation providers.

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.

Language identification (LID) is a fundamental step in curating multilingual corpora. However, LID models still perform poorly for many languages, especially on the noisy and heterogeneous web data often used to train multilingual language models. In this paper, we introduce CommonLID, a community-driven, human-annotated LID benchmark for the web domain, covering 109 languages. Many of the included languages have been previously under-served, making CommonLID a key resource for developing more representative high-quality text corpora. We show CommonLID’s value by using it, alongside five other common evaluation sets, to test eight popular LID models. We analyse our results to situate our contribution and to provide an overview of the state of the art. In particular, we highlight that existing evaluations overestimate LID accuracy for many languages in the web domain. We make CommonLID and the code used to create it available under an open, permissive license.

We present the results of the WMT 2025 shared task of the Open Language Data Initiative. Participants were invited to contribute to the massively multilingual open datasets (FLORES+, MT Seed, WMT24++) or create new such resources. We accepted 8 submissions, including 7 extensions or revisions of the existing datasets and one submission with a new parallel training dataset, SMOL.

We introduce the first highly multilingual speech and American Sign Language (ASL) comprehension dataset by extending BELEBELE. Our dataset covers 91 spoken languages at the intersection of BELEBELE and FLEURS, and one sign language (ASL). As a by-product we also extend the Automatic Speech Recognition Benchmark, FLEURS, by 20%. We evaluate 2M-BELEBELE dataset for both 5-shot and zero-shot settings and across languages, the speech comprehension accuracy is ≈ 10% average lower compared to reading comprehension.

BOUQuET is a multi-way, multicentric and multi-register/domain dataset and benchmark, and a broader collaborative initiative. This dataset is handcrafted in 8 non-English languages (i.e. Egyptian Arabic and Modern Standard Arabic, French, German, Hindi, Indonesian, Mandarin Chinese, Russian, and Spanish). Each of these source languages are representative of the most widely spoken ones and therefore they have the potential to serve as pivot languages that will enable more accurate translations. The dataset is multicentric to enforce representation of multilingual language features. In addition, the dataset goes beyond the sentence level, as it is organized in paragraphs of various lengths. Compared with related machine translation datasets, we show that BOUQuET has a broader representation of domains while simplifying the translation task for non-experts. Therefore, BOUQuET is specially suitable for crowd-source extension for which we are launching a call aim-ing at collecting a multi-way parallel corpus covering any written language. The dataset is freely available at https://huggingface.co/datasets/facebook/bouquet.

We present the results of the WMT 2024 shared task of the Open Language Data Initiative. Participants were invited to contribute to the FLORES+ and MT Seed multilingual datasets, two foundational open resources that facilitate the organic expansion of language technology’s reach. We accepted ten submissions covering 16 languages, which extended the range of languages included in the datasets and improved the quality of existing data.

Large multilingual machine translation efforts are driving improved access and performance for under-resourced languages, but often fail to translate culturally specific and local concepts. Additionally, translation from practically relevant input languages may flag behind those that are comparatively over-represented in the training dataset. In this work, we release a new corpus, ZenaMT, containing 7,561 parallel Ligurian-Italian sentences, nearly a fifth of which are also translated in English. This corpus spans five domains: local and international news, Ligurian literature, Genoese Ligurian linguistics concepts, traditional card game rules, and Ligurian geographic expressions. We find that a translation model augmented with ZenaMT improves a baseline by 20%, and by over 25% (BLEU) compared to NLLB-3.3B, which is over 50 times the size. Our results demonstrate the utility of creating data sets for MT that are specifically tailored for the cultural context of Ligurian speakers. We freely release ZenaMT and expect to periodically update the corpus to improve MT performance and domain coverage.

A major impediment to the advancement of sign language translation (SLT) is data scarcity. Much of the sign language data currently available on the web cannot be used for training supervised models due to the lack of aligned captions. Furthermore, scaling SLT using large-scale web-scraped datasets bears privacy risks due to the presence of biometric information, which the responsible development of SLT technologies should account for. In this work, we propose a two-stage framework for privacy-aware SLT at scale that addresses both of these issues. We introduce SSVP-SLT, which leverages self-supervised video pretraining on anonymized and unannotated videos, followed by supervised SLT finetuning on a curated parallel dataset. SSVP-SLT achieves state-of-the-art finetuned and zero-shot gloss-free SLT performance on the How2Sign dataset, outperforming the strongest respective baselines by over 3 BLEU-4. Based on controlled experiments, we further discuss the advantages and limitations of self-supervised pretraining and anonymization via facial obfuscation for SLT.

Mixture-of-experts (MoE) models that employ sparse activation have demonstrated effectiveness in significantly increasing the number of parameters while maintaining low computational requirements per token. However, recent studies have established that MoE models are inherently parameter-inefficient as the improvement in performance diminishes with an increasing number of experts. We hypothesize this parameter inefficiency is a result of all experts having equal capacity, which may not adequately meet the varying complexity requirements of different tokens or tasks. In light of this, we propose Stratified Mixture of Experts (SMoE) models, which feature a stratified structure and can assign dynamic capacity to different tokens. We demonstrate the effectiveness of SMoE on three multilingual machine translation benchmarks, containing 4, 15, and 94 language pairs, respectively. We show that SMoE outperforms multiple state-of-the-art MoE models with the same or fewer parameters.

Machine Translation systems can produce different types of errors, some of which are characterized as critical or catastrophic due to the specific negative impact that they can have on users. In this paper we focus on one type of critical error: added toxicity. We evaluate and analyze added toxicity when translating a large evaluation dataset (HOLISTICBIAS, over 472k sentences, covering 13 demographic axes) from English into 164 languages. An automatic toxicity evaluation shows that added toxicity across languages varies from 0% to 5%. The output languages with the most added toxicity tend to be low-resource ones, and the demographic axes with the most added toxicity include sexual orientation, gender and sex, and ability. We also perform human evaluation on a subset of 8 translation directions, confirming the prevalence of true added toxicity. We use a measurement of the amount of source contribution to the translation, where a low source contribution implies hallucination, to interpret what causes toxicity. Making use of the input attributions allows us to explain toxicity, because the source contributions significantly correlate with toxicity for 84% of languages studied. Given our findings, our recommendations to reduce added toxicity are to curate training data to avoid mistranslations, mitigate hallucination and check unstable translations.

Multilingual machine translation (MMT) benefits from cross-lingual transfer but is a challenging multitask optimization problem. This is partly because there is no clear framework to systematically learn language-specific parameters. Self-supervised learning (SSL) approaches that leverage large quantities of monolingual data (where parallel data is unavailable) have shown promise by improving translation performance as complementary tasks to the MMT task. However, jointly optimizing SSL and MMT tasks is even more challenging. In this work, we first investigate how to utilize intra-distillation to learn more language-specific parameters and then show the importance of these language-specific parameters. Next, we propose a novel but simple SSL task, concurrent denoising, that co-trains with the MMT task by concurrently denoising monolingual data on both the encoder and decoder. Finally, we apply intra-distillation to this co-training approach. Combining these two approaches significantly improves MMT performance, outperforming three state-of-the-art SSL methods by a large margin, e.g., 11.3% and 3.7% improvement on an 8-language and a 15-language benchmark compared with MASS, respectively.

For many languages, machine translation progress is hindered by the lack of reliable training data. Models are trained on whatever pre-existing datasets may be available and then augmented with synthetic data, because it is often not economical to pay for the creation of large-scale datasets. But for the case of low-resource languages, would the creation of a few thousand professionally translated sentence pairs give any benefit? In this paper, we show that it does. We describe a broad data collection effort involving around 6k professionally translated sentence pairs for each of 39 low-resource languages, which we make publicly available. We analyse the gains of models trained on this small but high-quality data, showing that it has significant impact even when larger but lower quality pre-existing corpora are used, or when data is augmented with millions of sentences through backtranslation.

We present the results of the WMT’22 SharedTask on Large-Scale Machine Translation Evaluation for African Languages. The shared taskincluded both a data and a systems track, alongwith additional innovations, such as a focus onAfrican languages and extensive human evaluation of submitted systems. We received 14system submissions from 8 teams, as well as6 data track contributions. We report a largeprogress in the quality of translation for Africanlanguages since the last iteration of this sharedtask: there is an increase of about 7.5 BLEUpoints across 72 language pairs, and the average BLEU scores went from 15.09 to 22.60.

We aim to investigate the performance of current OCR systems on low resource languages and low resource scripts. We introduce and make publicly available a novel benchmark, OCR4MT, consisting of real and synthetic data, enriched with noise, for 60 low-resource languages in low resource scripts. We evaluate state-of-the-art OCR systems on our benchmark and analyse most common errors. We show that OCR monolingual data is a valuable resource that can increase performance of Machine Translation models, when used in backtranslation. We then perform an ablation study to investigate how OCR errors impact Machine Translation performance and determine what is the minimum level of OCR quality needed for the monolingual data to be useful for Machine Translation.

Challenging problems such as open-domain question answering, fact checking, slot filling and entity linking require access to large, external knowledge sources. While some models do well on individual tasks, developing general models is difficult as each task might require computationally expensive indexing of custom knowledge sources, in addition to dedicated infrastructure. To catalyze research on models that condition on specific information in large textual resources, we present a benchmark for knowledge-intensive language tasks (KILT). All tasks in KILT are grounded in the same snapshot of Wikipedia, reducing engineering turnaround through the re-use of components, as well as accelerating research into task-agnostic memory architectures. We test both task-specific and general baselines, evaluating downstream performance in addition to the ability of the models to provide provenance. We find that a shared dense vector index coupled with a seq2seq model is a strong baseline, outperforming more tailor-made approaches for fact checking, open-domain question answering and dialogue, and yielding competitive results on entity linking and slot filling, by generating disambiguated text. KILT data and code are available at https://github.com/facebookresearch/KILT.

Retrieving relevant contexts from a large corpus is a crucial step for tasks such as open-domain question answering and fact checking. Although neural retrieval outperforms traditional methods like tf-idf and BM25, its performance degrades considerably when applied to out-of-domain data. Driven by the question of whether a neural retrieval model can be universal and perform robustly on a wide variety of problems, we propose a multi-task trained model. Our approach not only outperforms previous methods in the few-shot setting, but also rivals specialised neural retrievers, even when in-domain training data is abundant. With the help of our retriever, we improve existing models for downstream tasks and closely match or improve the state of the art on multiple benchmarks.

Recent years have seen a growing interest within the natural language processing (NLP) community in evaluating the ability of semantic models to capture human meaning representation in the brain. Existing research has mainly focused on applying semantic models to decode brain activity patterns associated with the meaning of individual words, and, more recently, this approach has been extended to sentences and larger text fragments. Our work is the first to investigate metaphor processing in the brain in this context. We evaluate a range of semantic models (word embeddings, compositional, and visual models) in their ability to decode brain activity associated with reading of both literal and metaphoric sentences. Our results suggest that compositional models and word embeddings are able to capture differences in the processing of literal and metaphoric sentences, providing support for the idea that the literal meaning is not fully accessible during familiar metaphor comprehension.

The structured representation for semantic parsing in task-oriented assistant systems is geared towards simple understanding of one-turn queries. Due to the limitations of the representation, the session-based properties such as co-reference resolution and context carryover are processed downstream in a pipelined system. In this paper, we propose a semantic representation for such task-oriented conversational systems that can represent concepts such as co-reference and context carryover, enabling comprehensive understanding of queries in a session. We release a new session-based, compositional task-oriented parsing dataset of 20k sessions consisting of 60k utterances. Unlike Dialog State Tracking Challenges, the queries in the dataset have compositional forms. We propose a new family of Seq2Seq models for the session-based parsing above, which also set state-of-the-art in ATIS, SNIPS, TOP and DSTC2. Notably, we improve the best known results on DSTC2 by up to 5 points for slot-carryover.

Recent work shows that distributional semantic models can be used to decode patterns of brain activity associated with individual words and sentence meanings. However, it is yet unclear to what extent such models can be used to study and decode fMRI patterns associated with specific aspects of semantic composition such as the negation function. In this paper, we apply lexical and compositional semantic models to decode fMRI patterns associated with negated and affirmative sentences containing hand-action verbs. Our results show reduced decoding (correlation) of sentences where the verb is in the negated context, as compared to the affirmative one, within brain regions implicated in action-semantic processing. This supports behavioral and brain imaging studies, suggesting that negation involves reduced access to aspects of the affirmative mental representation. The results pave the way for testing alternate semantic models of negation against human semantic processing in the brain.

Latent tree learning models represent sentences by composing their words according to an induced parse tree, all based on a downstream task. These models often outperform baselines which use (externally provided) syntax trees to drive the composition order. This work contributes (a) a new latent tree learning model based on shift-reduce parsing, with competitive downstream performance and non-trivial induced trees, and (b) an analysis of the trees learned by our shift-reduce model and by a chart-based model.