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/

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.

Benchmarks do more than report scores: they provide a shared basis for evaluating methods, comparing systems, and measuring progress. As language technologies are now widely integrated into everyday life, efforts should also be made to include minority languages in the resources that underpin language-technology development. If languages are absent from relevant benchmarks, they remain harder to evaluate, compare, and incorporate into future language technologies. Expanding benchmark coverage is therefore essential for ensuring fair evaluation, increasing visibility, and enabling continued development of language technologies for these communities. To support MT development for Ladin, we contribute a Ladin extension of BOUQuET to the WMT 2026 OLDI shared task, covering the Val Badia and Gherdëina varieties. We present a systematic reference-creation workflow combining human translation, expert revision, cross-variant alignment, and blind ranking to produce high-quality references. The dataset is released under the CC BY-SA 4.0 license.

Large Language Models (LLMs) have demonstrated strong capabilities in multilingual machine translation. However, they underperform for low-resource languages, indicating the need for more explicit instructional guidance. In this work, we introduce Fragment-Shot Prompting, a novel few-shot prompting method that aims to retrieve examples for every word occurring in the sentence to be translated, illustrating their use and meaning in context. We evaluate our method on translation between Italian, Ladin (Val Badia) and Ladin (Gherdëina) and compare its performance with zero-shot prompting, random few-shot prompting, as well as established lexical and semantic retrieval strategies. We conduct these experiments using state-of-the-art LLMs, including GPT-3.5, GPT-4o, o1-mini, LlaMA-3.3, and DeepSeek-R1. Our results demonstrate that LLMs can extract substantial value from limited data when translating from a low- to the high-resource language. However, this does not apply to translations into the low-resource languages, where the prompting method plays a much more important role. In particular, our method consistently delivers the best results and enables significant gains. Even though translation performance into Ladin remains limited with the available resources, our results highlight the importance of syntactic coverage for improving translation accuracy and ariant-specific adaptation in low-resource scenarios.

High-quality datasets are crucial for training effective state of the art machine translation systems. However, due to the data-intensive nature of these systems, they have to be trained on large amounts of text that can easily go beyond the scope of full human inspection. This makes the presence of noise that can degrade overall system performance a frequent and significant issue. While various approaches have been developed to identify and select only the highest-quality training examples, this is undesirable in scenarios where resources are limited. For this reason, we introduce AlignFix, an open-source tool for augmenting data, identifying and correcting errors in parallel corpora. Leveraging word alignments, AlignFix extracts consistent phrase pairs, enabling targeted replacements that can improve the dataset quality. Besides targeted replacements, the tool enables contextual augmentation by duplicating sentences and allowing users to substitute words with alternatives of their choice. The tool maintains and updates the underlying word alignments, thereby avoiding the costly recomputation. AlignFix runs locally in the browser, requires no installation, and ensures that all data remains entirely on the client side. It is released under Apache 2.0 license, encouraging broad adoption, reuse, and further development. A live demo is available at https://ifi-alignfix.uibk.ac.at.

Recent advances in neural machine translation (NMT) have opened new possibilities for developing translation systems also for smaller, so-called low-resource, languages. The rise of large language models (LLMs) has further revolutionized machine translation by enabling more flexible and context-aware generation. However, many challenges remain for low-resource languages, and the availability of high-quality, validated test data is essential to support meaningful development, evaluation, and comparison of translation systems. In this work, we present an extension of the FLORES+ dataset for two Ladin variants, Val Badia and Gherdëina, as a submission to the Open Language Data Initiative Shared Task 2025. To complement existing resources, we additionally release two parallel datasets for Gherdëina–Val Badia and Gherdëina–Italian. We validate these datasets by evaluating state-of-the-art LLMs and NMT systems on this test data, both with and without leveraging the newly released parallel data for fine-tuning and prompting. The results highlight the considerable potential for improving translation quality in Ladin, while also underscoring the need for further research and resource development, for which this contribution provides a basis.

While machine translation has taken great strides in recent years, thanks in large part to transformer language models, machine translation tools are designed primarily for plain text, and thus not equipped to deal with complex markup documents. Not even Large Language Models can reliably handle LaTeX source files, as non-standard structures are not captured by any available training data. Previous attempts to create translation engines for LaTeX either work on compiled documents, rely on document pre-processors which may lose critical semantic elements, or cannot distinguish between text and non-text content. In this paper we present LaTeXMT, a software solution for structure-preserving, source-to-source translation of LaTeX documents. All of the source code to LaTeXMT is provided under the LGPL-3.0 open-source licence and a web version is publicly available.

This paper explores the impact of different back-translation approaches on machine translation for Ladin, specifically the Val Badia variant. Given the limited amount of parallel data available for this language (only 18k Ladin-Italian sentence pairs), we investigate the performance of a multilingual neural machine translation model fine-tuned for Ladin-Italian. In addition to the available authentic data, we synthesise further translations by using three different models: a fine-tuned neural model, a rule-based system developed specifically for this language pair, and a large language model. Our experiments show that all approaches achieve comparable translation quality in this low-resource scenario, yet round-trip translations highlight differences in model performance.