We describe LT3’s submission to the WMT26 shared task on Multitask LLMs with Limited Resources for Upper Sorbian (hsb) and Lower Sorbian (dsb): one QLoRA-fine-tuned Qwen3.5-2B model serves machine translation, multiple-choice question answering, spell checking, grammar checking, and maths reasoning. Our approach centres on three components. First, retrieval augmentation: fuzzy-match (FM) exemplars, retrieved by character- and embedding-based similarity from authentic and back-translated sentence pools, are included in prompts at training and inference times. Second, synthetic task data: training data for spell- and grammar-checking is generated from dictionaries and monolingual text. Third, a scoring-based inference system: the model generates free text only for the open-ended tasks–translation and maths reasoning–while, for the other three tasks, it scores candidates from constrained spaces by ranking multiple-choice options in place and selecting spelling and grammar corrections from dictionary-derived candidates. Our primary system was the joint winner of the Sorbian track, ranking first or second across all five tasks and outperforming the next-best system by 7.7 accuracy points on question answering and 1.4 points on mathematical reasoning. These results show that, for low-resource languages, FM augmentation combined with back-translation can improve MT performance for a relatively small LLM, while inference-level strategies can further improve its performance on other tasks without compromising MT performance.

Before using GenAI models as EdTech tools, their pedagogical suitability should be corroborated. In this paper, we present ShAnEL-2, a novel multilingual dataset comprising 1,185 student responses to short-answer language learning exercises corrected by teachers. We use ShAnEL-2 to establish an initial benchmark of (1) “off-the-shelf” GenAI models and (2) retrieval-augmented generation (RAG) techniques for the automated correction of this exercise type. With an overall accuracy of 90% and recall of 95%, few-shot RAG (which adds previously corrected responses to the prompt) outperforms the off-the-shelf baseline and textbook RAG setup (which adds coursebook materials) by up to 7 (accuracy) and 5 (recall) percentage points. These results confirm that LLMs learn better from examples than from analysing context and highlight GenAI’s particular potential as a correction assistant for teachers.

Recent work on gender bias in the automatic translation of gender-ambiguous scenarios has highlighted the need for gender ambiguity detection. We address this gap with ContraGAND, a contrastive corpus of 4,852 natural gender-ambiguous English sentences, each minimally paired with a masculine and a feminine variant generated by a neurosymbolic pipeline. We use ContraGAND to perform and evaluate gender ambiguity detection in two settings: (i) a closed task, where the referent is given, and (ii) an open task, where it is not. For both tasks, we evaluate six open-weight large language models by prompting, and test selected models with fine-tuning and knowledge distillation. We show that (i) the closed task is saturated in current models, whereas (ii) the open task produces a gap that varies considerably by model. We find that two failure modes dominate: a binding failure (the model surfaces the referent but defaults to ’ambiguous’ when the gender cue is cross-clause) and a format failure that, on local cues, is primarily under-commitment rather than a binding deficit (long unparseable reasoning chains on small models). Further analysis reveals that both are forms of over-prediction of the ’ambiguous’ class rather than stereotype-driven misclassification. While explicit prompts and decoding interventions do not close the gap, both fine-tuning on the augmented contrastive variants and distillation from a 31B teacher are effective at teaching a 4B student to successfully perform gender ambiguity detection.

Effective communication in asylum reception settings requires reliable machine translation (MT) across many languages, including low-resource ones. Using data from the MaTIAS project, we compare retrieval-augmented LLM translation with adaptive Neural MT across 14 target languages with varying resource levels. Working with a very small translation memory of only 358 sentences, we evaluate fuzzy match (FM) augmentation as an in-context learning strategy for open-source and commercial LLMs and benchmark these against ModernMT with and without domain adaptation. In the LLM setting, FM-based example selection consistently outperforms random selection and zero-shot prompting, with the largest gains for low-resource languages. Adaptive NMT retains an overall advantage, although Gemini Pro approaches its performance and outperforms it on 6 of 14 languages, highlighting a trade-off between translation quality and data sovereignty in privacy-sensitive contexts. These findings show that FM augmentation remains effective under severe data constraints and emphasise the importance of language-specific evaluation in multilingual MT.

To enhance the accessibility of scientific literature in multiple languages and facilitate the exchange of information among scholars and a wider audience, there is a need for high-performing specialized machine translation (MT) engines. However, this requires efficient filtering and the use of domain-specific data. In this study, we investigate whether approaches for increasing training data using topic filtering and more efficient use of such data through exploiting fuzzy matches (i.e. similar translations to a given input; FMs) improve translation quality. We apply these techniques both to sequence-to-sequence MT models and off-the-shelf multilingual large language models (LLMs) in three scientific disciplines. Our results suggest that the combination of topic filtering and FM augmentation is an effective strategy for training neural machine translation (NMT) models from scratch, not only surpassing baseline NMT models but also delivering improved translation performance compared to smaller LLMs in terms of the number of parameters. Furthermore, we find that although FM augmentation through in-context learning generally improves LLM translation performance, limited domain-specific datasets can yield results comparable to those achieved with additional multi-domain datasets.