Term injection in machine translation is undergoing a paradigm shift in the era of large language models (LLMs). Although recent shared-task results suggest near-saturation for sentence-level term injection from pre-defined glossaries, it remains unclear whether this also holds in more challenging settings. We address this question with a custom test set for legal translation from Italian into South Tyrolean German, a non-dominant and under-resourced language variety. We cover three terminology challenges: simple term injection, localisation of abbreviated forms, and homonym disambiguation. We focus on Reasoning Models (RMs) leveraging Test-Time Scaling, comparing them with different architectures and contributing a human analysis of reasoning traces. We find that reasoning offers little benefit for simple term insertion, but yields clear gains for semantically complex cases such as homonym disambiguation. However, human evaluation of reasoning traces shows that these gains do not necessarily reflect robust and factually grounded translation-specific reasoning. We further show that without external terminological resources, even state-of-the-art RMs struggle to retrieve correct terminology for a non-dominant variety, while NMT small models remain competitive when trained on in-domain bilingual corpora. Based on these findings, we propose data collection strategies for inducing translation-specific reasoning, frameworks for adapting to and evaluating terminology across many language varieties, and terminology challenges beyond simple term injection.

We investigate the effect of terminology injection for terminology-constrained translation in a low-resource language variety, with a particular focus on off-the-shelf instruction-tuned Large Language Models (LLMs). We compare a total of 9 models: 4 instruction-tuned LLMs from the Tower and EuroLLM suites, which have been specifically trained for translation-related tasks; 2 generic open-weight LLMs (LLaMA-8B and Mistral-7B); 3 Neural Machine Translation (NMT) systems (an adapted version of MarianMT and ModernMT with and without the glossary function). To this end, we release LegISTyr, a manually curated test set of 2,000 Italian sentences from the legal domain, paired with source Italian terms and target terms in the South Tyrolean standard variety of German. We select only real-world sources and design constraints on length, syntactic clarity, and referential coherence to ensure high quality. LegISTyr includes a homonym subset, which challenges systems on the selection of the correct homonym where sense disambiguation is deducible from the context. Results show that while generic LLMs achieve the highest raw term insertion rates (approximately 64%), translation-specialized LLMs deliver superior fluency (∆ COMET up to 0.04), reduce incorrect homonym selection by half, and generate more controllable output. We posit that models trained on translation-related data are better able to focus on source-side information, producing more coherent translations.

English grammar Multiple-Choice Cloze (MCC) exercises are crucial for improving learners’ grammatical proficiency andcomprehension skills. However, creating these exercises is labour-intensive and requires expert knowledge. Effective MCCexercises must be contextually relevant and engaging, incorporating distractors—plausible but incorrect alternatives—tobalance difficulty and maintain learner motivation. Despite the increasing interest in utilizing large language models (LLMs)in education, their application in generating English grammar MCC exercises is still limited. Previous methods typicallyimpose constraints on LLMs, producing grammatically correct yet uncreative results. This paper explores the potentialof LLMs to independently generate diverse and contextually relevant MCC exercises without predefined limitations. Wehypothesize that LLMs can craft self-contained sentences that foster learner’s communicative competence. Our analysisof existing MCC exercise datasets revealed issues of diversity, completeness, and correctness. Furthermore, we addressthe lack of a standardized automatic metric for evaluating the quality of generated exercises. Our contributions includedeveloping an LLM-based solution for generating MCC exercises, curating a comprehensive dataset spanning 19 grammartopics, and proposing an automatic metric validated against human expert evaluations. This work aims to advance theautomatic generation of English grammar MCC exercises, enhancing both their quality and creativity.