We describe the TUMHN team’s systems for the WMT26 Shared Tasks on LLMs with Lim- ited Resources for Slavic Languages. We partic- ipated in the Ukrainian track, covering machine translation, question answering, spell checking, grammar checking, and math reasoning tasks. We restructured our training dataset into a chat-style format and fine-tuned the Qwen3.5-2B model for the Ukrainian track in a multitask learn- ing setting. Our model outperforms the base- line on three of the five tasks, with significant improvements on Question Answering, Spell Checking, and Grammar Checking. In Machine Translation, the model shows a slight improve- ment for the English–Ukrainian direction. For Spell Checking and Grammar Checking specif- ically, the largest gains were observed in spell or grammatical error detection as opposed to correction. In Math Reasoning, however, our model does not outperform the baseline. The fine-tuned model 1 and training dataset 2 are publicly available on Hugging Face.

This paper describes the TUM-Heilbronn sys- tem submission to the WMT26 Multilingual Instruction Shared Task (MIST). The MIST covers three tasks: context-based question an- swering, open-ended generation, and cross- lingual summarization to evaluate the multi- lingual instruction-following ability of large language models. We fine-tune Qwen3.5-9B- Instruct on 27 diverse languages, across 10 dif- ferent writing scripts. Our model outperformed 6 of the 15 models submitted to this shared task. Moreover, we ranked first in truthfulness (i.e., decline to answer when the context lacks sufficient information) on the context-based question-answering task. Cross-lingual sum- marization remains a challenging task for all submitted models, with the lowest performance observed among the three tasks.

LLMs have shown impressive linguistic competencies in a wide range of NLP tasks, yet morphological generalization remains challenging. Models can benefit from in-context learning, where the choice of few-shot examples significantly influences the outcome. We probe LLMs’ morphological capabilities by assessing how two targeted in-context exemplar selection strategies, Iterative Confidence-Based Exemplar Selection (i-CES) and Similarity-Based Exemplar Selection (SES), influence the model’s ability for inflectional and derivational modeling across multiple languages. Our experiments with LLaMA 3.3-70B-instruct and Qwen2.5-32B-Instruct models indicate that SES is more effective for accessing latent linguistic knowledge, showing a substantial performance advantage over i-CES in both derivational and inflectional tasks. In contrast, i-CES offers no clear advantage over random sampling.

This paper presents a study of the linguistic knowledge and generalization capabilities of Large Language Models (LLMs), focusing ontheir morphosyntactic competence. We design three diagnostic tasks: (i) labeling syntactic information at the sentence level - identifying subjects, objects, and indirect objects; (ii) derivational decomposition at the word level - identifying morpheme boundaries and labeling thedecomposed sequence; and (iii) in-depth study of morphological decomposition in German and Amharic. We evaluate prompting strategies in GPT-4o and LLaMA 3.3-70B to extract different types of linguistic structure for typologically diverse languages. Our results showthat GPT-4o consistently outperforms LLaMA in all tasks; however, both models exhibit limitations and show little evidence of abstract morphological rule learning. Importantly, we show strong evidence that the models fail to learn underlying morphological structures. Therefore,raising important doubts about their ability to generalize.