We present the Findings of the WMT 2026 Shared Task on Multitask LLMs with Limited Resources. This year, we extended the Shared Task to jointly model five diverse NLP tasks, Machine Translation (MT) for five language pairs (eight directions), Question Answering (QA), Spell Checking (SC), Grammar Checking (GC), and Mathematical Reasoning (MR), for two language tracks: Ukrainian and Sorbian (grouping Upper and Lower Sorbian). As the Shared Task is focused on low-resource conditions, the model is restricted to Qwen3.5-2B to remain below the 3B-parameter threshold.

In total, six teams participated across the two language tracks: two in the Ukrainian and four in the Sorbian track. No team participated in both language tracks. All systems have been uploaded to HuggingFace by the participants. Despite our strict restrictions on model and dataset availability, systems were diverse in the resources they used as well as their training and inference strategies. The submitted systems show that a single small model can handle all five tasks, with clear difficulty shown for the challenging MR only.

This work examines recent byte-level models and their failure to perfectly manipulate characters. State-of-the-art byte-level models use a hierarchical structure, starting at the byte level, downsampling to the word level, and then upsampling back to bytes. While this improves training and inference efficiency, we find that the hierarchical design itself limits character-level understanding, with pure byte-level models consistently outperforming hierarchical variants on character manipulation tasks. Ablating transformer layers into attention and feed-forward components further reveals that byte-level attention is the primary mechanism driving this behavior. Together, our results provide an explanation for the character-level failures of hierarchical byte models and establish a clear trade-off between computational efficiency and fine-grained character understanding.

Large language models (LLMs) demonstrate impressive performance across a wide range of general NLP tasks; however, their effectiveness in sensitive domains, such as hate speech detection, remains less clear. Prior studies comparing prompted LLMs with state-of-the-art encoder-based models (e.g., BERT variants (Roy et al., 2023; Dönmez et al., 2024)) have shown only marginal gains, suggesting that LLMs may not excel in hate speech detection or mitigation. In this work, we revisit this question through the lens of instruction tuning. By thoroughly unifying 36 English hate speech datasets spanning multiple labeling schemes, we fine-tune a generalist LLM, based on Qwen3 (Qwen Team, 2025), specifically for hate speech mitigation. Our results demonstrate not only state-of-the-art performance on in-domain benchmarks but also substantial improvements in cross-domain and cross-lingual generalization—areas where encoder-based specialist classifiers often struggle.

We investigate the effects of multilingual pre-training with limited data by addressing the problem of how language exposure should be distributed and ordered across English, Dutch, and Chinese. In particular, we compare balanced multilingual exposure to language-dominant exposure, where one language is exposed with more training data than the others. We also compare mixed training, where multilingual training data is used throughout training, with curriculum training, where monolingual exposure of the dominant language is followed by balanced multilingual exposure. Our results show that multilingual BabyLM performance varies with both the amount and order of language exposure. Mixed training outperforms curriculum learning across all 15 matched configurations, indicating that simultaneous multilingual exposure is more effective than staged exposure in our setting. For each primary language, the best language-dominant mixed training configuration outperforms the balanced configuration, suggesting that an equal distribution is not necessarily optimal under a limited data budget. Interestingly, the Dutch-primary configurations achieve the best overall performance.

The BabyLM shared task asks whether language models can acquire robust linguistic behavior from developmentally sized corpora. We study a causal transformer recipe for this setting that increases computation per token without adding a full stack of independent layers. Our RecGPT models reuse one transformer block recurrently, factorize the vocabulary maps to keep small models from being dominated by embeddings, and add a training-only next-latent auxiliary loss that encourages hidden states to support local transitions. We submit RecGPT-10M for Strict-Small and RecGPT-100M for Strict. Across the BabyLM 2026 zero-shot and finetuned evaluations, the submitted models improve over GPT-2 baselines and are competitive with GPT-BERT.

This work concerns our submission for the Multilingual track of the 2026 BabyLM Challenge. We improve on our baseline using data that has been summarized and translated by an LLM, showing that higher-quality data can come from synthetic sources. We further improve with 2 simple changes: using a larger model, and using a character-level vocabulary. Both of these changes require LoRA-style fine-tuning for realizing their full potential. We dub our final Qwen3-based model Qwenfant, which, at the time of submission, ranks 1st in the Multilingual track.

We investigate whether augmenting training data with LLM-generated rewrites can improve performance in small-data language model pretraining. We pair each training sentence with a meaning-preserving rewrite, so that masked tokens can be predicted using both the original and its paired version in the same training window. On FineWeb, we test three rewrite types: simplification (shorter and easier), explanation (longer and easier), and paraphrasing (similar length but harder to read). We also test simplification on OpenStax+BCCampus and Simple Wikipedia. We add a reverse masking curriculum: 7 epochs of whole-word masking (WWM), where entire words are masked instead of individual subword tokens, followed by 3 epochs of standard token-level masking (WWM7-to-Token3). The masking procedure is not rewrite-aware: it does not use alignments between the original sentence and its rewrite. All three FineWeb rewrite types improve EWoK and GLUE-style fine-tuning over source-matched baselines, and simplification improves Entity Tracking across all three sources. The EWoK and GLUE gains concentrate on concept-swap and sentence-pair tasks, suggesting that rewrite pairing especially helps models link meaning across different surface forms. Despite using only one-tenth as much training data, our best model surpasses the official BabyLM 100M GPT-2 baseline on EWoK, Entity Tracking, and GLUE, suggesting that paired rewrite augmentation can improve sample efficiency under a fixed small-data budget.

Training multilingual language models under strict data constraints remains a challenging problem. This paper presents our submission to the BabyLM 2026 Multilingual Track, focusing on English, Dutch, and Chinese. Three key challenges arise under the BabyLM constraints: efficiently sharing limited model capacity, maximizing sequence utilization, and allocating parameters efficiently under a fixed vocabulary budget. To address these challenges, we propose a three-stage curriculum learning strategy that dynamically adjusts language mixing ratios throughout training. Furthermore, we investigate data packing strategies, comparing naive chunking with document wrapping to maximize the learning signal within a strict 100M-token budget. Finally, we design a custom BPE tokenizer with an asymmetric sampling strategy. By reducing the size of the embedding matrix, the saved parameters are reallocated to the Transformer backbone, enabling highly efficient GPT-2 and BERT models. Experimental results show that combining our curriculum strategy with document packing consistently improves GPT-2 training convergence and achieves highly competitive multilingual performance.

We present the findings of the WMT 2025 Shared Task LLMs with Limited Resources for Slavic Languages. This shared task focuses on training LLMs using limited data and compute resources for three Slavic languages: Upper Sorbian (hsb), Lower Sorbian (dsb), and Ukrainian (uk), with the objective to develop and improve LLMs for these languages. We consider two tasks which are to be evaluated jointly: Machine Translation (MT) and Multiple-Choice Question Answering (QA). In total, three teams participated in this shared task, with submissions from all three teams for the Sorbian languages and one submission for Ukrainian. All submissions led to an improvement compared to the baseline Qwen2.5-3B model through varying fine-tuning strategies. We note, however, that training purely on MT degrades original QA capabilities. We also report further analyses on the submissions, including MT evaluation using advanced neural metrics for Ukrainian, as well as manual annotation and comparison to the current Sorbian machine translator.

The CUTE benchmark showed that LLMs struggle with character understanding in English. We extend it to more languages with diverse scripts and writing systems, introducing EXECUTE. Our simplified framework allows easy expansion to any language. Tests across multiple LLMs reveal that challenges in other languages are not always on the character level as in English. Some languages show word-level processing issues, some show no issues at all. We also examine sub-character tasks in Chinese, Japanese, and Korean to assess LLMs’ understanding of character components.

We describe our strategy for the 2025 edition of the BabyLM Challenge. Our main contribution is that of an improved form of Masked Language Modeling (MLM), which adapts the probabilities of the tokens masked according to the model’s ability to predict them. The results show a substantial increase in performance on (Super)GLUE tasks over the standard MLM. We also incorporate sub-token embeddings, finding that this increases the model’s morphological generalization capabilities. Our submission beats the baseline in the strict-small track.

Large Language Models (LLMs) typically track the order of tokens using positional encoding, which causes the following problems: positional bias, where the model is influenced by an ordering within the prompt, and a fixed context window, as models struggle to generalize to positions beyond those encountered during training. To address these limitations, we developed a novel method called set encoding. This method allows multiple pieces of text to be encoded in the same position, thereby eliminating positional bias entirely. Another promising use case for set encoding is to increase the size of the input an LLM can handle. Our experiments demonstrate that set encoding allows an LLM to solve tasks with far more tokens than without set encoding. To our knowledge, set encoding is the first technique to effectively extend an LLM’s context window without requiring any additional training.

Pretrained character-level and byte-level language models have been shown to be competitive with popular subword models across a range of Natural Language Processing tasks. However, there has been little research on their effectiveness for neural machine translation (NMT), particularly within the popular pretrain-then-finetune paradigm. This work performs an extensive comparison across multiple languages and experimental conditions of character- and subword-level pretrained models (ByT5 and mT5, respectively) on NMT. We show the effectiveness of character-level modeling in translation, particularly in cases where fine-tuning data is limited. In our analysis, we show how character models’ gains in translation quality are reflected in better translations of orthographically similar words and rare words. While evaluating the importance of source texts in driving model predictions, we highlight word-level patterns within ByT5, suggesting an ability to modulate word-level and character-level information during generation. We conclude by assessing the efficiency tradeoff of byte models, suggesting their usage in non-time-critical scenarios to boost translation quality.

Large Language Models (LLMs) show remarkable performance on a wide variety of tasks. Most LLMs split text into multi-character tokens and process them as atomic units without direct access to individual characters. This raises the question: To what extent can LLMs learn orthographic information? To answer this, we propose a new benchmark, CUTE, which features a collection of tasks designed to test the orthographic knowledge of LLMs. We evaluate popular LLMs on CUTE, finding that most of them seem to know the spelling of their tokens, yet fail to use this information effectively to manipulate text, calling into question how much of this knowledge is generalizable.

This paper describes a linguistically-motivated approach to the 2024 edition of the BabyLM Challenge. Rather than pursuing a first language learning (L1) paradigm, we approach the challenge from a second language (L2) learning perspective. In L2 learning, there is a stronger focus on learning explicit linguistic information, such as grammatical notions, definitions of words or different ways of expressing a meaning. This makes L2 learning potentially more efficient and concise. We approximate this using data from Wiktionary, grammar examples either generated by an LLM or sourced from grammar books, and paraphrase data.We find that explicit information about word meaning (in our case, Wiktionary) does not boost model performance, while grammatical information can give a small improvement. The most impactful data ingredient is sentence paraphrases, with our two best models being trained on 1) a mix of paraphrase data and data from the BabyLM pretraining dataset, and 2) exclusively paraphrase data.

Misogyny and sexism are growing problems in social media. Advances have been made in online sexism detection but the systems are often uninterpretable. SemEval-2023 Task 10 on Explainable Detection of Online Sexism aims at increasing explainability of the sexism detection, and our team participated in all the proposed subtasks. Our system is based on further domain-adaptive pre-training. Building on the Transformer-based models with the domain adaptation, we compare fine-tuning with multi-task learning and show that each subtask requires a different system configuration. In our experiments, multi-task learning performs on par with standard fine-tuning for sexism detection and noticeably better for coarse-grained sexism classification, while fine-tuning is preferable for fine-grained classification.

This paper describes our system created for the SemEval 2022 Task 3: Presupposed Taxonomies - Evaluating Neural-network Semantics. This task is focused on correctly recognizing taxonomic word relations in English, French and Italian. We developed various datageneration techniques that expand the originally provided train set and show that all methods increase the performance of modelstrained on these expanded datasets. Our final system outperformed the baseline system from the task organizers by achieving an average macro F1 score of 79.6 on all languages, compared to the baseline’s 67.4.

Subword-level models have been the dominant paradigm in NLP. However, character-level models have the benefit of seeing each character individually, providing the model with more detailed information that ultimately could lead to better models. Recent works have shown character-level models to be competitive with subword models, but costly in terms of time and computation. Character-level models with a downsampling component alleviate this, but at the cost of quality, particularly for machine translation. This work analyzes the problems of previous downsampling methods and introduces a novel downsampling method which is informed by subwords.This new downsampling method not only outperforms existing downsampling methods, showing that downsampling characters can be done without sacrificing quality, but also leads to promising performance compared to subword models for translation.

This paper describes the methods behind the systems submitted by the University of Groningen for the WMT 2021 Unsupervised Machine Translation task for German–Lower Sorbian (DE–DSB): a high-resource language to a low-resource one. Our system uses a transformer encoder-decoder architecture in which we make three changes to the standard training procedure. First, our training focuses on two languages at a time, contrasting with a wealth of research on multilingual systems. Second, we introduce a novel method for initializing the vocabulary of an unseen language, achieving improvements of 3.2 BLEU for DE->DSB and 4.0 BLEU for DSB->DE.Lastly, we experiment with the order in which offline and online back-translation are used to train an unsupervised system, finding that using online back-translation first works better for DE->DSB by 2.76 BLEU. Our submissions ranked first (tied with another team) for DSB->DE and third for DE->DSB.

This paper investigates very low resource language model pretraining, when less than 100 thousand sentences are available. We find that, in very low-resource scenarios, statistical n-gram language models outperform state-of-the-art neural models. Our experiments show that this is mainly due to the focus of the former on a local context. As such, we introduce three methods to improve a neural model’s performance in the low-resource setting, finding that limiting the model’s self-attention is the most effective one, improving on downstream tasks such as NLI and POS tagging by up to 5% for the languages we test on: English, Hindi, and Turkish.

Translating to and from low-resource polysynthetic languages present numerous challenges for NMT. We present the results of our systems for the English–Inuktitut language pair for the WMT 2020 translation tasks. We investigated the importance of correct morphological segmentation, whether or not adding data from a related language (Greenlandic) helps, and whether using contextual word embeddings improves translation. While each method showed some promise, the results are mixed.

This paper describes the methods behind the systems submitted by the University of Groningen for the WMT 2020 Unsupervised Machine Translation task for German–Upper Sorbian. We investigate the usefulness of data selection in the unsupervised setting. We find that we can perform data selection using a pretrained model and show that the quality of a set of sentences or documents can have a great impact on the performance of the UNMT system trained on it. Furthermore, we show that document-level data selection should be preferred for training the XLM model when possible. Finally, we show that there is a trade-off between quality and quantity of the data used to train UNMT systems.

Unsupervised Machine Translation has been advancing our ability to translate without parallel data, but state-of-the-art methods assume an abundance of monolingual data. This paper investigates the scenario where monolingual data is limited as well, finding that current unsupervised methods suffer in performance under this stricter setting. We find that the performance loss originates from the poor quality of the pretrained monolingual embeddings, and we offer a potential solution: dependency-based word embeddings. These embeddings result in a complementary word representation which offers a boost in performance of around 1.5 BLEU points compared to standard word2vec when monolingual data is limited to 1 million sentences per language. We also find that the inclusion of sub-word information is crucial to improving the quality of the embeddings.

This paper presents the systems submitted by the University of Groningen to the English– Kazakh language pair (both translation directions) for the WMT 2019 news translation task. We explore the potential benefits of (i) morphological segmentation (both unsupervised and rule-based), given the agglutinative nature of Kazakh, (ii) data from two additional languages (Turkish and Russian), given the scarcity of English–Kazakh data and (iii) synthetic data, both for the source and for the target language. Our best submissions ranked second for Kazakh→English and third for English→Kazakh in terms of the BLEU automatic evaluation metric.