Vision-language models (VLMs) have achieved strong performance on tasks such as image captioning, visual question answering, and image-to-text generation. However, they are predominantly trained on English-centric data, which limits their ability to handle culturally grounded visual understanding and leads to failures in interpreting region-specific meanings, symbolic content, and context-dependent visual cues. Existing benchmarks for cultural competence are often template-driven and focused on surface-level recognition, making them insufficient for evaluating deeper linguistic and pragmatic understanding in culturally situated settings. We introduce PoVisLE, a monocultural vision-language benchmark for Polish designed to evaluate culturally grounded multimodal understanding under a grounded evaluation paradigm, where language is interpreted in interaction with visual context. The dataset contains 1,117 images and 2,366 manually annotated VQA pairs. Overall, our dataset provides a controlled and challenging resource for assessing culturally grounded vision-language understanding beyond surface-level recognition.1

Large language models (LLMs) in medicine are mainly evaluated using multiple-choice question answering (MCQA), which can overestimate real clinical ability due to guessing strategies and answer biases. To address these limitations, we introduce an expanded and more challenging benchmark based on Polish medical exams, adding over 15,000 questions, two new domains, and four structural modifications that reduce MCQA-specific artifacts and better test reasoning. We evaluate 21 LLMs and show that evaluation design strongly affects results. Under our harder setup, the best model (Qwen3.5-122B) drops by 28.4 and 31 pp on English and Polish exams, respectively. Despite low evidence of data contamination, standard MCQA scores do not reliably reflect true medical competence. To facilitate further research, we make our benchmark publicly available: https://github.com/NASK-NLP/Polish-Medical-Exams

Comparative analyses of NLP research, such as identifying who is shaping a subfield, how a topic is shifting geographically, where method- ological gaps lie between research communi- ties, require institutional and topical metadata at a density and granularity the ACL Anthol- ogy does not itself provide. Large-scale re- sources offer partial support: OpenAlex pro- vides high-quality institutional metadata where it can match ACL papers (32% of our gold set), but the DOI dependency limits its cov- erage on older Anthology publications; and the S2ORC bulk release contains 52,013 ACL Anthology fulltexts but populates structured author–affiliation annotations on a small fraction of them in its current state. We present ANTHEM (ANTHology with Enriched Meta- data), a complementary resource that narrows this gap for NLP publications. ANTHEM links 117,386 ACL Anthology papers to Research Or- ganization Registry (ROR)-normalized author affiliations and Computer Science Ontology (CSO) topic labels via a multi-parser, LLM- based extraction pipeline that retains evidence spans for every decision, achieving F1hier of 0.760 on a 781-paper decade-stratified gold set, where parent–child ROR identifiers are treated as equivalent (e.g., a department and its parent university). On top of the corpus we release a benchmark of 99 structured comparison ques- tions across five categories, evaluated under a programmatic gold-set construction in which ANTHEM’s own metadata defines the evalu- ation ground truth. All data, code, and infras- tructure are released under CC-BY-SA-NC.

Dialectal machine translation remains challenging due to limited data and strong linguistic variation not captured by standard benchmarks, which often assume standardized and well–edited text. We study Polish–Silesian MT using neural and rule–based systems, evaluating on SiLTT — a new Pol–Szl testset, alongside established BOUQuET and FLORES benchmarks. Results show our rule–based system is consistently strongest on SiLTT and BOUQuET datasets and that TranslateGemma fine–tuned on a curated dataset improves over strong neural baselines but does not surpass the rule–based system in dialectal settings. We release SiLTT and our best neural model to support further research.

As Large Language Models (LLMs) are deployed globally, ensuring their safety and alignment across multiple languages becomes paramount. However, safety behaviors often vary unpredictably between languages, posing significant challenges for consistent and ethical AI. In this work, we systematically investigate the dynamics of multilingual alignment, exploring whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general knowledge capabilities. We introduce RefusEU a novel refusal alignment dataset covering 12 European languages, including a dedicated test set for evaluating current state-of-the-art models. Our controlled Direct Preference Optimization (DPO) experiments provide two key insights: aligning models exclusively in English is insufficient to ensure cross-lingual safety, even for the same harm categories, whereas training on multilingual datasets can improve safety without degrading general performance, as measured by the Global MMLU benchmark.

Most vision-language models (VLMs) are trained on English-centric data, limiting their performance in other languages and cultural contexts. This restricts their usability for non-English-speaking users and hinders the development of multimodal systems that reflect diverse linguistic and cultural realities. In this work, we reproduce and adapt the LLaVA-Next methodology to create a set of Polish VLMs. We rely on a fully automated pipeline for translating and filtering existing multimodal datasets, and complement this with synthetic Polish data for OCR and culturally specific tasks. Despite relying almost entirely on automatic translation and minimal manual intervention, our approach yields strong results: we observe a +9.5 pp improvement over LLaVA-1.6-Vicuna-13B on a Polish-adapted MMBench, along with higher-quality captions in generative evaluations, as measured by human annotators in terms of linguistic correctness. These findings highlight that large-scale automated translation, combined with lightweight filtering, can effectively bootstrap high-quality multimodal models for low-resource languages. Some challenges remain, particularly in cultural coverage and evaluation. To facilitate further research, we release our models, code, and datasets.

As Large Language Models (LLMs) continue to evolve, ensuring their safety across multiple languages has become a critical concern. While LLMs demonstrate impressive capabilities in English, their safety mechanisms may not generalize effectively to other languages, leading to disparities in toxicity detection, bias mitigation, and harm prevention. This systematic review examines the multilingual safety of LLMs by synthesizing findings from recent studies that evaluate their robustness across diverse linguistic and cultural contexts beyond English language. Our review explores the methodologies used to assess multilingual safety, identifies challenges such as dataset availability and evaluation biases. Based on our analysis we highlight gaps in multilingual safety research and provide recommendations for future work. This review aims to contribute to the development of fair and effective safety mechanisms for LLMs across all languages. We provide the extracted data in an interactive Streamlit dashboard, enabling transparent access to the raw data and allowing for continuous updates.

The development of multilingual large language models (LLMs) presents challenges in evaluating their safety across all supported languages. Enhancing safety in one language (e.g., English) may inadvertently introduce vulnerabilities in others. To address this issue, we implement a methodology for the automatic creation of red-teaming datasets for safety evaluation in Polish language. Our approach generates both harmful and non-harmful prompts by sampling different risk categories and attack styles. We test several open-source models, including those trained on Polish data, and evaluate them using metrics such as Attack Success Rate (ASR) and False Reject Rate (FRR). The results reveal clear gaps in safety performance between models and show that better testing across languages is needed.

Pretrained language models are integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

We present a benchmark dataset for evaluating language model safety in Polish, addressing the underrepresentation of medium-resource languages in existing safety assessments. Our dataset includes both original and adversarially perturbed examples. We fine-tune and evaluate multiple models—LlamaGuard-3-8B, a HerBERT-based classifier, and PLLuM—and find that the HerBERT-based model outperforms others, especially under adversarial conditions.

Neural Machine Translation (NMT) has improved translation by using Transformer-based models, but it still struggles with word ambiguity and context. This problem is especially important in domain-specific applications, which often have problems with unclear sentences or poor data quality. Our research explores how adding information to models can improve translations in the context of e-commerce data. To this end we create ConECT– a new Czech-to-Polish e-commerce product translation dataset coupled with images and product metadata consisting of 11,400 sentence pairs. We then investigate and compare different methods that are applicable to context-aware translation. We test a vision-language model (VLM), finding that visual context aids translation quality. Additionally, we explore the incorporation of contextual information into text-to-text models, such as the product’s category path or image descriptions. The results of our study demonstrate that the incorporation of contextual information leads to an improvement in the quality of machine translation. We make the new dataset publicly available.

Taking note of the current challenges of the peer review system, this paper inventories the research tasks for analysing and possibly automating parts of the reviewing process, like matching submissions with a reviewer’s domain of expertise. For each of these tasks we list their associated datasets, analysing their quality in terms of available documentation of creation and use. Building up on this, we give a set of recommendations to take into account when collecting and releasing data.

Model interpretability in toxicity detection greatly profits from token-level annotations. However, currently, such annotations are only available in English. We introduce a dataset annotated for offensive language detection sourced from a news forum, notable for its incorporation of the Austrian German dialect, comprising 4,562 user comments. In addition to binary offensiveness classification, we identify spans within each comment constituting vulgar language or representing targets of offensive statements. We evaluate fine-tuned Transformer models as well as large language models in a zero- and few-shot fashion. The results indicate that while fine-tuned models excel in detecting linguistic peculiarities such as vulgar dialect, large language models demonstrate superior performance in detecting offensiveness in AustroTox.

This paper describes the HEVS-TUW team submission to the SemEval-2023 Task 8: Causal Claims. We participated in two subtasks: (1) causal claims detection and (2) PIO identification. For subtask 1, we experimented with an ensemble of weakly supervised question detection and fine-tuned Transformer-based models. For subtask 2 of PIO frame extraction, we used a combination of deep representation learning and a rule-based approach. Our best model for subtask 1 ranks fourth with an F1-score of 65.77%. It shows moderate benefit from ensembling models pre-trained on independent categories. The results for subtask 2 warrant further investigation for improvement.

We present a new gold-standard dataset and a benchmark for the Research Theme Identification task, a sub-task of the Scholarly Knowledge Graph Generation shared task, at the 3rd Workshop on Scholarly Document Processing. The objective of the shared task was to label given research papers with research themes from a total of 36 themes. The benchmark was compiled using data drawn from the largest overall assessment of university research output ever undertaken globally (the Research Excellence Framework - 2014). We provide a performance comparison of a transformer-based ensemble, which obtains multiple predictions for a research paper, given its multiple textual fields (e.g. title, abstract, reference), with traditional machine learning models. The ensemble involves enriching the initial data with additional information from open-access digital libraries and Argumentative Zoning techniques (CITATION). It uses a weighted sum aggregation for the multiple predictions to obtain a final single prediction for the given research paper. Both data and the ensemble are publicly available on https://www.kaggle.com/competitions/sdp2022-scholarly-knowledge-graph-generation/data?select=task1_test_no_label.csv and https://github.com/ProjectDoSSIER/sdp2022, respectively.

This paper describes our contribution to the Answer Localization track of the MedVidQA 2022 Shared Task. We propose two answer localization approaches that use only textual information extracted from the video. In particular, our approaches exploit the text extracted from the video’s transcripts along with the text displayed in the video’s frames to create a set of features. Having created a set of features that represents a video’s textual information, we employ four different models to measure the similarity between a video’s segment and a corresponding question. Then, we employ two different methods to obtain the start and end times of the identified answer. One of them is based on a random forest regressor, whereas the other one uses an unsupervised peak detection model to detect the answer’s start time. Our findings suggest that for this task, leveraging only text-related features (transmitted either verbally or visually) and using a small amount of training data, lead to significant improvements over the benchmark Video Span Localization model that is based on deep neural networks.

Large-scale language modeling and natural language prompting have demonstrated exciting capabilities for few and zero shot learning in NLP. However, translating these successes to specialized domains such as biomedicine remains challenging, due in part to biomedical NLP’s significant dataset debt – the technical costs associated with data that are not consistently documented or easily incorporated into popular machine learning frameworks at scale. To assess this debt, we crowdsourced curation of datasheets for 167 biomedical datasets. We find that only 13% of datasets are available via programmatic access and 30% lack any documentation on licensing and permitted reuse. Our dataset catalog is available at: https://tinyurl.com/bigbio22.

This paper evaluates the impact of various event extraction systems on automatic pathway curation using the popular mTOR pathway. We quantify the impact of training data sets as well as different machine learning classifiers and show that some improve the quality of automatically extracted pathways.