Desmond Elliott

Ilias Chalkidis, Xiang Dai, Manos Fergadiotis et al.

Non-hierarchical sparse attention Transformer-based models, such as Longformer and Big Bird, are popular approaches to working with long documents. There are clear benefits to these approaches compared to the original Transformer in terms of efficiency, but Hierarchical Attention Transformer (HAT) models are a vastly understudied alternative. We develop and release fully pre-trained HAT models that use segment-wise followed by cross-segment encoders and compare them with Longformer models and partially pre-trained HATs. In several long document downstream classification tasks, our best HAT model outperforms equally-sized Longformer models while using 10-20% less GPU memory and processing documents 40-45% faster. In a series of ablation studies, we find that HATs perform best with cross-segment contextualization throughout the model than alternative configurations that implement either early or late cross-segment contextualization. Our code is on GitHub: https://github.com/coastalcph/hierarchical-transformers.

Rita Ramos, Bruno Martins, Desmond Elliott et al.

Recent advances in image captioning have focused on scaling the data and model size, substantially increasing the cost of pre-training and finetuning. As an alternative to large models, we present SmallCap, which generates a caption conditioned on an input image and related captions retrieved from a datastore. Our model is lightweight and fast to train, as the only learned parameters are in newly introduced cross-attention layers between a pre-trained CLIP encoder and GPT-2 decoder. SmallCap can transfer to new domains without additional finetuning and can exploit large-scale data in a training-free fashion since the contents of the datastore can be readily replaced. Our experiments show that SmallCap, trained only on COCO, has competitive performance on this benchmark, and also transfers to other domains without retraining, solely through retrieval from target-domain data. Further improvement is achieved through the training-free exploitation of diverse human-labeled and web data, which proves to be effective for a range of domains, including the nocaps benchmark, designed to test generalization to unseen visual concepts.

Xiang Dai, Ilias Chalkidis, Sune Darkner et al.

The recent literature in text classification is biased towards short text sequences (e.g., sentences or paragraphs). In real-world applications, multi-page multi-paragraph documents are common and they cannot be efficiently encoded by vanilla Transformer-based models. We compare different Transformer-based Long Document Classification (TrLDC) approaches that aim to mitigate the computational overhead of vanilla transformers to encode much longer text, namely sparse attention and hierarchical encoding methods. We examine several aspects of sparse attention (e.g., size of local attention window, use of global attention) and hierarchical (e.g., document splitting strategy) transformers on four document classification datasets covering different domains. We observe a clear benefit from being able to process longer text, and, based on our results, we derive practical advice of applying Transformer-based models on long document classification tasks.

Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello et al.

Language models are defined over a finite set of inputs, which creates a vocabulary bottleneck when we attempt to scale the number of supported languages. Tackling this bottleneck results in a trade-off between what can be represented in the embedding matrix and computational issues in the output layer. This paper introduces PIXEL, the Pixel-based Encoder of Language, which suffers from neither of these issues. PIXEL is a pretrained language model that renders text as images, making it possible to transfer representations across languages based on orthographic similarity or the co-activation of pixels. PIXEL is trained to reconstruct the pixels of masked patches instead of predicting a distribution over tokens. We pretrain the 86M parameter PIXEL model on the same English data as BERT and evaluate on syntactic and semantic tasks in typologically diverse languages, including various non-Latin scripts. We find that PIXEL substantially outperforms BERT on syntactic and semantic processing tasks on scripts that are not found in the pretraining data, but PIXEL is slightly weaker than BERT when working with Latin scripts. Furthermore, we find that PIXEL is more robust than BERT to orthographic attacks and linguistic code-switching, further confirming the benefits of modelling language with pixels.

Rita Ramos, Desmond Elliott, Bruno Martins

Inspired by retrieval-augmented language generation and pretrained Vision and Language (V&L) encoders, we present a new approach to image captioning that generates sentences given the input image and a set of captions retrieved from a datastore, as opposed to the image alone. The encoder in our model jointly processes the image and retrieved captions using a pretrained V&L BERT, while the decoder attends to the multimodal encoder representations, benefiting from the extra textual evidence from the retrieved captions. Experimental results on the COCO dataset show that image captioning can be effectively formulated from this new perspective. Our model, named EXTRA, benefits from using captions retrieved from the training dataset, and it can also benefit from using an external dataset without the need for retraining. Ablation studies show that retrieving a sufficient number of captions (e.g., k=5) can improve captioning quality. Our work contributes towards using pretrained V&L encoders for generative tasks, instead of standard classification tasks.

Chen Qiu, Dan Oneata, Emanuele Bugliarello et al.

Most vision-and-language pretraining research focuses on English tasks. However, the creation of multilingual multimodal evaluation datasets (e.g. Multi30K, xGQA, XVNLI, and MaRVL) poses a new challenge in finding high-quality training data that is both multilingual and multimodal. In this paper, we investigate whether machine translating English multimodal data can be an effective proxy for the lack of readily available multilingual data. We call this framework TD-MML: Translated Data for Multilingual Multimodal Learning, and it can be applied to any multimodal dataset and model. We apply it to both pretraining and fine-tuning data with a state-of-the-art model. In order to prevent models from learning from low-quality translated text, we propose two metrics for automatically removing such translations from the resulting datasets. In experiments on five tasks across 20 languages in the IGLUE benchmark, we show that translated data can provide a useful signal for multilingual multimodal learning, both at pretraining and fine-tuning.

Nadav Borenstein, Phillip Rust, Desmond Elliott et al.

The digitisation of historical documents has provided historians with unprecedented research opportunities. Yet, the conventional approach to analysing historical documents involves converting them from images to text using OCR, a process that overlooks the potential benefits of treating them as images and introduces high levels of noise. To bridge this gap, we take advantage of recent advancements in pixel-based language models trained to reconstruct masked patches of pixels instead of predicting token distributions. Due to the scarcity of real historical scans, we propose a novel method for generating synthetic scans to resemble real historical documents. We then pre-train our model, PHD, on a combination of synthetic scans and real historical newspapers from the 1700-1900 period. Through our experiments, we demonstrate that PHD exhibits high proficiency in reconstructing masked image patches and provide evidence of our model's noteworthy language understanding capabilities. Notably, we successfully apply our model to a historical QA task, highlighting its usefulness in this domain.

Jonas F. Lotz, Elizabeth Salesky, Phillip Rust et al.

Pixel-based language models process text rendered as images, which allows them to handle any script, making them a promising approach to open vocabulary language modelling. However, recent approaches use text renderers that produce a large set of almost-equivalent input patches, which may prove sub-optimal for downstream tasks, due to redundancy in the input representations. In this paper, we investigate four approaches to rendering text in the PIXEL model (Rust et al., 2023), and find that simple character bigram rendering brings improved performance on sentence-level tasks without compromising performance on token-level or multilingual tasks. This new rendering strategy also makes it possible to train a more compact model with only 22M parameters that performs on par with the original 86M parameter model. Our analyses show that character bigram rendering leads to a consistently better model but with an anisotropic patch embedding space, driven by a patch frequency bias, highlighting the connections between image patch- and tokenization-based language models.

Galadrielle Humblot-Renaux, Mohammad N. S. Jahromi, Rohat Bakuri-Jørgensen et al.

Off-the-shelf large language models (LLMs) are increasingly used to automate text annotation, yet their effectiveness remains underexplored for underrepresented languages and specialized domains where the class definition requires subtle expert understanding. We investigate LLM-based annotation for a novel legal NLP task: identifying the presence and sentiment of credibility assessments in asylum decision texts. We introduce RAB-Cred, a Danish text classification dataset featuring high-quality, expert annotations and valuable metadata such as annotator confidence and asylum case outcome. We benchmark 21 open-weight models and 30 system-user prompt combinations for this task, and systematically evaluate the effect of model and prompt choice for zero-shot and few-shot classification. We zoom in on the errors made by top-performing models and prompts, investigating error consistency across LLMs, inter-class confusion, correlation with human confidence and sample-wise difficulty and severity of LLM mistakes. Our results confirm the potential of LLMs for cost-effective labeling of asylum decisions, but highlight the imperfect and inconsistent nature of LLM annotators, and the need to look beyond the predictions of a single, arbitrarily chosen model. The RAB-Cred dataset and code are available at https://github.com/glhr/RAB-Cred

Mike Zhang, Ali Basirat, Desmond Elliott

Prior work establishes that controlled contrastiveness between self-generated responses from large language models, set via reward scores, improves downstream preference tuning in English. We extend this method to multiple languages and evaluate two models across a total of 14 high and low-resource languages on a diverse set of tasks. Our central finding is that cross-lingual contrastive preference tuning on self-generations (CroCo) transfers without language-specific preference annotation. A reward model trained on English preferences (atop a multilingual base) produces useful within-language rankings across most languages, and pairing in either a monolingual or multilingual setting improves over each model on the majority of setups while preventing the catastrophic forgetting of supervised fine-tuning. We observe that the gains require on-policy data. Off-policy responses reduce the benefit and online preference optimization fails to improve over the offline variant. Specifically, on structured tasks, our method matches or exceeds the base in 6/7 languages for EuroLLM-9B and 4/7 settings for Aya-3B. On open-ended generation, both tuned models win against their respective base across 11 evaluated languages. Overall, we show promising directions for multilingual preference tuning.

Ingo Ziegler, Abdullatif Köksal, Desmond Elliott et al.

Building high-quality datasets for specialized tasks is a time-consuming and resource-intensive process that often requires specialized domain knowledge. We propose Corpus Retrieval and Augmentation for Fine-Tuning (CRAFT), a method for generating synthetic datasets, given a small number of user-written few-shots that demonstrate the task to be performed. Given these examples, CRAFT uses large-scale public web-crawled corpora and similarity-based document retrieval to find other relevant human-written documents. Lastly, instruction-tuned large language models (LLMs) augment the retrieved documents into custom-formatted task samples, which then can be used for fine-tuning. We demonstrate that CRAFT can efficiently generate large-scale task-specific training datasets for four diverse tasks: biology, medicine, and commonsense question-answering (QA), as well as summarization. Our experiments show that CRAFT-based models outperform or match general LLMs on QA tasks, while exceeding models trained on human-curated summarization data by 46 preference points. CRAFT outperforms other synthetic dataset generation methods such as Self- and Evol-Instruct, and remains robust even when the quality of the initial few-shots varies.

Vincent Beliveau, Helene Kaas, Martin Prener et al.

Natural language processing (NLP) in the medical domain can underperform in real-world applications involving small datasets in a non-English language with few labeled samples and imbalanced classes. There is yet no consensus on how to approach this problem. We evaluated a set of NLP models including BERT-like transformers, few-shot learning with sentence transformers (SetFit), and prompted large language models (LLM), using three datasets of radiology reports on magnetic resonance images of epilepsy patients in Danish, a low-resource language. Our results indicate that BERT-like models pretrained in the target domain of radiology reports currently offer the optimal performances for this scenario. Notably, the SetFit and LLM models underperformed compared to BERT-like models, with LLM performing the worst. Importantly, none of the models investigated was sufficiently accurate to allow for text classification without any supervision. However, they show potential for data filtering, which could reduce the amount of manual labeling required.

Yiyi Chen, Qiongkai Xu, Desmond Elliott et al.

Image embeddings are generally assumed to pose limited privacy risk. We challenge this assumption by formalizing semantic leakage as the ability to recover semantic structures from compressed image embeddings. Surprisingly, we show that semantic leakage does not require exact reconstruction of the original image. Preserving local semantic neighborhoods under embedding alignment is sufficient to expose the intrinsic vulnerability of image embeddings. Crucially, this preserved neighborhood structure allows semantic information to propagate through a sequence of lossy mappings. Based on this conjecture, we propose Semantic Leakage from Image Embeddings (SLImE), a lightweight inference framework that reveals semantic information from standalone compressed image embeddings, incorporating a locally trained semantic retriever with off-the-shelf models, without training task-specific decoders. We thoroughly validate each step of the framework empirically, from aligned embeddings to retrieved tags, symbolic representations, and grammatical and coherent descriptions. We evaluate SLImE across a range of open and closed embedding models, including GEMINI, COHERE, NOMIC, and CLIP, and demonstrate consistent recovery of semantic information across diverse inference tasks. Our results reveal a fundamental vulnerability in image embeddings, whereby the preservation of semantic neighborhoods under alignment enables semantic leakage, highlighting challenges for privacy preservation.1

Laura Cabello, Emanuele Bugliarello, Stephanie Brandl et al.

Pretrained machine learning models are known to perpetuate and even amplify existing biases in data, which can result in unfair outcomes that ultimately impact user experience. Therefore, it is crucial to understand the mechanisms behind those prejudicial biases to ensure that model performance does not result in discriminatory behaviour toward certain groups or populations. In this work, we define gender bias as our case study. We quantify bias amplification in pretraining and after fine-tuning on three families of vision-and-language models. We investigate the connection, if any, between the two learning stages, and evaluate how bias amplification reflects on model performance. Overall, we find that bias amplification in pretraining and after fine-tuning are independent. We then examine the effect of continued pretraining on gender-neutral data, finding that this reduces group disparities, i.e., promotes fairness, on VQAv2 and retrieval tasks without significantly compromising task performance.

Danae Sánchez Villegas, Samuel Lewis-Lim, Nikolaos Aletras et al.

Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear. We analyze reasoning dynamics in 18 VLMs covering instruction-tuned and reasoning-trained models from two different model families. We track confidence over Chain-of-Thought (CoT), measure the corrective effect of reasoning, and evaluate the contribution of intermediate reasoning steps. We find that models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, we show that models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues, but their longer and fluent CoTs can still appear visually grounded while actually following textual cues, obscuring modality reliance. In contrast, instruction-tuned models refer to the cues less explicitly, but their shorter traces reveal inconsistencies with the visual input. Taken together, these findings indicate that CoT provides only a partial view of how different modalities drive VLM decisions, with important implications for the transparency and safety of multimodal systems.

Alice Schiavone, Marco Fraccaro, Lea Marie Pehrson et al.

Radiology reports contain rich clinical information that can be used to train imaging models without relying on costly manual annotation. However, existing approaches face critical limitations: rule-based methods struggle with linguistic variability, supervised models require large annotated datasets, and recent LLM-based systems depend on closed-source or resource-intensive models that are unsuitable for clinical use. Moreover, current solutions are largely restricted to English and single-modality, single-taxonomy datasets. We introduce MOSAIC, a multilingual, taxonomy-agnostic, and computationally efficient approach for radiological report classification. Built on a compact open-access language model (MedGemma-4B), MOSAIC supports both zero-/few-shot prompting and lightweight fine-tuning, enabling deployment on consumer-grade GPUs. We evaluate MOSAIC across seven datasets in English, Spanish, French, and Danish, spanning multiple imaging modalities and label taxonomies. The model achieves a mean macro F1 score of 88 across five chest X-ray datasets, approaching or exceeding expert-level performance, while requiring only 24 GB of GPU memory. With data augmentation, as few as 80 annotated samples are sufficient to reach a weighted F1 score of 82 on Danish reports, compared to 86 with the full 1600-sample training set. MOSAIC offers a practical alternative to large or proprietary LLMs in clinical settings. Code and models are open-source. We invite the community to evaluate and extend MOSAIC on new languages, taxonomies, and modalities.

Wenyan Li, Xinyu Zhang, Jiaang Li et al.

Food is a rich and varied dimension of cultural heritage, crucial to both individuals and social groups. To bridge the gap in the literature on the often-overlooked regional diversity in this domain, we introduce FoodieQA, a manually curated, fine-grained image-text dataset capturing the intricate features of food cultures across various regions in China. We evaluate vision-language Models (VLMs) and large language models (LLMs) on newly collected, unseen food images and corresponding questions. FoodieQA comprises three multiple-choice question-answering tasks where models need to answer questions based on multiple images, a single image, and text-only descriptions, respectively. While LLMs excel at text-based question answering, surpassing human accuracy, the open-sourced VLMs still fall short by 41% on multi-image and 21% on single-image VQA tasks, although closed-weights models perform closer to human levels (within 10%). Our findings highlight that understanding food and its cultural implications remains a challenging and under-explored direction.

Laura De Grazia, Danae Sánchez Villegas, Desmond Elliott et al.

Online sexism appears in various forms, which makes its detection challenging. Although automated tools can enhance the identification of sexist content, they are often restricted to binary classification. Consequently, more subtle manifestations of sexism may remain undetected due to the lack of fine-grained, context-sensitive labels. To address this issue, we make the following contributions: (1) we present FineMuSe, a new multimodal sexism detection dataset in Spanish that includes both binary and fine-grained annotations; (2) we introduce a comprehensive hierarchical taxonomy that encompasses forms of sexism, non-sexism, and rhetorical devices of irony and humor; and (3) we evaluate a wide range of LLMs for both binary and fine-grained sexism detection. Our findings indicate that multimodal LLMs perform competitively with human annotators in identifying nuanced forms of sexism; however, they struggle to capture co-occurring sexist types when these are conveyed through visual cues.

Constanza Fierro, Negar Foroutan, Desmond Elliott et al.

Large Language Models (LLMs) store and retrieve vast amounts of factual knowledge acquired during pre-training. Prior research has localized and identified mechanisms behind knowledge recall; however, it has only focused on English monolingual models. The question of how these mechanisms generalize to non-English languages and multilingual LLMs remains unexplored. In this paper, we address this gap by conducting a comprehensive analysis of three multilingual LLMs. First, we show that previously identified recall mechanisms in English largely apply to multilingual contexts, with nuances based on language and architecture. Next, through patching intermediate representations, we localize the role of language during recall, finding that subject enrichment is language-independent, while object extraction is language-dependent. Additionally, we discover that the last token representation acts as a Function Vector (FV), encoding both the language of the query and the content to be extracted from the subject. Furthermore, in decoder-only LLMs, FVs compose these two pieces of information in two separate stages. These insights reveal unique mechanisms in multilingual LLMs for recalling information, highlighting the need for new methodologies -- such as knowledge evaluation, fact editing, and knowledge acquisition -- that are specifically tailored for multilingual LLMs.

Nadav Borenstein, Greta Warren, Desmond Elliott et al.

Two commonly employed strategies to combat the rise of misinformation on social media are (i) fact-checking by professional organisations and (ii) community moderation by platform users. Policy changes by Twitter/X and, more recently, Meta, signal a shift away from partnerships with fact-checking organisations and towards an increased reliance on crowdsourced community notes. However, the extent and nature of dependencies between fact-checking and helpful community notes remain unclear. To address these questions, we use language models to annotate a large corpus of Twitter/X community notes with attributes such as topic, cited sources, and whether they refute claims tied to broader misinformation narratives. Our analysis reveals that community notes cite fact-checking sources up to five times more than previously reported. Fact-checking is especially crucial for notes on posts linked to broader narratives, which are twice as likely to reference fact-checking sources compared to other sources. Our results show that successful community moderation relies on professional fact-checking and highlight how citizen and professional fact-checking are deeply intertwined.

Israfel Salazar, Manuel Fernández Burda, Shayekh Bin Islam et al. · mila

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

Laura De Grazia, Pol Pastells, Mauro Vázquez Chas et al.

Sexism is generally defined as prejudice and discrimination based on sex or gender, affecting every sector of society, from social institutions to relationships and individual behavior. Social media platforms amplify the impact of sexism by conveying discriminatory content not only through text but also across multiple modalities, highlighting the critical need for a multimodal approach to the analysis of sexism online. With the rise of social media platforms where users share short videos, sexism is increasingly spreading through video content. Automatically detecting sexism in videos is a challenging task, as it requires analyzing the combination of verbal, audio, and visual elements to identify sexist content. In this study, (1) we introduce MuSeD, a new Multimodal Spanish dataset for Sexism Detection consisting of $\approx$ 11 hours of videos extracted from TikTok and BitChute; (2) we propose an innovative annotation framework for analyzing the contributions of textual, vocal, and visual modalities to the classification of content as either sexist or non-sexist; and (3) we evaluate a range of large language models (LLMs) and multimodal LLMs on the task of sexism detection. We find that visual information plays a key role in labeling sexist content for both humans and models. Models effectively detect explicit sexism; however, they struggle with implicit cases, such as stereotypes, instances where annotators also show low agreement. This highlights the inherent difficulty of the task, as identifying implicit sexism depends on the social and cultural context.

Danae Sánchez Villegas, Ingo Ziegler, Desmond Elliott

Reasoning over sequences of images remains a challenge for multimodal large language models (MLLMs). While recent models incorporate multi-image data during pre-training, they still struggle to recognize sequential structures, often treating images independently. This work introduces ImageChain, a framework that enhances MLLMs with sequential reasoning capabilities over image data by modeling visual sequences as a multi-turn conversation. In ImageChain, images are interleaved with corresponding textual descriptions to form a controlled dialogue that explicitly captures temporal dependencies and narrative progression. Our method optimizes for the task of next-scene description, where the model generates a context-aware description of an upcoming scene based on preceding visual and textual cues. We demonstrate that our approach improves performance on the next-scene description task -- achieving an average improvement from 3.7% to 19% in SimRate, a metric that quantifies semantic similarity to human-annotated ground truths. Moreover, ImageChain achieves robust zero-shot out-of-domain performance in applications ranging from comics to robotics. Extensive experiments validate that instruction-tuning in a multimodal, multi-turn conversation design is key to bridging the gap between static image understanding and temporally-aware reasoning.

Konstantin Dobler, Desmond Elliott, Gerard de Melo

Current language models rely on static vocabularies determined at pretraining time, which can lead to decreased performance and increased computational cost for domains underrepresented in the original vocabulary. New tokens can be added to solve this problem, when coupled with a good initialization for their new embeddings. However, existing embedding initialization methods require expensive further training or pretraining of additional modules. In this paper, we propose Token Distillation and show that by distilling representations obtained using the original tokenization, we can quickly learn high-quality input embeddings for new tokens. Experimental results with a wide range of open-weight models show that Token Distillation outperforms even strong baselines.

Dan Oneata, Desmond Elliott, Stella Frank

Human learning and conceptual representation is grounded in sensorimotor experience, in contrast to state-of-the-art foundation models. In this paper, we investigate how well such large-scale models, trained on vast quantities of data, represent the semantic feature norms of concrete object concepts, e.g. a ROSE is red, smells sweet, and is a flower. More specifically, we use probing tasks to test which properties of objects these models are aware of. We evaluate image encoders trained on image data alone, as well as multimodally-trained image encoders and language-only models, on predicting an extended denser version of the classic McRae norms and the newer Binder dataset of attribute ratings. We find that multimodal image encoders slightly outperform language-only approaches, and that image-only encoders perform comparably to the language models, even on non-visual attributes that are classified as "encyclopedic" or "function". These results offer new insights into what can be learned from pure unimodal learning, and the complementarity of the modalities.

Ram Mohan Rao Kadiyala, Siddhant Gupta, Jebish Purbey et al.

Vision-Language Models (VLMs) have demonstrated impressive capabilities across a range of tasks, yet concerns about their potential biases exist. This work investigates the extent to which prominent VLMs exhibit cultural biases by evaluating their performance on an image-based country identification task at a country level. Utilizing the geographically diverse Country211 dataset, we probe several large vision language models (VLMs) under various prompting strategies: open-ended questions, multiple-choice questions (MCQs) including challenging setups like multilingual and adversarial settings. Our analysis aims to uncover disparities in model accuracy across different countries and question formats, providing insights into how training data distribution and evaluation methodologies might influence cultural biases in VLMs. The findings highlight significant variations in performance, suggesting that while VLMs possess considerable visual understanding, they inherit biases from their pre-training data and scale that impact their ability to generalize uniformly across diverse global contexts.

Semih Yagcioglu, Osman Batur İnce, Aykut Erdem et al.

The rise of large-scale multimodal models has paved the pathway for groundbreaking advances in generative modeling and reasoning, unlocking transformative applications in a variety of complex tasks. However, a pressing question that remains is their genuine capability for stronger forms of generalization, which has been largely underexplored in the multimodal setting. Our study aims to address this by examining sequential compositional generalization using \textsc{CompAct} (\underline{Comp}ositional \underline{Act}ivities)\footnote{Project Page: \url{http://cyberiada.github.io/CompAct}}, a carefully constructed, perceptually grounded dataset set within a rich backdrop of egocentric kitchen activity videos. Each instance in our dataset is represented with a combination of raw video footage, naturally occurring sound, and crowd-sourced step-by-step descriptions. More importantly, our setup ensures that the individual concepts are consistently distributed across training and evaluation sets, while their compositions are novel in the evaluation set. We conduct a comprehensive assessment of several unimodal and multimodal models. Our findings reveal that bi-modal and tri-modal models exhibit a clear edge over their text-only counterparts. This highlights the importance of multimodality while charting a trajectory for future research in this domain.

Amit Moryossef, Clara Meister, Pavel Stepachev et al.

We present UTF8Tokenizer, a minimalist byte-level tokenizer that maps text exactly to IDs corresponding to the bytes underlying the text's UTF-8 encoding (e.g., byte x09 is token ID 9). Unlike prior byte-level approaches (Xue et al., 2021; Pagnoni et al., 2025), our implementation never introduces out-of-range IDs (i.e. there is no token ID 256) or auxiliary tokens: all special behavior (e.g., padding, boundaries, conversation structure, attention segments, tool calling, "thinking" spans, etc.) is encoded using C0 control bytes - just as ASCII was originally designed to embed control information alongside printable text. These design principles yield practical benefits: (1) faster tokenization (14x) and significantly lower host-device transfer (8x less than int64); (2) simple, shareable 256*d embedding tables that can be aligned across models; and (3) a training-time enhancement via bit-biased embeddings, which exposes per-byte bit structure and can be added to the embedding table post-training, removing inference costs. Our HuggingFace-compatible implementation improves language modeling convergence.

Mehmet Onurcan Kaya, Desmond Elliott, Dim P. Papadopoulos

Small Vision-Language Models (VLMs) provide a computationally efficient alternative to larger models, at the cost of weaker generalization abilities and downstream task performance. These shortcomings could be addressed by test-time scaling techniques, but existing methods are typically computationally demanding, contradicting the resource-efficient design goals of small models. To address these limitations, we propose two novel and efficient test-time scaling strategies that leverage the model-internal features rather than external supervision: (i) Test-Time Augmentation (TTAug), which generates multiple augmented inputs and aggregates outputs at the token level without parameter updates, and (ii) Test-Time Adaptation (TTAdapt), which adapts model parameters during inference using consensus-based pseudolabels from TTAug. Through extensive experiments across nine benchmarks, we demonstrate consistent performance improvements while maintaining computational efficiency suitable for resource-constrained environments. The generality of our approach is demonstrated both within models at different scales and across different VLMs without additional tuning.

Israfel Salazar, Desmond Elliott, Yova Kementchedjhieva

Contrastive vision-language models (VLMs) have made significant progress in binding visual and textual information, but understanding long, dense captions remains an open challenge. We hypothesize that compositionality, the capacity to reason about object-attribute bindings and inter-object relationships, is key to understanding longer captions. In this paper, we investigate the interaction between compositionality and long-caption understanding, asking whether training for one property enhances the other. We train and evaluate a range of models that target each of these capabilities. Our results reveal a bidirectional relationship: compositional training improves performance on long-caption retrieval, and training on long captions promotes compositionality. However, these gains are sensitive to data quality and model design. We find that training on poorly structured captions, or with limited parameter updates, fails to support generalization. Likewise, strategies that aim at retaining general alignment, such as freezing positional embeddings, do not improve compositional understanding. Overall, we find that compositional understanding and long-caption understanding are intertwined capabilities that can be jointly learned through training on dense, grounded descriptions. Despite these challenges, we show that models trained on high-quality, long-caption data can achieve strong performance in both tasks, offering practical guidance for improving VLM generalization.

Kenneth Enevoldsen, Kristian Nørgaard Jensen, Jan Kostkan et al.

Large-scale datasets are foundational for research and development in natural language processing. However, current approaches face three key challenges: (1) reliance on ambiguously licensed sources restricting use, sharing, and derivative works; (2) static dataset releases that prevent community contributions and diminish longevity; and (3) quality assurance processes restricted to publishing teams rather than leveraging community expertise. To address these limitations, we introduce two contributions: the Dynaword approach and Danish Dynaword. The Dynaword approach is a framework for creating large-scale, open datasets that can be continuously updated through community collaboration. Danish Dynaword is a concrete implementation that validates this approach and demonstrates its potential. Danish Dynaword contains over four times as many tokens as comparable releases, is exclusively openly licensed, and has received multiple contributions across industry and research. The repository includes light-weight tests to ensure data formatting, quality, and documentation, establishing a sustainable framework for ongoing community contributions and dataset evolution.

Ilker Kesen, Jonas F. Lotz, Ingo Ziegler et al.

Pixel language models operate directly on images of rendered text, eliminating the need for a fixed vocabulary. While these models have demonstrated strong capabilities for downstream cross-lingual transfer, multilingual pretraining remains underexplored. We introduce PIXEL-M4, a model pretrained on four visually and linguistically diverse languages: English, Hindi, Ukrainian, and Simplified Chinese. Multilingual evaluations on semantic and syntactic tasks show that PIXEL-M4 outperforms an English-only counterpart on non-Latin scripts. Word-level probing analyses confirm that PIXEL-M4 captures rich linguistic features, even in languages not seen during pretraining. Furthermore, an analysis of its hidden representations shows that multilingual pretraining yields a semantic embedding space closely aligned across the languages used for pretraining. This work demonstrates that multilingual pretraining substantially enhances the capability of pixel language models to effectively support a diverse set of languages.

Niels Horn, Desmond Elliott

We study how features emerge, disappear, and persist across models fine-tuned on different domains of text. More specifically, we start from a base one-layer Transformer language model that is trained on a combination of the BabyLM corpus, and a collection of Python code from The Stack. This base model is adapted to two new domains of text: TinyStories, and the Lua programming language, respectively; and then these two models are merged using these two models using spherical linear interpolation. Our exploration aims to provide deeper insights into the stability and transformation of features across typical transfer-learning scenarios using small-scale models and sparse auto-encoders.

Anna Bavaresco, Raffaella Bernardi, Leonardo Bertolazzi et al.

There is an increasing trend towards evaluating NLP models with LLMs instead of human judgments, raising questions about the validity of these evaluations, as well as their reproducibility in the case of proprietary models. We provide JUDGE-BENCH, an extensible collection of 20 NLP datasets with human annotations covering a broad range of evaluated properties and types of data, and comprehensively evaluate 11 current LLMs, covering both open-weight and proprietary models, for their ability to replicate the annotations. Our evaluations show substantial variance across models and datasets. Models are reliable evaluators on some tasks, but overall display substantial variability depending on the property being evaluated, the expertise level of the human judges, and whether the language is human or model-generated. We conclude that LLMs should be carefully validated against human judgments before being used as evaluators.

Wenyan Li, Jiaang Li, Rita Ramos et al.

Recent advances in retrieval-augmented models for image captioning highlight the benefit of retrieving related captions for efficient, lightweight models with strong domain-transfer capabilities. While these models demonstrate the success of retrieval augmentation, retrieval models are still far from perfect in practice: the retrieved information can sometimes mislead the model, resulting in incorrect generation and worse performance. In this paper, we analyze the robustness of a retrieval-augmented captioning model SmallCap. Our analysis shows that the model is sensitive to tokens that appear in the majority of the retrieved captions, and the input attribution shows that those tokens are likely copied into the generated output. Given these findings, we propose to train the model by sampling retrieved captions from more diverse sets. This decreases the chance that the model learns to copy majority tokens, and improves both in-domain and cross-domain performance.

Rita Ramos, Bruno Martins, Desmond Elliott

Multilingual image captioning has recently been tackled by training with large-scale machine translated data, which is an expensive, noisy, and time-consuming process. Without requiring any multilingual caption data, we propose LMCap, an image-blind few-shot multilingual captioning model that works by prompting a language model with retrieved captions. Specifically, instead of following the standard encoder-decoder paradigm, given an image, LMCap first retrieves the captions of similar images using a multilingual CLIP encoder. These captions are then combined into a prompt for an XGLM decoder, in order to generate captions in the desired language. In other words, the generation model does not directly process the image, instead processing retrieved captions. Experiments on the XM3600 dataset of geographically diverse images show that our model is competitive with fully-supervised multilingual captioning models, without requiring any supervised training on any captioning data.

Wenyan Li, Jonas F. Lotz, Chen Qiu et al.

Image captioning models are typically trained by treating all samples equally, neglecting to account for mismatched or otherwise difficult data points. In contrast, recent work has shown the effectiveness of training models by scheduling the data using curriculum learning strategies. This paper contributes to this direction by actively curating difficult samples in datasets without increasing the total number of samples. We explore the effect of using three data curation methods within the training process: complete removal of an sample, caption replacement, or image replacement via a text-to-image generation model. Experiments on the Flickr30K and COCO datasets with the BLIP and BEiT-3 models demonstrate that these curation methods do indeed yield improved image captioning models, underscoring their efficacy.

Emanuele Bugliarello, Fangyu Liu, Jonas Pfeiffer et al.

Reliable evaluation benchmarks designed for replicability and comprehensiveness have driven progress in machine learning. Due to the lack of a multilingual benchmark, however, vision-and-language research has mostly focused on English language tasks. To fill this gap, we introduce the Image-Grounded Language Understanding Evaluation benchmark. IGLUE brings together - by both aggregating pre-existing datasets and creating new ones - visual question answering, cross-modal retrieval, grounded reasoning, and grounded entailment tasks across 20 diverse languages. Our benchmark enables the evaluation of multilingual multimodal models for transfer learning, not only in a zero-shot setting, but also in newly defined few-shot learning setups. Based on the evaluation of the available state-of-the-art models, we find that translate-test transfer is superior to zero-shot transfer and that few-shot learning is hard to harness for many tasks. Moreover, downstream performance is partially explained by the amount of available unlabelled textual data for pretraining, and only weakly by the typological distance of target-source languages. We hope to encourage future research efforts in this area by releasing the benchmark to the community.

Fangyu Liu, Emanuele Bugliarello, Edoardo Maria Ponti et al.

The design of widespread vision-and-language datasets and pre-trained encoders directly adopts, or draws inspiration from, the concepts and images of ImageNet. While one can hardly overestimate how much this benchmark contributed to progress in computer vision, it is mostly derived from lexical databases and image queries in English, resulting in source material with a North American or Western European bias. Therefore, we devise a new protocol to construct an ImageNet-style hierarchy representative of more languages and cultures. In particular, we let the selection of both concepts and images be entirely driven by native speakers, rather than scraping them automatically. Specifically, we focus on a typologically diverse set of languages, namely, Indonesian, Mandarin Chinese, Swahili, Tamil, and Turkish. On top of the concepts and images obtained through this new protocol, we create a multilingual dataset for {M}ulticultur{a}l {R}easoning over {V}ision and {L}anguage (MaRVL) by eliciting statements from native speaker annotators about pairs of images. The task consists of discriminating whether each grounded statement is true or false. We establish a series of baselines using state-of-the-art models and find that their cross-lingual transfer performance lags dramatically behind supervised performance in English. These results invite us to reassess the robustness and accuracy of current state-of-the-art models beyond a narrow domain, but also open up new exciting challenges for the development of truly multilingual and multicultural systems.

Rasmus Kær Jørgensen, Mareike Hartmann, Xiang Dai et al.

Domain adaptive pretraining, i.e. the continued unsupervised pretraining of a language model on domain-specific text, improves the modelling of text for downstream tasks within the domain. Numerous real-world applications are based on domain-specific text, e.g. working with financial or biomedical documents, and these applications often need to support multiple languages. However, large-scale domain-specific multilingual pretraining data for such scenarios can be difficult to obtain, due to regulations, legislation, or simply a lack of language- and domain-specific text. One solution is to train a single multilingual model, taking advantage of the data available in as many languages as possible. In this work, we explore the benefits of domain adaptive pretraining with a focus on adapting to multiple languages within a specific domain. We propose different techniques to compose pretraining corpora that enable a language model to both become domain-specific and multilingual. Evaluation on nine domain-specific datasets-for biomedical named entity recognition and financial sentence classification-covering seven different languages show that a single multilingual domain-specific model can outperform the general multilingual model, and performs close to its monolingual counterpart. This finding holds across two different pretraining methods, adapter-based pretraining and full model pretraining.

Stella Frank, Emanuele Bugliarello, Desmond Elliott

Pretrained vision-and-language BERTs aim to learn representations that combine information from both modalities. We propose a diagnostic method based on cross-modal input ablation to assess the extent to which these models actually integrate cross-modal information. This method involves ablating inputs from one modality, either entirely or selectively based on cross-modal grounding alignments, and evaluating the model prediction performance on the other modality. Model performance is measured by modality-specific tasks that mirror the model pretraining objectives (e.g. masked language modelling for text). Models that have learned to construct cross-modal representations using both modalities are expected to perform worse when inputs are missing from a modality. We find that recently proposed models have much greater relative difficulty predicting text when visual information is ablated, compared to predicting visual object categories when text is ablated, indicating that these models are not symmetrically cross-modal.

Emanuele Bugliarello, Desmond Elliott

Image captioning has focused on generalizing to images drawn from the same distribution as the training set, and not to the more challenging problem of generalizing to different distributions of images. Recently, Nikolaus et al. (2019) introduced a dataset to assess compositional generalization in image captioning, where models are evaluated on their ability to describe images with unseen adjective-noun and noun-verb compositions. In this work, we investigate different methods to improve compositional generalization by planning the syntactic structure of a caption. Our experiments show that jointly modeling tokens and syntactic tags enhances generalization in both RNN- and Transformer-based models, while also improving performance on standard metrics.

Emanuele Bugliarello, Ryan Cotterell, Naoaki Okazaki et al.

Large-scale pretraining and task-specific fine-tuning is now the standard methodology for many tasks in computer vision and natural language processing. Recently, a multitude of methods have been proposed for pretraining vision and language BERTs to tackle challenges at the intersection of these two key areas of AI. These models can be categorised into either single-stream or dual-stream encoders. We study the differences between these two categories, and show how they can be unified under a single theoretical framework. We then conduct controlled experiments to discern the empirical differences between five V&L BERTs. Our experiments show that training data and hyperparameters are responsible for most of the differences between the reported results, but they also reveal that the embedding layer plays a crucial role in these massive models.

Tejas Srinivasan, Ramon Sanabria, Florian Metze et al.

Visual context has been shown to be useful for automatic speech recognition (ASR) systems when the speech signal is noisy or corrupted. Previous work, however, has only demonstrated the utility of visual context in an unrealistic setting, where a fixed set of words are systematically masked in the audio. In this paper, we simulate a more realistic masking scenario during model training, called RandWordMask, where the masking can occur for any word segment. Our experiments on the Flickr 8K Audio Captions Corpus show that multimodal ASR can generalize to recover different types of masked words in this unstructured masking setting. Moreover, our analysis shows that our models are capable of attending to the visual signal when the audio signal is corrupted. These results show that multimodal ASR systems can leverage the visual signal in more generalized noisy scenarios.

Bertrand Higy, Desmond Elliott, Grzegorz Chrupała

Visually-grounded models of spoken language understanding extract semantic information directly from speech, without relying on transcriptions. This is useful for low-resource languages, where transcriptions can be expensive or impossible to obtain. Recent work showed that these models can be improved if transcriptions are available at training time. However, it is not clear how an end-to-end approach compares to a traditional pipeline-based approach when one has access to transcriptions. Comparing different strategies, we find that the pipeline approach works better when enough text is available. With low-resource languages in mind, we also show that translations can be effectively used in place of transcriptions but more data is needed to obtain similar results.

Tejas Srinivasan, Ramon Sanabria, Florian Metze et al.

Multimodal automatic speech recognition systems integrate information from images to improve speech recognition quality, by grounding the speech in the visual context. While visual signals have been shown to be useful for recovering entities that have been masked in the audio, these models should be capable of recovering a broader range of word types. Existing systems rely on global visual features that represent the entire image, but localizing the relevant regions of the image will make it possible to recover a larger set of words, such as adjectives and verbs. In this paper, we propose a model that uses finer-grained visual information from different parts of the image, using automatic object proposals. In experiments on the Flickr8K Audio Captions Corpus, we find that our model improves over approaches that use global visual features, that the proposals enable the model to recover entities and other related words, such as adjectives, and that improvements are due to the model's ability to localize the correct proposals.

Alessandro Suglia, Ioannis Konstas, Andrea Vanzo et al.

Approaches to Grounded Language Learning typically focus on a single task-based final performance measure that may not depend on desirable properties of the learned hidden representations, such as their ability to predict salient attributes or to generalise to unseen situations. To remedy this, we present GROLLA, an evaluation framework for Grounded Language Learning with Attributes with three sub-tasks: 1) Goal-oriented evaluation; 2) Object attribute prediction evaluation; and 3) Zero-shot evaluation. We also propose a new dataset CompGuessWhat?! as an instance of this framework for evaluating the quality of learned neural representations, in particular concerning attribute grounding. To this end, we extend the original GuessWhat?! dataset by including a semantic layer on top of the perceptual one. Specifically, we enrich the VisualGenome scene graphs associated with the GuessWhat?! images with abstract and situated attributes. By using diagnostic classifiers, we show that current models learn representations that are not expressive enough to encode object attributes (average F1 of 44.27). In addition, they do not learn strategies nor representations that are robust enough to perform well when novel scenes or objects are involved in gameplay (zero-shot best accuracy 50.06%).

Mostafa Abdou, Vinit Ravishankar, Maria Barrett et al.

Large-scale pretrained language models are the major driving force behind recent improvements in performance on the Winograd Schema Challenge, a widely employed test of common sense reasoning ability. We show, however, with a new diagnostic dataset, that these models are sensitive to linguistic perturbations of the Winograd examples that minimally affect human understanding. Our results highlight interesting differences between humans and language models: language models are more sensitive to number or gender alternations and synonym replacements than humans, and humans are more stable and consistent in their predictions, maintain a much higher absolute performance, and perform better on non-associative instances than associative ones. Overall, humans are correct more often than out-of-the-box models, and the models are sometimes right for the wrong reasons. Finally, we show that fine-tuning on a large, task-specific dataset can offer a solution to these issues.

Umut Sulubacak, Ozan Caglayan, Stig-Arne Grönroos et al.

Multimodal machine translation involves drawing information from more than one modality, based on the assumption that the additional modalities will contain useful alternative views of the input data. The most prominent tasks in this area are spoken language translation, image-guided translation, and video-guided translation, which exploit audio and visual modalities, respectively. These tasks are distinguished from their monolingual counterparts of speech recognition, image captioning, and video captioning by the requirement of models to generate outputs in a different language. This survey reviews the major data resources for these tasks, the evaluation campaigns concentrated around them, the state of the art in end-to-end and pipeline approaches, and also the challenges in performance evaluation. The paper concludes with a discussion of directions for future research in these areas: the need for more expansive and challenging datasets, for targeted evaluations of model performance, and for multimodality in both the input and output space.

Ákos Kádár, Grzegorz Chrupała, Afra Alishahi et al.

Recent work has highlighted the advantage of jointly learning grounded sentence representations from multiple languages. However, the data used in these studies has been limited to an aligned scenario: the same images annotated with sentences in multiple languages. We focus on the more realistic disjoint scenario in which there is no overlap between the images in multilingual image--caption datasets. We confirm that training with aligned data results in better grounded sentence representations than training with disjoint data, as measured by image--sentence retrieval performance. In order to close this gap in performance, we propose a pseudopairing method to generate synthetically aligned English--German--image triplets from the disjoint sets. The method works by first training a model on the disjoint data, and then creating new triples across datasets using sentence similarity under the learned model. Experiments show that pseudopairs improve image--sentence retrieval performance compared to disjoint training, despite requiring no external data or models. However, we do find that using an external machine translation model to generate the synthetic data sets results in better performance.