arXiv subjects

Albert Gatt

Publications and source records attributed to Albert Gatt.

At least 19 recordsLinked to original sources

arXiv2026

From Vision to Language: Investigating Causal Information Flow in Multimodal Decision-Making

Vision-Language Models are commonly evaluated through their final predictions, but understanding whether these decisions are grounded in visual evidence requires tracing how visual information contributes to language-based decisions. With this purpose in mind, we investigate cross-modal information flow in a video-based generative multiple-choice-like setting by applying a layer-wise causal intervention on video-text attention pathways. We target spatial, causal, and temporal visual reasoning. Our results show that visual information is mainly integrated while the model processes the candidate answer options, which serve as the primary textual grounding sites for the final decision. We further show that nouns play an important role as semantic anchors during multimodal enrichment, while verbs are more relevant when temporal relations are processed. Finally, we identify a distinct pattern in temporal reasoning, suggesting that VLMs struggle to reconstruct sequential information across video frames, but we remark that such fragility may also reflect linguistic biases associated with specific temporal expressions used for defining the relation between events within a scene.

cs.CL↗

arXiv2026

Language Re-generation: An investigation into information locality effects on reconstruction

Information locality, the tendency for syntactically related words to appear close together, shapes both human language processing and language model learning. While prior work has examined whether language models can acquire impossible languages, it remains unclear whether they can recover natural language from such input and what this reveals about their inductive biases. We address this by complementing learnability-based approaches with a reconstruction framework: fine-tuning GPT-2 models pre-trained on impossible languages to reconstruct natural English from three perturbation types. Our findings show that the recovered structures exhibit shorter dependency lengths than the original text, mirroring the locality preference observed in unconstrained language model generation and providing a quantitative signature of an architectural bias that learnability experiments alone do not reveal. Recovery difficulty increases with the degree of locality disruption. Structural recovery (dependency Triple F1) dissociates from surface recovery (Exact Match), while fluency dissociates from faithful reconstruction under global shuffling. Sentence length further modulates performance: longer sentences facilitate recovery when local structure is preserved but lead to complete collapse under global shuffling. Finally, recovery difficulty tracks learnability difficulty across perturbation types, suggesting that information locality is the shared constraint governing both.

cs.CL↗

arXiv2026

Beyond wheelchairs and blindfolds: Investigating disability stereotypes in T2I models with INCLUDE-BENCH

Text-to-image (T2I) models have been shown to exhibit social biases. Prior work has mainly focused on gender, skin tone, and cultural representation within restricted occupational associations, and emerging benchmarks increasingly incorporate these dimensions. However, disability remains systematically underexplored. Current evaluation practices often fail to align with sociologically grounded definitions of stereotyping, limiting principled assessment of representational harms toward people with disabilities (PWD). To address this, we introduce INCLUDE-BENCH, the first large-scale benchmark for evaluating disability-related bias in T2I models. INCLUDE-BENCH comprises 119K generated images based on prompt design across multiple bias dimensions and both static and dynamic contexts. We evaluate 15 open-source and 2 closed-source models. Our key findings reveal that: (1) mobility-impaired and default disability prompts predominantly yield wheelchair depictions across all models; (2) disability-conditioned generations consistently exhibit less diversity; (3) stereotypical portrayals demonstrate stronger disability-text alignment; and (4) we introduce the Stereotype Content Model (SCM) Score, demonstrating that T2I models reflect real-world stereotypical associations.

cs.CV↗

arXiv2026

Seeing Is Not Sharing: Some Vision-Language Models Overestimate Common Ground in Asymmetric Dialogue

In collaborative dialogue, shared perception does not guarantee shared interpretation. Mutual understanding must be established through interaction. We investigate whether vision-language models (VLMs) can distinguish what could be shared from what has been shared between dialogue participants through grounding. We formulate this as an interpretation-matching task on 13,077 annotated reference expressions from HCRC MapTask dialogues, and evaluate VLMs under systematically controlled manipulations of dialogue context and map-information access. Our results show that providing authentic map images improves overall performance but shifts models toward over-predicting alignment. Textual descriptions of the same map content reproduce this bias, while non-informative images suppress alignment predictions entirely, indicating that the bias is driven by task-relevant map content, not the visual channel. This improvement comes at the cost of degraded accuracy on non-aligned cases. Calibration analysis and reference-chain tracking further suggest that models rely on static referential cues on the maps rather than tracking how grounding unfolds through dialogue history. We observe these patterns most clearly in Qwen3-VL-8B-Instruct and, to varying degrees, in four additional models from two architecture families. In models that exhibit the bias, map content, whether presented visually or textually, is treated as evidence of mutual understanding, conflating potential with established common ground.

cs.CL↗

arXiv2026

When Models Decide and When They Bind: A Two-Stage Computation for Multiple-Choice Question-Answering

Multiple-choice question answering (MCQA) is easy to evaluate but adds a meta-task: models must both solve the problem and output the symbol that *represents* the answer, conflating reasoning errors with symbol-binding failures. We study how language models implement MCQA internally using representational analyses (PCA, linear probes) as well as causal interventions. We find that option-boundary (newline) residual states often contain strong linearly decodable signals related to per-option correctness. Winner-identity probing reveals a two-stage progression: the winning *content position* becomes decodable immediately after the final option is processed, while the *output symbol* is represented closer to the answer emission position. Tests under symbol and content permutations support a two-stage mechanism in which models first select a winner in content space and then bind or route that winner to the appropriate symbol to emit.

cs.CL↗

arXiv2025

Don't Learn, Ground: A Case for Natural Language Inference with Visual Grounding

We propose a zero-shot method for Natural Language Inference (NLI) that leverages multimodal representations by grounding language in visual contexts. Our approach generates visual representations of premises using text-to-image models and performs inference by comparing these representations with textual hypotheses. We evaluate two inference techniques: cosine similarity and visual question answering. Our method achieves high accuracy without task-specific fine-tuning, demonstrating robustness against textual biases and surface heuristics. Additionally, we design a controlled adversarial dataset to validate the robustness of our approach. Our findings suggest that leveraging visual modality as a meaning representation provides a promising direction for robust natural language understanding.

cs.CL↗

arXiv2025

Grounded Misunderstandings in Asymmetric Dialogue: A Perspectivist Annotation Scheme for MapTask

Collaborative dialogue relies on participants incrementally establishing common ground, yet in asymmetric settings they may believe they agree while referring to different entities. We introduce a perspectivist annotation scheme for the HCRC MapTask corpus (Anderson et al., 1991) that separately captures speaker and addressee grounded interpretations for each reference expression, enabling us to trace how understanding emerges, diverges, and repairs over time. Using a scheme-constrained LLM annotation pipeline, we obtain 13k annotated reference expressions with reliability estimates and analyze the resulting understanding states. The results show that full misunderstandings are rare once lexical variants are unified, but multiplicity discrepancies systematically induce divergences, revealing how apparent grounding can mask referential misalignment. Our framework provides both a resource and an analytic lens for studying grounded misunderstanding and for evaluating (V)LLMs' capacity to model perspective-dependent grounding in collaborative dialogue.

cs.CL↗

arXiv2025

Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning

As synthetic data becomes widely used in language model development, understanding its impact on model behavior is crucial. This paper investigates the impact of the diversity of sources of synthetic data on fine-tuned large language models. We focus on three key dimensions: distribution collapse, adversarial robustness, and self-preference bias. Our findings reveal that fine-tuning models on synthetic data from diverse sources can mitigate distribution collapse, preserving the breadth of the output distribution and the diversity of the output text. Furthermore, while both human and synthetic fine-tuning data can remove safeguards, we observe a tendency for higher output quality in the latter case, thus making outputs potentially more usable and dangerous. Finally, we also find evidence that fine-tuning reduces self-preference bias, with human data being the most effective, followed by multi-source synthetic data.

cs.CL↗

arXiv2025

Evaluating LLM-Generated Versus Human-Authored Responses in Role-Play Dialogues

Evaluating large language models (LLMs) in long-form, knowledge-grounded role-play dialogues remains challenging. This study compares LLM-generated and human-authored responses in multi-turn professional training simulations through human evaluation ($N=38$) and automated LLM-as-a-judge assessment. Human evaluation revealed significant degradation in LLM-generated response quality across turns, particularly in naturalness, context maintenance and overall quality, while human-authored responses progressively improved. In line with this finding, participants also indicated a consistent preference for human-authored dialogue. These human judgements were validated by our automated LLM-as-a-judge evaluation, where Gemini 2.0 Flash achieved strong alignment with human evaluators on both zero-shot pairwise preference and stochastic 6-shot construct ratings, confirming the widening quality gap between LLM and human responses over time. Our work contributes a multi-turn benchmark exposing LLM degradation in knowledge-grounded role-play dialogues and provides a validated hybrid evaluation framework to guide the reliable integration of LLMs in training simulations.

cs.CL↗

arXiv2025

Do LLMs exhibit the same commonsense capabilities across languages?

This paper explores the multilingual commonsense generation abilities of Large Language Models (LLMs). To facilitate this investigation, we introduce MULTICOM, a novel benchmark that extends the COCOTEROS dataset to four languages: English, Spanish, Dutch, and Valencian. The task involves generating a commonsensical sentence that includes a given triplet of words. We evaluate a range of open-source LLMs, including LLaMA, Qwen, Gemma, EuroLLM, and Salamandra, on this benchmark. Our evaluation combines automatic metrics, LLM-as-a-judge approaches (using Prometheus and JudgeLM), and human annotations. Results consistently show superior performance in English, with significantly lower performance in less-resourced languages. While contextual support yields mixed results, it tends to benefit underrepresented languages. These findings underscore the current limitations of LLMs in multilingual commonsense generation. The dataset is publicly available at https://huggingface.co/datasets/gplsi/MULTICOM.

cs.CL↗

arXiv2025

Correlates of Image Memorability in Vision Encoders: Activations, Attention Entropy, Patch Uniformity and Autoencoder Losses

Images vary in how memorable they are to humans. Inspired by findings from cognitive science and computer vision, we explore correlates of image memorability in pretrained transformer-based vision encoders for the first time. Focusing initially on activations, attention distributions, and the uniformity of image patches, we find that these features correlate with memorability to some extent. Additionally, we explore sparse autoencoder loss over the representations of vision encoders as a proxy for memorability, which yields results outperforming past methods using convolutional neural network representations. Our results shed light on the relationship between model-internal features and memorability. They show that some features are informative predictors of what makes images memorable to humans; revealing that, in particular, the reconstruction loss from our autoencoders is a strong correlate of image memorability.

cs.CV↗

arXiv2025

From Image Captioning to Visual Storytelling

Visual Storytelling is a challenging multimodal task between Vision & Language, where the purpose is to generate a story for a stream of images. Its difficulty lies on the fact that the story should be both grounded to the image sequence but also narrative and coherent. The aim of this work is to balance between these aspects, by treating Visual Storytelling as a superset of Image Captioning, an approach quite different compared to most of prior relevant studies. This means that we firstly employ a vision-to-language model for obtaining captions of the input images, and then, these captions are transformed into coherent narratives using language-to-language methods. Our multifarious evaluation shows that integrating captioning and storytelling under a unified framework, has a positive impact on the quality of the produced stories. In addition, compared to numerous previous studies, this approach accelerates training time and makes our framework readily reusable and reproducible by anyone interested. Lastly, we propose a new metric/tool, named ideality, that can be used to simulate how far some results are from an oracle model, and we apply it to emulate human-likeness in visual storytelling.

cs.CL↗

arXiv2025

Disentangling the Roles of Representation and Selection in Data Pruning

Data pruning, selecting small but impactful subsets, offers a promising way to efficiently scale NLP model training. However, existing methods often involve many different design choices, which have not been systematically studied. This limits future developments. In this work, we decompose data pruning into two key components: the data representation and the selection algorithm, and we systematically analyze their influence on the selection of instances. Our theoretical and empirical results highlight the crucial role of representations: better representations, e.g., training gradients, generally lead to a better selection of instances, regardless of the chosen selection algorithm. Furthermore, different selection algorithms excel in different settings, and none consistently outperforms the others. Moreover, the selection algorithms do not always align with their intended objectives: for example, algorithms designed for the same objective can select drastically different instances, highlighting the need for careful evaluation.

cs.CL↗

arXiv2025

Common Objects Out of Context (COOCo): Investigating Multimodal Context and Semantic Scene Violations in Referential Communication

To what degree and under what conditions do VLMs rely on scene context when generating references to objects? To address this question, we introduce the $\textit{Common Objects Out-of-Context (COOCo)}$ dataset and conduct experiments on several VLMs under different degrees of scene-object congruency and noise. We find that models leverage scene context adaptively, depending on scene-object semantic relatedness and noise level. Based on these consistent trends across models, we turn to the question of how VLM attention patterns change as a function of target-scene semantic fit, and to what degree these patterns are predictive of categorisation accuracy. We find that successful object categorisation is associated with increased mid-layer attention to the target. We also find a non-monotonic dependency on semantic fit, with attention dropping at moderate fit and increasing for both low and high fit. These results suggest that VLMs dynamically balance local and contextual information for reference generation. Dataset and code are available here: $\href{https://github.com/cs-nlp-uu/scenereg}{https://github.com/cs-nlp-uu/scenereg}$.

cs.CV↗

arXiv2025

References Matter: Investigating the Impact of Reference Set Variation on Summarization Evaluation

Human language production exhibits remarkable richness and variation, reflecting diverse communication styles and intents. However, this variation is often overlooked in summarization evaluation. While having multiple reference summaries is known to improve correlation with human judgments, the impact of the reference set on reference-based metrics has not been systematically investigated. This work examines the sensitivity of widely used reference-based metrics in relation to the choice of reference sets, analyzing three diverse multi-reference summarization datasets: SummEval, GUMSum, and DUC2004. We demonstrate that many popular metrics exhibit significant instability. This instability is particularly concerning for n-gram-based metrics like ROUGE, where model rankings vary depending on the reference sets, undermining the reliability of model comparisons. We also collect human judgments on LLM outputs for genre-diverse data and examine their correlation with metrics to supplement existing findings beyond newswire summaries, finding weak-to-no correlation. Taken together, we recommend incorporating reference set variation into summarization evaluation to enhance consistency alongside correlation with human judgments, especially when evaluating LLMs.

cs.CL↗

arXiv2025

Burn After Reading: Do Multimodal Large Language Models Truly Capture Order of Events in Image Sequences?

This paper introduces the TempVS benchmark, which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models (MLLMs) in image sequences. TempVS consists of three main tests (i.e., event relation inference, sentence ordering and image ordering), each accompanied with a basic grounding test. TempVS requires MLLMs to rely on both visual and linguistic modalities to understand the temporal order of events. We evaluate 38 state-of-the-art MLLMs, demonstrating that models struggle to solve TempVS, with a substantial performance gap compared to human capabilities. We also provide fine-grained insights that suggest promising directions for future research. Our TempVS benchmark data and code are available at https://github.com/yjsong22/TempVS.

cs.CL↗

arXiv2025

VAQUUM: Are Vague Quantifiers Grounded in Visual Data?

Vague quantifiers such as "a few" and "many" are influenced by various contextual factors, including the number of objects present in a given context. In this work, we evaluate the extent to which vision-and-language models (VLMs) are compatible with humans when producing or judging the appropriateness of vague quantifiers in visual contexts. We release a novel dataset, VAQUUM, containing 20,300 human ratings on quantified statements across a total of 1089 images. Using this dataset, we compare human judgments and VLM predictions using three different evaluation methods. Our findings show that VLMs, like humans, are influenced by object counts in vague quantifier use. However, we find significant inconsistencies across models in different evaluation settings, suggesting that judging and producing vague quantifiers rely on two different processes.

cs.CL↗

arXiv2024

Probing Omissions and Distortions in Transformer-based RDF-to-Text Models

In Natural Language Generation (NLG), important information is sometimes omitted in the output text. To better understand and analyse how this type of mistake arises, we focus on RDF-to-Text generation and explore two methods of probing omissions in the encoder output of BART (Lewis et al, 2020) and of T5 (Raffel et al, 2019): (i) a novel parameter-free probing method based on the computation of cosine similarity between embeddings of RDF graphs and of RDF graphs in which we removed some entities and (ii) a parametric probe which performs binary classification on the encoder embeddings to detect omitted entities. We also extend our analysis to distorted entities, i.e. entities that are not fully correctly mentioned in the generated text (e.g. misspelling of entity, wrong units of measurement). We found that both omitted and distorted entities can be probed in the encoder's output embeddings. This suggests that the encoder emits a weaker signal for these entities and therefore is responsible for some loss of information. This also shows that probing methods can be used to detect mistakes in the output of NLG models.

cs.CL↗