Simone Alghisi

Seyed Mahed Mousavi, Christian Moiola, Massimo Rizzoli et al.

Vision-Language Models (VLMs) are trained on data snapshots of documents, including images and texts. Their training data and evaluation benchmarks are typically static, implicitly treating factual knowledge as time-invariant. However, real-world facts are intrinsically time-sensitive and subject to erratic and periodic changes, causing model predictions to become outdated. We present V-DyKnow, a Visual Dynamic Knowledge benchmark for evaluating time-sensitive factual knowledge in VLMs. Using V-DyKnow, we benchmark closed- and open-source VLMs and analyze a) the reliability (correctness and consistency) of model responses across modalities and input perturbations; b) the efficacy of knowledge editing and multi-modal RAG methods for knowledge updates across modalities; and c) the sources of outdated predictions, through data and mechanistic analysis. Our results show that VLMs frequently output outdated facts, reflecting outdated snapshots used in the (pre-)training phase. Factual reliability degrades from textual to visual stimuli, even when entities are correctly recognized. Besides, existing alignment approaches fail to consistently update the models' knowledge across modalities. Together, these findings highlight fundamental limitations in how current VLMs acquire and update time-sensitive knowledge across modalities. We release the benchmark, code, and evaluation data.

Seyed Mahed Mousavi, Simone Alghisi, Giuseppe Riccardi

LLMs acquire knowledge from massive data snapshots collected at different timestamps. Their knowledge is then commonly evaluated using static benchmarks. However, factual knowledge is generally subject to time-sensitive changes, and static benchmarks cannot address those cases. We present an approach to dynamically evaluate the knowledge in LLMs and their time-sensitiveness against Wikidata, a publicly available up-to-date knowledge graph. We evaluate the time-sensitive knowledge in twenty-four private and open-source LLMs, as well as the effectiveness of four editing methods in updating the outdated facts. Our results show that 1) outdatedness is a critical problem across state-of-the-art LLMs; 2) LLMs output inconsistent answers when prompted with slight variations of the question prompt; and 3) the performance of the state-of-the-art knowledge editing algorithms is very limited, as they can not reduce the cases of outdatedness and output inconsistency.

Simone Alghisi, Massimo Rizzoli, Seyed Mahed Mousavi et al.

Pointing increases the accuracy and explainability of Large Vision-Language Models (LVLMs) by modeling grounding and reasoning as explicit sequential steps. The model grounds the objects mentioned in the natural-language query by predicting their coordinates, and then generates an answer conditioned on these points. While pointing has been shown to increase LVLMs' accuracy, it is unclear which mechanism supports these gains and its relevance in cognitive tasks. In addition, the reliability of the intermediate points remains understudied, limiting their use as visual explanations. In this work, we study the role of pointing in a cognitive task: zero-shot counting from a visual scene. We fine-tune state-of-the-art LVLMs following two approaches: Direct Counting, where models only predict the total number of objects, and Point-then-Count, where LVLMs generate the target objects' coordinates followed by their count. The results show that Point-then-Count achieves higher out-of-distribution generalization, suggesting that coordinates help LVLMs learn skills rather than overfitting on narrow tasks. Although predicted points are accurately grounded in the image in over 89\% of cases (as measured by F1), performance varies across image regions, revealing spatial biases. Finally, mechanistic analyses show that gains in counting arise from the spatial information encoded in the coordinates.

Seyed Mahed Mousavi, Simone Alghisi, Giuseppe Riccardi

LLMs' sources of knowledge are data snapshots containing factual information about entities collected at different timestamps and from different media types (e.g. wikis, social media, etc.). Such unstructured knowledge is subject to change due to updates through time from past to present. Equally important are the inconsistencies and inaccuracies occurring in different information sources. Consequently, the model's knowledge about an entity may be perturbed while training over the sequence of snapshots or at inference time, resulting in inconsistent and inaccurate model performance. In this work, we study the appropriateness of Large Language Models (LLMs) as repositories of factual knowledge. We consider twenty-four state-of-the-art LLMs that are either closed-, partially (weights), or fully (weight and training data) open-source. We evaluate their reliability in responding to time-sensitive factual questions in terms of accuracy and consistency when prompts are perturbed. We further evaluate the effectiveness of state-of-the-art methods to improve LLMs' accuracy and consistency. We then propose "ENtity-Aware Fine-tuning" (ENAF), a soft neurosymbolic approach aimed at providing a structured representation of entities during fine-tuning to improve the model's performance.

Seyed Mahed Mousavi, Simone Alghisi, Giuseppe Riccardi

Continual Pre-Training (CPT) is widely used for acquiring and updating factual knowledge in LLMs. This practice treats loss as a proxy for knowledge learning, while offering no grounding into how it changes during training. We study CPT as a knowledge learning process rather than a solely optimization problem. We construct a controlled, distribution-matched benchmark of factual documents and interleave diagnostic probes directly into the CPT loop, enabling epoch-level measurement of knowledge acquisition dynamics and changes in Out-Of-Domain (OOD) general skills (e.g., math). We further analyze how CPT reshapes knowledge circuits during training. Across three instruction-tuned LLMs and multiple CPT strategies, optimization and learning systematically diverge as loss decreases monotonically while factual learning is unstable and non-monotonic. Acquired facts are rarely consolidated, learning is strongly conditioned on prior exposure, and OOD performance degrades from early epochs. Circuit analysis reveals rapid reconfiguration of knowledge pathways across epochs, providing an explanation for narrow acquisition windows and systematic forgetting. These results show that loss optimization is misaligned with learning progress in CPT and motivate evaluation of stopping criteria based on task-level learning dynamics.

Massimo Rizzoli, Simone Alghisi, Seyed Mahed Mousavi et al.

Fine-tuning Vision-Language Models (VLMs) is a common strategy to improve performance following an ad-hoc data collection and annotation of real-world scenes. However, this process is often prone to biases, errors, and distribution imbalance, resulting in overfitting and imbalanced performance. Although a few studies have tried to address this problem by generating synthetic data, they lacked control over distribution bias and annotation quality. To address these challenges, we redesign the fine-tuning process in two ways. First, we control the generation of data and its annotations, ensuring it is free from bias, distribution imbalance, and annotation errors. We automatically construct the dataset by comprehensively sampling objects' attributes, including color, shape, size, and position within the scene. Secondly, using this annotated dataset, we fine-tune state-of-the-art VLMs and assess performance transferability to real-world data on the absolute position task. We conduct exhaustive evaluations on both synthetic and real-world benchmarks. Our experiments reveal two key findings: 1) fine-tuning on balanced synthetic data yields uniform performance across the visual scene and mitigates common biases; and 2) fine-tuning on synthetic stimuli significantly improves performance on real-world data (COCO), outperforming models fine-tuned in the matched setting.

Seyed Mahed Mousavi, Gabriel Roccabruna, Simone Alghisi et al.

Large Pre-Trained Language Models have demonstrated state-of-the-art performance in different downstream tasks, including dialogue state tracking and end-to-end response generation. Nevertheless, most of the publicly available datasets and benchmarks on task-oriented dialogues focus on written conversations. Consequently, the robustness of the developed models to spoken interactions is unknown. In this work, we have evaluated the performance of LLMs for spoken task-oriented dialogues on the DSTC11 test sets. Due to the lack of proper spoken dialogue datasets, we have automatically transcribed a development set of spoken dialogues with a state-of-the-art ASR engine. We have characterized the ASR-error types and their distributions and simulated these errors in a large dataset of dialogues. We report the intrinsic (perplexity) and extrinsic (human evaluation) performance of fine-tuned GPT-2 and T5 models in two subtasks of response generation and dialogue state tracking, respectively. The results show that LLMs are not robust to spoken noise by default, however, fine-tuning/training such models on a proper dataset of spoken TODs can result in a more robust performance.

Massimo Rizzoli, Simone Alghisi, Olha Khomyn et al.

While Vision-Language Models (VLMs) have achieved competitive performance in various tasks, their comprehension of the underlying structure and semantics of a scene remains understudied. To investigate the understanding of VLMs, we study their capability regarding object properties and relations in a controlled and interpretable manner. To this scope, we introduce CIVET, a novel and extensible framework for systematiC evaluatIon Via controllEd sTimuli. CIVET addresses the lack of standardized systematic evaluation for assessing VLMs' understanding, enabling researchers to test hypotheses with statistical rigor. With CIVET, we evaluate five state-of-the-art VLMs on exhaustive sets of stimuli, free from annotation noise, dataset-specific biases, and uncontrolled scene complexity. Our findings reveal that 1) current VLMs can accurately recognize only a limited set of basic object properties; 2) their performance heavily depends on the position of the object in the scene; 3) they struggle to understand basic relations among objects. Furthermore, a comparative evaluation with human annotators reveals that VLMs still fall short of achieving human-level accuracy.

Simone Alghisi, Gabriel Roccabruna, Massimo Rizzoli et al.

Vision-Language Models (VLMs) have recently gained attention due to their competitive performance on multiple downstream tasks, achieved by following user-input instructions. However, VLMs still exhibit several limitations in visual reasoning, such as difficulties in identifying relations (e.g., spatial, temporal, and among objects), understanding temporal sequences (e.g., frames), and counting objects. In this work, we go beyond score-level benchmark evaluations of VLMs by investigating the underlying causes of their failures and proposing a targeted approach to improve their reasoning capabilities. We study the reasoning skills of seven state-of-the-art VLMs in the counting task under controlled experimental conditions. Our experiments show that VLMs are highly sensitive to the number and type of objects, their spatial arrangement, and the co-occurrence of distractors. A layer-wise analysis reveals that errors are due to incorrect mapping of the last-layer representation into the output space. Our targeted training shows that fine-tuning just the output layer improves accuracy by up to 21%. We corroborate these findings by achieving consistent improvements on real-world datasets.

Simone Alghisi, Massimo Rizzoli, Gabriel Roccabruna et al.

We study the limitations of Large Language Models (LLMs) for the task of response generation in human-machine dialogue. Several techniques have been proposed in the literature for different dialogue types (e.g., Open-Domain). However, the evaluations of these techniques have been limited in terms of base LLMs, dialogue types and evaluation metrics. In this work, we extensively analyze different LLM adaptation techniques when applied to different dialogue types. We have selected two base LLMs, Llama-2 and Mistral, and four dialogue types Open-Domain, Knowledge-Grounded, Task-Oriented, and Question Answering. We evaluate the performance of in-context learning and fine-tuning techniques across datasets selected for each dialogue type. We assess the impact of incorporating external knowledge to ground the generation in both scenarios of Retrieval-Augmented Generation (RAG) and gold knowledge. We adopt consistent evaluation and explainability criteria for automatic metrics and human evaluation protocols. Our analysis shows that there is no universal best-technique for adapting large language models as the efficacy of each technique depends on both the base LLM and the specific type of dialogue. Last but not least, the assessment of the best adaptation technique should include human evaluation to avoid false expectations and outcomes derived from automatic metrics.