In 2026, we held the fourth iteration of the SHROOM Shared Task series: SHROOM-Visions (Shared-task on Hallucinations and Related Observable Overgeneration Mistakes in Vision language models), which is hosted at the UncertaiNLP Workshop co-located with EMNLP 2026. Following the success of the 2024 and 2025 tasks, this time we aim to tackle hallucinations through a model-agnostic detection task focused on large vision-language models. Building on the recently introduced SHEEP dataset, designed for long-term evaluation across model generations, the task invites participants to detect and classify fine-grained hallucination spans in image-conditioned text generation (VQA, image captioning, etc.). The evaluation uses a five-class taxonomy of hallucinations spanning four languages: Chinese, English, French, and Italian. The shared task generated strong interest in the NLP community worldwide, with 27 teams contributing 600+ system submissions. The best systems achieve average scores of 0.58 in character-level correlation, 0.46 in label-conditioned correlation, and 0.51 in intersection-over-union (IoU) across four languages, outperforming the baselines by 30–40 points.

While large language models (LLMs) excel at semantic reasoning, their discrete token-based outputs introduce limitations for fine-grained regression tasks requiring continuous scoring. We address graded word-sense plausibility estimation by reformulating it as a Natural Language Inference (NLI) regression problem, adapting DeBERTa-v3-large with NLI pretraining and a regression head to predict continuous plausibility scores from story-sense pairs. We compare this model against BERT, vanilla DeBERTa, SmolLM variants and state-of-the art LLMs under various prompting strategies, and show that the NLI-finetuned model achieves superior rank correlation and alignment with human judgments. While several baselines collapse toward mean predictions and LLMs show unstable prompting sensitivity, our findings establish NLI-informed pretraining as highly effective for narrative plausibility regression, highlighting fundamental LLM limitations for word sense disambiguation.

In an age of rapid model turnover, how do we make hallucination evaluation more perennial? We explore whether human-written hallucination samples could take the place of model-generated hallucinations, in order to make benchmarking detection independent of particular models. To this end, we construct SHEEP, a dataset of 1,600 human-written samples, spanning four languages (Chinese, English, French, Italian), and 18,400 samples from five vision-and-language models, all annotated for hallucinations using a fine-grained span-level labeling scheme. We find that human-written samples result in higher agreement and allow greater control of dataset contents, while remaining distributionally similar to samples derived from vision-and-language samples and providing a reasonable portrayal of detection capabilities — suggesting that human data is a viable substitute for model-based hallucination benchmarks.

Idiomatic expressions present significant challenges for natural language understanding systems as their meaning often diverge from the literal interpretation. While prior works have focused on textual idiom detection, the role of visual content in reasoning about idiomaticity remains underexplored. This study introduces a Chain-of-Thought reasoning framework that enhances idiomatic comprehension by ranking images based on their relevance to a compound expression in context, requiring the system to distinguish between idiomatic and literal meanings.We comprehensively evaluate our approach by quantitatively analyzing the performance improvements achieved integrating textual and visual information in the ranking process through different prompting settings. Our empirical findings provide insights into the capabilities of visual Large Language Models to establish meaningful correlations between idiomatic content and its visual counterpart, suggesting promising directions for multimodal language understanding.

Video summarization aims to generate a condensed textual version of an original video. Summaries may consist of either plain text or a shortlist of salient events, possibly including temporal or spatial references. Video Large Language Models (VLLMs) exhibit impressive zero-shot capabilities in video analysis. However, their performance varies significantly according to the LLM prompt, the characteristics of the video, and the properties of the training data and LLM architecture.In this work, we thoroughly evaluate the zero-shot summarization performance of four state-of-the-art open-source VLLMs specifically designed to address spatial and temporal reasoning. In light of the detected summarization issues, we propose different cost-effective mitigation strategies, based on Chain-of-Thought prompting, that involve the injection of knowledge extracted by external, lightweight models. To perform the VLLM evaluation, we design a new video summarization benchmark consisting of 100 videos with varying characteristics in terms of domain, duration, and spatio-temporal properties. Videos are manually annotated by three independent human experts with plain text, event-based, and spatio-temporal summaries. The experimental evaluation shows that VLLMs significantly benefit from prompting a list of recognized actions, whereas injecting automatically recognized objects and scene changes respectively improve spatially contextualized and event-based summaries in specific cases.

Large language models (LLMs) have recently obtained strong performance on complex reasoning tasks. However, their capabilities in specialized domains like law remain relatively unexplored. We present CLUEDO, a system to tackle a novel legal reasoning task that involves determining if a provided answer correctly addresses a legal question derived from U.S. civil procedure cases. CLUEDO utilizes multiple collaborator models that are trained using multiple-choice prompting to choose the right label and generate explanations. These collaborators are overseen by a final “detective” model that identifies the most accurate answer in a zero-shot manner. Our approach achieves an F1 macro score of 0.74 on the development set and 0.76 on the test set, outperforming individual models. Unlike the powerful GPT-4, CLUEDO provides more stable predictions thanks to the ensemble approach. Our results showcase the promise of tailored frameworks to enhance legal reasoning capabilities in LLMs.

In the banking and finance sectors, members of the business units focused on Trend and Risk Analysis daily process internal and external visually-rich documents including text, images, and tables. Given a facet (i.e., topic) of interest, they are particularly interested in retrieving the top trending keywords related to it and then use them to annotate the most relevant document elements (e.g., text paragraphs, images or tables). In this paper, we explore the use of both open-source and proprietary Large Language Models to automatically generate lists of facet-relevant keywords, automatically produce free-text descriptions of both keywords and multimedia document content, and then annotate documents by leveraging textual similarity approaches. The preliminary results, achieved on English and Italian documents, show that OpenAI GPT-4 achieves superior performance in keyword description generation and multimedia content annotation, while the open-source Meta AI Llama2 model turns out to be highly competitive in generating additional keywords.

This paper presents a groundbreaking multimodal, multi-task, multi-teacher joint-grained knowledge distillation model for visually-rich form document understanding. The model is designed to leverage insights from both fine-grained and coarse-grained levels by facilitating a nuanced correlation between token and entity representations, addressing the complexities inherent in form documents. Additionally, we introduce new inter-grained and cross-grained loss functions to further refine diverse multi-teacher knowledge distillation transfer process, presenting distribution gaps and a harmonised understanding of form documents. Through a comprehensive evaluation across publicly available form document understanding datasets, our proposed model consistently outperforms existing baselines, showcasing its efficacy in handling the intricate structures and content of visually complex form documents.

Visual-Word Sense Disambiguation (V-WSD) entails resolving the linguistic ambiguity in a text by selecting a clarifying image from a set of (potentially misleading) candidates. In this paper, we address V-WSD using a state-of-the-art Image-Text Retrieval system, namely CLIP. We propose to alleviate the linguistic ambiguity across multiple domains and languages via text and image augmentation. To augment the textual content we rely on back-translation with the aid of a variety of auxiliary languages. The approach based on finetuning CLIP on the full phrases is effective in accurately disambiguating words and incorporating back-translation enhance the system’s robustness and performance on the test samples written in Indo-European languages.

The use of Natural Language Processing techniques in the legal domain has become established for supporting attorneys and domain experts in content retrieval and decision-making. However, understanding the legal text poses relevant challenges in the recognition of domain-specific entities and the adaptation and explanation of predictive models. This paper addresses the Legal Entity Name Recognition (L-NER) and Court judgment Prediction (CPJ) and Explanation (CJPE) tasks. The L-NER solution explores the use of various transformer-based models, including an entity-aware method attending domain-specific entities. The CJPE proposed method relies on hierarchical BERT-based classifiers combined with local input attribution explainers. We propose a broad comparison of eXplainable AI methodologies along with a novel approach based on NER. For the L-NER task, the experimental results remark on the importance of domain-specific pre-training. For CJP our lightweight solution shows performance in line with existing approaches, and our NER-boosted explanations show promising CJPE results in terms of the conciseness of the prediction explanations.

Gender discrimination is a serious and widespread problem on social media and online in general. Besides offensive messages, memes are one of the main means of dissemination for such content. With these premises, the MAMI task was proposed at the SemEval-2022, which consists of identifying memes with misogynous characteristics. In this work, we propose a solution to this problem based on Mask R-CNN and VisualBERT that leverages the multimodal nature of the task. Our study focuses on observing how the two sources of data in memes (text and image) and their possible combinations impact performances. Our best result slightly exceeds the higher baseline, but the experiments allowed us to draw important considerations regarding the importance of correctly exploiting the visual information and the relevance of the elements present in the memes images.