Generative agents are rapidly advancing in sophistication, raising urgent questions about how they might coordinate when deployed in online ecosystems. This is particularly consequential in information operations (IOs), influence campaigns that aim to manipulate public opinion on social media. While traditional IOs have been orchestrated by human operators and relied on manually crafted tactics, agentic AI promises to make campaigns more automated, adaptive, and difficult to detect. This work presents the first systematic study of emergent coordination among generative agents in simulated IO campaigns. Using generative agent-based modeling, we instantiate IO and organic agents in a simulated environment and evaluate coordination across operational regimes, from simple goal alignment to team knowledge and collective decision-making. As operational regimes become more structured, IO networks become denser and more clustered, interactions more reciprocal and positive, narratives more homogeneous, amplification more synchronized, and hashtag adoption faster and more sustained. Remarkably, simply revealing to agents which other agents share their goals can produce coordination levels nearly equivalent to those achieved through explicit deliberation and collective voting. Overall, we show that generative agents, even without human guidance, can reproduce coordination strategies characteristic of real-world IOs, underscoring the societal risks posed by increasingly automated, self-organizing IOs.

Cyber-Physical Systems (CPS) drive modern critical infrastructures by tightly integrating physical processes with computation and communication networks. This convergence exposes CPS to sophisticated cyber threats propagating across physical, control, and network layers, where stealthy and multi-stage attacks manifest through weak and distributed signals, challenging centralized intrusion detection systems and monolithic AI models that struggle to scale across heterogeneous subsystems and lack the interpretability required in safety-critical environments. This paper proposes a hierarchical multi-agent AI framework for automated cybersecurity assessment with human-understandable explanations in networked CPS. The framework dynamically instantiates a task-specific agent hierarchy from a natural language description of the target system, aligning the security analysis process with the underlying CPS architecture. Specialized agents perform fine-grained analysis of heterogeneous data sources, subsystem supervisors aggregate and contextualize local findings, and a deliberative round-table consensus mechanism enables cross-subsystem correlation for detecting coordinated and stealthy attacks. The framework is evaluated using four LLMs (Qwen 3 4B, Qwen 3 8B, Llama 3.1 8B, and Llama 3.3 70B) across three datasets (PicoDomain, CERT r5.2, and SWaT), achieving perfect recall for models with 8B parameters and above (i.e., zero missed attacks), while reaching an accuracy of 90.18% on PicoDomain, 92.77% on CERT r5.2, and 96.97% on SWaT. An ablation study confirms the effectiveness of the cross-subsystem consensus mechanism, demonstrating substantial precision gains when collaborative deliberation is enabled. Overall, this work establishes hierarchical multi-agent architectures as a scalable, interpretable, and structure-aware foundation for AI-driven cybersecurity in networked CPS.

Vision-Language Models (VLMs) excel in multimodal tasks, yet their effectiveness in specialized medical applications remains underexplored. Accurate interpretation of medical images and text is crucial for clinical decision support, particularly in multiple-choice question answering (MCQA). To address the lack of benchmarks in this domain, we introduce the Multilingual Multimodal Medical Exam Dataset (MMMED), designed to assess VLMs' ability to integrate visual and textual information for medical reasoning. MMMED includes 582 MCQA pairs from Spanish medical residency exams (MIR), with multilingual support (Spanish, English, Italian) and paired medical images. We benchmark state-of-theart VLMs, analyzing their strengths and limitations across languages and modalities. The dataset is publicly available on Hugging Face (https://huggingface.co/datasets/praiselab-picuslab/MMMED), with experimental code on GitHub (https://github.com/PRAISELab-PicusLab/MMMED).

Artificial Intelligence (AI) systems are transforming critical sectors such as healthcare, finance, and transportation, enhancing operational efficiency and decision-making processes. However, their deployment in high-stakes domains has exposed vulnerabilities that can result in significant societal harm. To systematically study and mitigate these risk, initiatives like the AI Incident Database (AIID) have emerged, cataloging over 3,000 real-world AI failure reports. Currently, associating a new report with the appropriate AI Incident relies on manual expert intervention, limiting scalability and delaying the identification of emerging failure patterns. To address this limitation, we propose a retrieval-based framework that automates the association of new reports with existing AI Incidents through semantic similarity modeling. We formalize the task as a ranking problem, where each report—comprising a title and a full textual description—is compared to previously documented AI Incidents based on embedding cosine similarity. Benchmarking traditional lexical methods, cross-encoder architectures, and transformer-based sentence embedding models, we find that the latter consistently achieve superior performance. Our analysis further shows that combining titles and descriptions yields substantial improvements in ranking accuracy compared to using titles alone. Moreover, retrieval performance remains stable across variations in description length, highlighting the robustness of the framework. Finally, we find that retrieval performance consistently improves as the training set expands. Our approach provides a scalable and efficient solution for supporting the maintenance of the AIID.

Harmful meme detection poses a critical challenge for online moderation, as the multimodal and context-dependent nature of memes undermines the effectiveness of traditional unimodal classifiers. In this work, we propose an agent-driven architecture that combines multimodal decomposition with multiagent reasoning. The system extracts complementary information from memes through text extraction, image captioning, and contextual visual description, which are integrated into a unified textual representation. This representation is then analyzed by a set of LLM-powered agents, each instantiated with a distinct interpretative persona, whose assessments are consolidated by a decision aggregation module. Experimental results on the Facebook Hateful Memes dataset demonstrate that the multiagent approach significantly improves performance: accuracy increases from$5 4. 2 5 \%$(single-agent baseline) to$6 7. 2 8 \%$, while the true positive rate rises from 40.32% to 72.78%. These findings highlight the effectiveness of integrating multimodal decomposition with agent-based perspectives for harmful meme detection, ensuring both robustness and explainability in the decision-making process.

Geopolitical news provides vast amounts of information essential for understanding international relations and political events. However, organizing this information into a coherent, structured format poses challenges due to the complexity and dynamic nature of the domain. This paper introduces a scalable system leveraging Large Language Models to build continuously updated Knowledge Graphs from Italian geopolitical news. The system features a modular architecture, including a Collector Node for scalable article extraction, a Redis-based reliable queue to manage large-scale data ingestion, and a Named Entity Recognition/Relation Extraction Engine to standardize entity-relation triples. The framework addresses key challenges, such as continuous updating and hallucination mitigation, ensuring the reliability of the graph. Our evaluations demonstrate significant improvements in scalability, uniformity of extracted triples, and graph accuracy, making this architecture particularly suitable for real-time geopolitical analysis.

Legal Artificial Intelligence has emerged as an essential field, focusing on AI technologies that facilitate various legal tasks and alleviate the workload of legal professionals. Despite advancements in Legal Artificial Intelligence, there remains a critical gap in systems that can provide both comprehensive and contextually accurate retrieval tailored to the intricate structure of EU legislation. We propose EuropeanLawAdvisor, an efficient and user-friendly legal information retrieval system designed to deliver tailored responses to legal queries. This system utilizes open-source Large Language Models within a Retrieval-Augmented Generation framework, facilitating precise and relevant information retrieval. The system employs a robust retrieval approach that integrates multi-match, k-nearest neighbors, hybrid methods, and TF-IDF search strategies across both complete documents and segmented text indexes, ensuring comprehensive retrieval for diverse query types. The implementation of the framework has demonstrated significant improvements in the accuracy and relevance of responses to EU legal queries, enhancing both the retrieval of relevant legal documents and the generation of precise responses. We show that EuropeanLawAdvisor, leveraging open-source models like Phi3-mini-3B and LLaMa-3-8B, achieves competitive Faithfulness and Relevance compared to GPT-4-Turbo. The performance gap narrows significantly in zero-shot scenarios, and our approach outperforms GPT-4-Turbo in the percentage of answered questions. We publicly release our code on GitHub: https://github.com/raffaele-russo/EuropeanLawAdvisor.