Most existing approaches to AI-Generated Text Detection (AIGTD) treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. However, this perspective overlooks the inherently dynamic nature of autoregressive generation, where content evolves progressively through the latent space. In this paper, we reformulate AIGTD as the problem of distinguishing between latent generation trajectories. Instead of relying on static representations, we model how textual representations evolve across the sequence. To this end, we propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation. GTCL then applies contrastive learning to these trajectories to learn geometric regularities associated with the autoregressive generation. Evaluations performed on three different benchmarks and several approaches show that GTCL outperforms detection baselines consistently, which implies that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains. These results suggest that modeling trajectory differences could improve detection and open up a dynamic direction that has been underexplored in previous AIGTD literature.

Retrieval-Augmented Generation (RAG) mitigates hallucinations in Large Language Models (LLMs) by grounding the generation process on external knowledge. However, standard RAG approaches struggle with multi-hop reasoning. While recent graph-based RAG methods improve the retrieval of interconnected chunks, they often rely on computationally expensive and error-prone LLM-based extraction pipelines. To address these issues, we propose TIGRAG (Token-Induced GraphRAG), an efficient graph-augmented RAG framework based on a token co-occurrence Knowledge Graph. TIGRAG directly models topological relationships between tokens using sliding-window co-occurrence statistics, thus enabling scalable graph construction. During inference, it combines graph-based semantic expansion and neural reranking to retrieve interconnected evidence for multi-hop reasoning. Specifically, it introduces an iterative entity-driven retrieval strategy that progressively expands the query using bridging entities extracted from previously retrieved contexts. We evaluated TIGRAG on three widely adopted multi-hop Question Answering (QA) benchmarks. Experimental results demonstrated that our framework consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks, while substantially reducing indexing time, inference latency, and prompt footprint.

Connectome analysis investigates the connections in the brain to understand how brain regions communicate with each other and how brain structure relates to its function. In recent years, researchers have reconstructed the structural connectome of several organisms, the most complex being Drosophila melanogaster. Two research groups have reconstructed the larval and adult connectomes of this organism and have applied network analysis to learn more about the Drosophila brain and its behavior. In this paper, we aim to continue the work of these two research groups at the larval and adult stages. Specifically, we construct several derived network representations and define a set of techniques that use the main concepts and measures of complex network analysis to extract new knowledge about Drosophila connectomes at the larval and adult stages. First, we conduct an Exploratory Data Analysis on the larval and adult connectomes to detect similarities and differences between them. Then, we define the concept of power neurons and illustrate an approach to detect them. Next, we demonstrate that power neurons represent a limited set of highly interconnected neurons that form a backbone and that, given their peculiar connectivity properties, may play a strategic role in brain functions. Finally, we extract a set of connectome motifs that allow us to learn about various features characterizing power neurons. We demonstrate that complex network analysis can allow the extraction of relevant knowledge about connectomes. Furthermore, we show that a very small number of power neurons can strongly influence all other neurons in the Drosophila brain.

In this paper, we propose Perturbation Investigation Framework for Multilayer Networks (PIF-MN), a framework for analyzing perturbations on multilayer networks. A perturbation is a generic action, directed toward a target, whose effect is to alter the state of the multilayer network, i.e., the set of its nodes, intralayer and interlayer arcs. The study of perturbations in multilayer networks is important because these networks are increasingly used to model heterogeneous and often critical scenarios where any perturbation can have dramatic cascading effects. PIF-MN considers a wide range of perturbation strategies, both single and combined. It also proposes several structural and connectivity measures allowing for the evaluation of the effects of each type of perturbation on a multilayer network and the resilience of this network to them. After introducing PIF-MN, we describe a set of experiments for testing it on both synthetic multilayer networks, built by applying two distinct random models, namely Erd & odblac;s-R & eacute;nyi and Barab & aacute;si-Albert, and a real-world multilayer network. The results of these experiments highlight the robustness of multilayer networks against potential perturbations.

Due to their outstanding performance, Vision Transformers (ViTs) are becoming one of the most widely used architectures for computer vision tasks. However, the interpretation of the output returned by a ViT is still a challenging problem, and yet its solution is crucial for users to gain confidence in these new architectures. In this setting, some methods to address this issue rely only on attention scores; they cannot offer any causal guarantee because the highlighted patches do not necessarily drive the prediction. Other methods use mask-based perturbations; they are computationally expensive because they must repeatedly test many mask combinations of the image and rerun the model to measure confidence changes. To address these limitations, we present SImilarity-based GRAphs for Vision Transformer Explainability (SIGRATE), a hybrid framework that combines attention cues with graph-based mask generation to leverage the strengths of both approaches and overcome their weaknesses. For each attention layer of a ViT, SIGRATE extracts the embeddings of the image patches and constructs the corresponding similarity graph; then, it generates a set of binary masks starting from specific patches and performing walks in the similarity graph. Finally, it merges the masks of all attention layers into a heatmap using the coverage bias formula. We have tested SIGRATE on two ViT models (i.e., ViT-Base and DeiT-Base) and on two datasets, i.e., a subset of the ImageNet validation set and BloodMNIST. Our experiments show that SIGRATE provides promising performance in terms of Insertion, Deletion, and Pointing Game. Finally, we have performed a qualitative analysis demonstrating SIGRATE’s capabilities, as well as a hyperparameter analysis and an ablation study showing the impact of design choices on SIGRATE’s performance and efficiency.

In this article, we present PostGage, a new framework that identifies the most suitable publishers of target content on social media, based on the level of user engagement they are expected to generate. PostGage uses a network-based data model that considers the characteristics of publishers, as well as the degree to which the semantic content and lifespan of the posts to be published overlap with those of the posts already published. Leveraging this data model, PostGage builds a multinomial classifier based on a graph neural network (GNN). The alignment between the graph-based data model and the GNN technology allows a precise classification of publishers based on their potential to generate user engagement for target content. This feature enables a more precise definition of publishing strategies on social media, as it allows the selection of the most promising publishers for a given post. The network-based model, the post ageing mechanism and the adoption of the GNN as a multiclass classifier represent PostGage's key contributions. We also present a series of tests performed on X that demonstrate the high quality of PostGage's results (e.g. 92% precision in predicting high-engagement posts), establishing it as an invaluable tool for various applications, including marketing and social media management.

In this paper, we introduce EgoFine, a new approach for fine-tuning Large Language Models (LLMs) based on ego networks extracted from Knowledge Graphs (KGs). EgoFine first identifies the most informative nodes in the KG using degree centrality. It then extracts their ego networks and generates structured training data through random paths within them, thus enabling LLMs to learn domain-specific knowledge. It further constructs negative samples to explicitly model the ab sence of relationships between entities. We present an experimental campaign involving three KGs (PrimeKG, WN18RR, and YAGO3) and four LLMs (Minerva-350M, Llama3.2-1B, Qwen2-1.5B, and Ministral-3B). This campaign demonstrates that EgoFine outperforms traditional embedding-based methods (e.g., AutoSF, BoxE, NodePiece, PairRE, and TransE), two state-of-the-art approaches integrating KGs and LLMs (e.g., GNN-RAG and KG-Adapter), as well as a baseline approach operating on the same principle as EgoFine but without exploiting the contribution of ego net works. Compared with this last approach, EgoFine improves Hit@1 values by up to 47.37%, Mean Reciprocal Rank (MRR) values by up to 46.87%, F1-Score values by up to 36.59%, and Accuracy values by up to 30.91%. The paper also presents an ablation study devoted to evaluating several design choices underlying EgoFine, as well as an analysis of the EgoFine's behavior when applied to dynamic or noisy KGs. This way of proceeding makes EgoFine particularly beneficial for a vari ety of real-world applications, including understanding complex biological mechanisms, reasoning about the relationships between legislative sources and court cases, interpreting and explaining complex industrial maintenance and production processes, and understanding the connections between attacks, exploits, and countermeasures in the context of cybersecurity.

Since the introduction of GPT-2, Large Language Models (LLMs) have proven to be able to handle various tasks with impressive performance. However, they sometimes generate incorrect output or even hallucinations. To overcome this problem, many researchers have investigated the possibility of integrating external factual knowledge, such as that encoded in Knowledge Graphs (KGs), into LLMs. Although there are many approaches in the existing literature that integrate KGs and LLMs in different ways, few of them use KGs to fine-tune LLMs, and none of them systematically use KG substructures. In this paper, we propose CoFine (Community-Based Fine-Tuner), an approach to fine-tune an LLM using the communities of a KG. CoFine works as follows: it first divides the KG into communities, each of which contains a homogeneous portion of the knowledge expressed by the KG. It then uses these communities to fine-tune the LLM. This way of proceeding allows LLM fine-tuning to focus on specific homogeneous information contained in the KG expressed by each community. CoFine allows the LLM to achieve a very high accuracy in knowledge completion tasks. This is evidenced by comparisons between CoFine and a baseline LLM fine-tuning approach, which showed that our approach achieves better results for all metrics considered with several KG.

Vision Transformers (ViTs) are difficult to interpret because current methods of relevance propagation and attention flow do not fully consider some key architectural features, such as the uneven importance of attention heads and residual connections. Prior approaches typically assume uniform importance across attention heads; furthermore, they model skip connections as identity paths, leading to inaccurate relevance attribution. To address these issues, we introduce GradSkip, a novel relevance propagation method for ViTs based on adaptive head weighting and skip-aware propagation. GradSkip models the different importance of the attention heads and dynamically distributes relevance between the attention and residual paths. Experiments on ImageNet1K and BloodMNIST demonstrate a state-of-the-art faithfulness of GradSkip while requiring over 14 times fewer GFLOPs than the best-performing existing approaches. Additional evaluations using transformer-based segmentation confirm improved localization and alignment with ground-truth regions.

Vision Transformers (ViTs) have demonstrated outstanding performance across different computer vision tasks thanks to their self-attention mechanism that captures long-range dependencies effectively. However, the inherent complexity of ViTs presents significant challenges in explaining their outputs, which is fundamental in safety-critical domains. To tackle the challenge of explaining ViT outputs, this paper presents Grad-Mask, a novel method that integrates gradients into the mask generation process to create explanation heatmaps. GradMask uses the query, key, and value matrices from each attention layer and computes their gradients with respect to a target class. Afterward, it uses these gradients to generate binary masks, which are then weighted by the corresponding ViT’s confidence scores. Finally, it combines the weighted masks to generate the resulting heatmap. Experimental evaluations on an ImageNet subset with ViT and DeiT (Data-efficient Image Transformer) architectures show that GradMask achieves competitive performance according to standard explainability metrics, such as Insertion, Deletion, and Pointing Game. A hyperparameter analysis confirms the high computational efficiency of GradMask, while an ablation study highlights the importance of combining gradients and masks for the generation of the explanation heatmap. Finally, a qualitative analysis shows the improved explainability of GradMask compared to existing methods, making it a promising approach for understanding ViTs.

Vision Transformers (ViTs) have gained recognition in computer vision due to their outstanding performance. Despite their success, the explainability of ViT outputs is still a challenging issue. To address it, we propose a novel explainability method that leverages image patch embeddings from each attention layer of a ViT to construct similarity graphs. The latter are used to generate binary masks by exploring paths starting from specific patches. The masks from all layers are then aggregated into a comprehensive heatmap using the coverage bias formula. We tested our method on two Vision Transformer architectures (ViT-Base and DeiT-Base) and a subset of the ImageNet validation set. Using Insertion and Deletion metrics, we demonstrate the effectiveness of our proposed method compared to similar ones in the literature. Finally, we include a qualitative analysis that shows the capabilities of our method to make ViTs more interpretable.

This paper introduces a neuro-symbolic framework designed to predict and explain subsequent facts from current observations. Facts are generated through causal relationships, which can be modeled by a set of propositional logic rules representing the domain knowledge. However, these rules remain unknown to the agent. By observing the facts, the agent constructs an approximation of them, which is then used to predict and explain new facts. The proposed framework can learn and adapt to different environments modeled by various forms of logic programs, also handling negation and recursion. Most notably, it can handle dynamic environments whose structure evolves over time. In these scenarios, the agent modifies its understanding of the environment to capture new observations, guaranteeing that its model of the domain knowledge remains up-to-date. To achieve this goal, our approach leverages the A2C (Advantage Actor-Critic) reinforcement learning algorithm. This choice allows us to integrate reinforcement learning principles into our logic framework. Through this research, we aspire to contribute to the development of explainable neuro-symbolic Artificial Intelligence systems in dynamic environments.

In online communities, polarization refers to the phenomenon in which individuals become more divided and extreme in their opinions due to their exposure to specific content. In this paper, we present a network-based framework for evaluating polarization levels in Online Social Networks (OSNs). Starting from a dataset of comments, our framework creates a network of user interactions and leverages the Louvain algorithm, the Rao’s Quadratic Entropy, and ego networks to assess the polarization level of communities and the most influential users. To test our framework, we leveraged a dataset of tweets about climate change. After performing Extraction, Transformation and Loading activities on the dataset, we evaluated its labels, identified communities, and analyzed their polarization level and that of the most influential users. We also analyzed the ego networks of believers and deniers and the aggressiveness of the corresponding tweets. Our analysis revealed the existence of polarized communities and homophily among the most influential users. It also showed that the type of communication used to disseminate information influences the polarization level of both communities and individual users. These results demonstrate our framework’s ability to support the polarization analysis in OSNs.

In this paper, we propose Multilayer network-based Visual Interpreter (MuVI), a framework for visual interpretability of Convolutional Neural Networks (CNNs) based on their mapping into multilayer networks. The peculiarity of MuVI is that it constructs a pixel-level heatmap of the salient parts of an image processed by a CNN, where the importance of each pixel depends on all layers of the CNN and not only on the final ones, as in the existing approaches in the literature. MuVI first maps the CNN into a multilayer network. It then uses this representation to identify the parts of the CNN that most influence the prediction results by extracting those paths within the multilayer network whose nodes correspond to the most active areas of the feature maps. The weight of the paths is given by the sum of the weights of the arcs corresponding to the activations across all feature maps of the CNN; this characteristic allows MuVI to consider all layers of the CNN, not just the last ones. Finally, MuVI constructs the visual interpretability heatmap by selecting the paths with the highest weights. The experimental tests performed show that MuVI is able to achieve very satisfactory results in terms of AUC insertion (0.25), AUC deletion (0.11), % Increase in Confidence (12.32), Average Drop % (51.22), Pointing Game Accuracy (0.28) and Computation time (26.226s). These results, taking all these measures together, are better than those obtained by the classical approaches already proposed in the literature, such as SmoothGrad, Grad-CAM, Grad-CAM++, and RISE. They are also comparable to state-of-the-art approaches in the literature, such as Score-CAM and HSIC.

Threads is a new social network that was launched by Meta in July 2023 and conceived as a direct alternative to X. It is a unique case study in the social network landscape, as it is content-based like X, but has an Instagram-based growth model, which makes it significantly different from X. As it was launched recently, studies on Threads are still scarce. One of the most common investigations in social networks regards power users (also called influencers, lead users, influential users, etc.), i.e., those users who can significantly influence information dissemination, user behavior, and ultimately the current dynamics and future development of a social network. In this paper, we want to contribute to the knowledge of Threads by showing that there are indeed power users in this social network and then attempt to understand the main features that characterize them. The definition of power users that we adopt here is novel and leverages the four classical centrality measures of Social Network Analysis. This ensures that our study of power users can benefit from the enormous knowledge on centrality measures that has accumulated in the literature over the years. In order to conduct our analysis, we had to build a Threads dataset, as none existed in the literature that contained the information necessary for our studies. Once we built such a dataset, we decided to make it open and thus available to all researchers who want to perform analyses on Threads. This dataset, the new definition of power users, and the characterization of Threads power users are the main contributions of this paper.

In this paper, we propose Automatic Intent Detector (AID), a framework for automatic intent detection to facilitate the creation of a conversational agent. AID follows an eight-step process incorporating best practices from the current literature and introducing innovative approaches in certain steps. The most notable innovation within AID is the automatic labeling of clusters, which is based on detailed and sophisticated rules derived from linguistics. These rules focus on morphosyntactic analysis, while also taking into account an aspect of semantic role theory. Furthermore, as for the overall validation of the results obtained, it provides an approach based on the concepts of semantic coherence, variability, and label appropriateness. After describing AID at the technical level, we illustrate the experiments we conducted both on a dataset widely used as benchmark in the literature and on a real corporate dataset. Finally, we present a critical discussion on the results obtained.

In recent years, polarization on social media has risen significantly. Social platforms often feature a range of topics that give rise to communities of users with diametrically opposed views, who tend to avoid engaging with others having different perspectives. We call these types of communities “diverging communities”. Examples include communities of supporters and skeptics of climate change or COVID-19 vaccines. In this paper, we aim to investigate this phenomenon. To do so, we first propose a formal definition of discording communities. We then present a framework for investigating the behavior of users of discording communities on a social platform. Our framework is general in that it can be adapted to any social platform where users discuss a topic that polarizes them into communities with diametrically opposed viewpoints rejecting confrontation. Our framework considers not only the structure of communities but also the content of the messages posted by their users. Finally, it can also handle the temporal evolution of the polarization level of both communities and their users. In addition to proposing a formal definition of diverging communities and presenting our framework, we illustrate the results of an extensive experimental campaign carried out on two case studies involving Reddit and X and show how our framework is able to identify a number of features that distinguish the users of one diverging community from the users of the other.