Large Language Models (LLMs) have been widely adopted to enhance Task-Oriented Dialogue Systems (TODS) by modeling complex language patterns and delivering contextually appropriate responses. However, this integration introduces significant privacy risks, as LLMs, functioning as soft knowledge bases that compress extensive training data into rich knowledge representations, can inadvertently memorize training dialogue data containing not only identifiable information such as phone numbers but also entire dialogue-level events like complete travel schedules. Despite the critical nature of this privacy concern, how LLM memorization is inherited in developing task bots remains unexplored. In this work, we address this gap through a systematic quantitative study that involves evaluating existing training data extraction attacks, analyzing key characteristics of task-oriented dialogue modeling that render existing methods ineffective, and proposing novel attack techniques tailored for LLM-based TODS that enhance both response sampling and membership inference. Experimental results demonstrate the effectiveness of our proposed data extraction attack. Our method can extract thousands of training labels of dialogue states with best-case precision exceeding 70%. Furthermore, we provide an in-depth analysis of training data memorization in LLM-based TODS by identifying and quantifying key influencing factors and discussing targeted mitigation strategies.

With the emergence of multi-modal knowledge graphs, multi-modal entity alignment is flourishing to compensate for entity knowledge from different data sources. Currently, most methods focus on reducing the noise of the image modality, which usually derives from fuzzy information generated by low-quality data sources. However, few methods consider modality conflict, which can be attributed to the semantic inconsistency between the dominant structure modality and auxiliary modalities. This modality conflict introduces contradictory signals, ultimately degrading overall performance. To tackle these bottlenecks, we propose CALM, a robust multi-modal entity alignment framework against modality conflict, including conflict-aware modality arbitration and multi-modal fusion with masking contrastive learning. We formally define modality conflict and design a metric to quantify conflict strength. Consequently, our proposed modality arbitration mechanism calculates regularized weights to adaptively adjust different modality contributions. Furthermore, our designed masking multi-modal fusion method further captures synergic and unique modal information. Finally, we utilize the large language model to generate entity attributes to assist in extending an existing EA dataset into a multi-modal version to explore the effectiveness of CALM under modality conflict. Experimental results on multiple benchmark datasets and the extended multi-modal data demonstrate the effectiveness and robustness of CALM, which consistently outperforms state-of-the-art methods. Our code is available at https://anonymous.4open.science/r/Anonymization-43F9.

Blockchain data is temporal. Every transaction carries a timestamp and the chain imposes a total order, so on-chain data forms per-source time-series streams. However, existing systems support only basic lookups on blocks and transactions, and cannot answer time-series queries such as time-range retrieval and windowed aggregation. Offloading queries off-chain restores expressiveness, but the off-chain query layer is untrusted, so results must be verifiable. To this end, we propose VeriTS, the first verifiable time-series query framework for blockchain systems. It supports efficient range and aggregation queries without altering blockchain storage structures. VeriTS maintains an off-chain query layer that represents each stream through an authenticated aggregate interval tree. The tree serves as the query index and as the authenticated data structure at once, so a windowed aggregate is answered by folding a logarithmic number of node aggregates. VeriTS verifies completeness through a minimum covering set and soundness through aggregate folding. It extends both guarantees to an approximate path over model segments, redefining completeness and soundness under bounded error. Miners validate a model's residual rather than replay its computation, so even an adversarial encoder can inflate proof size and answer width but never correctness. Experiments offer evidence that on windowed aggregation, VeriTS improves verification efficiency by more than two orders of magnitude over per-record proofs. Range-query proofs shrink by up to 14.5x.

Deliberation plays a crucial role in collaboration; when humans work together, they naturally engage in communication to align information and reach an agreement. In this paper, we investigate deliberative large language model (LLM) agents under partially observable joint decision-making tasks. We formalize deliberative collaboration as a cooperative joint decision problem with partial and asymmetric observations, and introduce a scalable benchmark that instantiates this problem across multiple task settings and domains in which agents must exchange information through deliberation to reach a joint decision with a shared reward. We then instantiate a reference scaffold and evaluation protocol for deliberative agents and conduct a systematic evaluation of a range of representative LLMs. The results reveal that complex deliberative collaboration tasks continue to challenge state-of-the-art language models. Even with the aid of external mathematical tools, language models may fail in either the deliberation process for aligning information or the complex reasoning process for making the decision. On the other hand, diagnostic analysis reveals that the deliberation process may also provide opportunities for reflection and error correction, sometimes improving performance over centralized baselines. Altogether, our work establishes a foundation for evaluating and improving LLM agents in deliberative collaboration and provides insights into the strengths, limitations, and properties of current LLM-based multi-agent systems.

Financial reporting fraud poses a considerable threat to financial market stability. Traditional detection methods predominantly rely on quantitative data analysis while neglecting the informational value of textual disclosures, and existing text analysis approaches suffer from 3 critical limitations: inadequate interpretability, poor domain adaptability, and absence of temporal continuity modeling. This paper proposes the financial statement fraud detection large language model (FSFDLLM) framework, which leverages large language models (LLMs) to achieve in-depth textual feature extraction and interpretable reasoning. Our methodology addresses 3 key challenges: (a) Specifically, we develop a structured prompting mechanism to guide the LLM in conducting the fraud risk assessment of financial texts while generating explanatory rationales. (b) We construct a hybrid learning architecture that integrates LLM-generated pseudo-labels and rationale chains as enhanced features with structured financial data for joint classification model training. (c) The framework has 3 distinctive features: First, it overcomes the domain transfer bottleneck in conventional natural language processing (NLP) models by harnessing the LLM’s semantic comprehension to accurately capture domain-specific financial terminology. Second, it establishes an auditable decision evidence chain through LLM-generated explanatory rationales, thereby enhancing regulatory compliance. Third, it captures continuous fraud patterns by modeling temporal financial indicators and inter-period textual semantic correlations. Experimental evaluations on a real-world listed company dataset demonstrate that our framework considerably outperforms baseline models in key metrics including F1-score and area under the receiver operating characteristic curve (AUC), while the generated explanatory text provides verifiable semantic evidence for regulatory scrutiny. This study presents a cross-modal solution for financial statement fraud detection (FSFD) that harmonizes high accuracy with interpretability, offering practical value for enhancing financial regulatory technology tools.

With the expansion of the capital market, financial fraud incidents have become increasingly frequent. Existing financial fraud detection methods have made some progress, but most approaches overlook at least one critical issue: First, they rely solely on single data sources or static data, failing to effectively capture complex dynamic relationships between companies and collaborative behaviors across time. Second, they struggle to address “disguised behavior” in financial fraud, where fraudulent companies establish relationships with legitimate entities to conceal their actions, leading to data inconsistencies. Third, they lack modeling of temporal dependencies, failing to adequately capture changes in corporate behavior across different time periods. To address these issues, we propose FFD-DHG, a financial fraud detection method based on dynamic heterogeneous graph representation learning. The method uses a fraud-aware graph convolution encoder to capture static topological structure information at each discrete time snapshot and aggregates neighborhood difference information for target nodes to mitigate the inconsistency caused by fraud camouflage while enhancing the distinguishability between normal and fraudulent nodes. Subsequently, we propose a feature fusion module using reinforcement learning, aiming to adaptively adjust the filtering threshold during different node feature aggregation processes to reduce noise interference. To capture temporal dependencies, we introduce a sliding window mechanism and temporal encoding function to model cross-temporal context similarity, thereby more accurately identifying concealed fraudulent behaviors. To validate the practical effectiveness of the proposed financial risk identification method, experiments are conducted on a real-world dataset. Furthermore, the model interpretability analysis of FFD-DHG reveals key features affecting financial fraud detection, providing a decision-making basis for regulatory authorities.

Trajectory similarity computation is a key functionality in many applications, such as abnormal trajectory detection and traffic-mode classification. However, existing methods either incompletely exploit trajectory features or overlook the effects of noisy and critical points, compromising the accuracy of trajectory similarity computation. This paper presents SimCFT, a novel trajectory representation learning framework for accurate trajectory similarity computation. SimCFT consists of two parallel channels. In the dynamic status channel, point-level features (e.g., position, speed, movement distance, movement angle, timestamp) are extracted and fed into a Dynamic Gated Recurrent Unit (DGRU) with an attention mechanism to capture temporal dependencies and dynamic variations. In the spatial context channel, a weighted grid graph is built from historical trajectory data, and Node2Vec is employed to learn the grid embeddings. A Transformer encoder with a noise recognition mechanism (TransNR) is followed to identify and de-emphasize noisy trajectory points. The two channels are fused via a learnable coefficient to produce the final trajectory embeddings. The model is then optimized with the unsupervised InfoNCE contrastive loss, which pulls similar trajectories together while pushing dissimilar ones apart. Experiments on three real-world datasets show that SimCFT significantly outperforms six state-of-the-art methods in similarity accuracy. The source code is available at https://github.com/xianghz123/SIMCFT.git.

Generating discriminative node representations for heterogeneous graphs based on contrastive learning with graph neural networks has become an important topic in data mining. Many existing contrastive learning methods for heterogeneous graphs rely on meta-path information to establish contrastive views and train graph encoders. However, these algorithms have not yet taken full advantage of the wealth of information contained in heterogeneous graphs, encompassing diverse node connections and semantic insights from various meta-paths. Moreover, they do not pay sufficient attention to the encoder adaptation to model properties and the selection of appropriate positive samples between views. To address these problems, we propose a novel Multi-view interactive heterogeneous Graph Contrastive learning (MiCo) model, which leverages positive samples derived from meta-path topology to capture intricate information within a heterogeneous graph effectively. Specifically, we first adopt a multi-view strategy to characterize the various levels of information from a heterogeneous graph, while also selecting suitable encoders tailored to specific views. Then, each unique positive instance is determined according to the topological structure of the target node on each meta-path. Finally, we employ temperature coefficients at different stages to distinguish between the challenging and straightforward negative samples within the negative sample set. Extensive experiments show that MiCo consistently outperforms state-of-the-art baselines on four real-world datasets: ACM, DBLP, AMiner, and IMDB.

Graph alignment, the task of identifying corresponding nodes across different graphs, is crucial for applications ranging from social network analysis to bioinformatics. Although most existing methods leverage graph neural networks (GNNs) to learn node embeddings for attributed graphs and match them based on node similarity, they often rely on objectives designed for node classification or link prediction. These approaches preserve node proximity within individual graphs but fail to capture cross-graph correspondence knowledge, leading to suboptimal alignment performance. This paper presents SGA (Self-boosting Graph Alignment), a novel framework that introduces a self-supervised objective and an iterative refinement scheme for robustly aligning attributed graphs. SGA begins with an unsupervised Graph Convolutional Network (GCN) to learn initial node representations. Then, a contrastive loss is designed to train the GCN using stably aligned node pairs derived from these initial embeddings. This constraint distinguishes closely connected nodes within individual graphs by leveraging cross-graph correspondence knowledge, thereby promoting accurate alignment across graphs. To handle graphs with incomplete topology, SGA incorporates an iterative refinement process enhanced by a neighborhood consistency metric. This metric evaluates the reliability of complementary edges, ensuring that only reliable connections are used to refine the graph structures. The refined graphs are then fed back into the GCN for the subsequent iterations, with the process terminating when no additional pseudo-aligned node pairs are identified or a maximum iteration count is reached. Extensive experiments on four real-world and two synthetic datasets demonstrate that SGA outperforms state-of-the-art graph alignment methods by an average of 5.8% in alignment accuracy. Our code is publicly available at https://github.com/XNetLab/SGA.

Recently, hypergraph contrastive learning (HCL) has received increasing attention due to its effectiveness in addressing high labeling costs and improving generalization. However, existing HCL methods often rely on random topology augmentation strategies and prioritize a single granularity, making it challenging to preserve critical structural information, high-order relationships, and a balance between global consistency and local feature dynamics during the representation process. To overcome these challenges, this paper proposes a multi-view multi-granularity hypergraph contrastive learning. Specifically, we first design a new sampling strategy that generates three complementary views to enhance hypergraph representation and improve diversity. In addition, we construct a hybrid attention encoder that integrates static homogeneity attention with dynamic self-attention to align global constraints with local adaptability. By jointly optimizing contrastive losses at both node and hyperedge levels, our framework effectively captures structural nuances and semantic correlations, significantly boosting both classification accuracy and clustering quality. This method is well-suited for scenarios with higher-order dependencies, such as session-based recommendation, social group interaction modeling, and biomedical multi-entity networks. Extensive experiments on multiple benchmark datasets demonstrate that M $^{2}$ HyCo improves node classification accuracy by an average of 0.4 percentage points and clustering performance by 5.72 and 8.72 points in NMI and ARI, respectively, over recent strong baselines, especially on Cora-A and Pubmed, thereby evidencing its effectiveness under high-order structural dependencies.

In the era of digital economies, data has emerged as a critical production factor, driving innovation and economic growth. Governments worldwide have implemented policies and frameworks to facilitate data circulation while ensuring privacy protection. Data trading systems play a crucial role in data circulation. However, existing centralized data trading systems face significant challenges, including trust deficits, transparency issues, and vulnerability to data breaches. To address these critical issues, this paper presents TDTS (Trusted Data Trading System), an innovative blockchain-based framework that integrates data identification and private information retrieval mechanisms. Our system architecture leverages three fundamental technologies: blockchain smart contracts for transaction automation, decentralized data identification for provenance verification, and private information retrieval (PIR) techniques for pre-trading data verification. This comprehensive approach establishes a robust pipeline for data flow management, which includes data identification, upload, browsing, and trusted transaction, thereby ensuring secure and verifiable data circulation within a decentralized ecosystem. The implementation utilizes Hyperledger Fabric to construct a consortium blockchain network, specifically designed for data entities engaged in trading activities. We introduce two novel pre-trading verification schemes based on PIR techniques: (1) a randomized data verification format retrieval mechanism, and (2) a multi-query data authenticity verification protocol. Through the implementation of a fully functional TDTS prototype, we conduct extensive performance evaluations and practical assessments, demonstrating the system’s effectiveness and efficiency in addressing trust and security concerns in decentralized data trading environments.

Graph anomaly detection is a crucial task in graph mining that aims to identify nodes that deviate significantly from the majority in terms of structure or attributes. Recent self-supervised learning (SSL) approaches employing graph neural networks (GNNs) have shown promise, yet they suffer from two critical limitations. First, in camouflage scenarios, anomalous nodes strategically connect to normal ones, misleading GNNs to aggregate harmful information and thereby poisoning the learned representations, a phenomenon often termed the "inconsistency problem." Second, most existing methods narrowly focus on node-level information, overlooking vital organizational patterns at higher structural levels such as communities and the global graph, which limits their detection capability. To address these challenges, we propose GraphAD-CMCL, a novel community-aware multi-granularity contrastive learning framework for graph anomaly detection. Inspired by the observation that normal nodes often form cohesive communities with consistent patterns, our method first constructs robust normal patterns by partitioning the graph into communities using both topology and node features. To further capture complex structural deviations, we introduce a hierarchical contrastive learning scheme with dual-level contrasts: node-community and node-graph. This enables the model to evaluate each node's conformity to both its local community and the global graph structure, allowing the detection of sophisticated anomalies that may emerge at different semantic scales. This design not only defends against camouflaged anomalies by mitigating the influence of corrupted neighborhoods but also systematically exploits richer graph information beyond local regions. Comprehensive experiments on five real-world benchmark datasets demonstrate that GraphAD-CMCL significantly outperforms most state-of-the-art methods, validating its effectiveness in detecting complex and camouflaged anomaly patterns.

Although many existing works have studied probabilistic or dynamic environments where the objects used in daily life may be moved due to human activities, the scale of datasets is usually limited due to the cost of human annotation or manual configuration. This paper introduces a framework that simulates human activities and corresponding object dynamics using Large Language Models (LLMs) and applies the simulated human residents to embodied scenes to generate dynamic scenes. Using this framework, we craft a dataset named DynamicTHOR with 50 characters and 100 dynamic scenes, which can be easily extended in scale. A user study comparing our generated scene dynamics with a baseline approach and human annotations validates that our framework successfully produces believable, diversified data of a quality comparable to human annotations. The novel framework and dataset can facilitate the study of embodied intelligence, such as the navigation task in dynamic scenarios.

Reaching a consensus on the team plans is vital to human-AI coordination. Although previous studies provide approaches through communication in various ways, it could still be hard to coordinate when the AI has no explainable plan to communicate. To cover this gap, we suggest incorporating external models to assist humans in understanding the intentions of AI agents. In this paper, we propose a two-stage paradigm that first trains an external trajectory prediction model from collected offline trajectories of the target agent and then utilizes the model in human-AI collaboration by displaying the future action predictions of the target agent in real-time. Such a paradigm leaves the AI agent as a black box and thus is available for improving any agents. To test our paradigm, we further implement a transformer-based predictor and develop an online human-AI collaboration platform based on the Overcooked environment for experiments. The comprehensive experimental results verify that human-AI teams can achieve better performance with the help of our model. A user assessment attached to the experiment further demonstrates that our paradigm can significantly enhance the situational awareness of humans. Our study presents the potential to augment the ability of humans via external assistance in human-AI collaboration, which may further inspire future research.

Edgar Allan Poe noted, “Truth often lurks in the shadow of error,” highlighting the deep complexity intrinsic to the interplay between truth and falsehood, notably under conditions of cognitive and informational asymmetry. This dynamic is strikingly evident in large language models (LLMs). Despite their impressive linguistic generation capabilities, LLMs sometimes produce information that appears factually accurate but is, in reality, fabricated, an issue often referred to as “hallucinations.” The prevalence of these hallucinations can mislead users, affecting their judgments and decisions. In sectors such as finance, law, and healthcare, such misinformation risks causing substantial economic losses, legal disputes, and health risks, with wide-ranging consequences. In our research, we have methodically categorized and analyzed the causes, detection methods, and solutions related to LLM hallucinations. Our efforts specifically focus on elucidating the root causes of hallucinations and evaluating current strategies, thereby laying the groundwork for developing robust mitigation techniques.

Self-evolving LLMs that iteratively train a problem-generating Questioner and a problem-solving Solver present a promising path toward reducing human supervision. Existing approaches optimize each role independently with short-term, separate rewards, neglecting the co-evolutionary dynamics of the Questioner–Solver system as a whole. This leads to model collapse, where the Solver’s performance plateaus or degrades after only a few iterations. To address this, we propose S-BGM, a framework for stable self-evolving LLMs based on bipartite graph modeling. S-BGM models Questioner–Solver interactions as a Cognitive Bipartite Graph, where the two partitions discretize the question space into semantic clusters and the response space into uncertainty intervals. The long-term interactions are encoded into the graph topology, facilitating the tracking of holistic system status. A novel structural entropy is further integrated into both roles' training objectives to measure system-level stability, preventing excessive concentration on the few global interaction structures while preserving local reward optimization. To keep this stability signal dynamically aligned with ongoing co-evolution, the Cognitive Bipartite Graph is updated via a sliding window mechanism that reinforces recent interactions and discards outdated ones. Extensive experiments on ten datasets demonstrate the superiority of S-BGM.

Outlier detection is a critical method in data mining as it has a wide range of practical uses in various fields. Spectral clustering has attracted much attention for its advantage in coping with the curse of dimensionality in high-dimensional spaces. While spectral clustering has been applied to outlier detection, existing methods typically rely on eigenvectors and lack a unified mechanism to simultaneously address both local and global outliers. This paper presents a new spectral analysis based outlier detection (SAOD) framework. We first construct the k-nearest neighbors (kNN) graphs of each sample and compute the eigenvalues of the Laplacian matrices of these graphs. Our findings demonstrate that the distribution of eigenvalues for outliers differs from that of inliers. Based on this observation, we employ the multivariate Gaussian kernel density estimation to empirically estimate the eigenvalue distribution of a given data set. We then introduce a new outlier factor, based on the eigenvalue probabilities of samples to assess their degree. The experimental results, obtained from both synthetic and real-world data sets, show that our spectral analysis based outlier detection method outperforms the state-of-the-art approaches. The code for these experiments is made available at: https://github.com/laetella/SAOD.

The emergence of multi-agent systems built from large language models (LLMs) offers a promising paradigm for scalable collective intelligence and self-evolution. Ideally, such systems would achieve continuous self-improvement in a fully closed loop while maintaining robust safety alignment–a combination we term the self-evolution trilemma. However, we demonstrate both theoretically and empirically that an agent society satisfying continuous self-evolution, complete isolation, and safety invariance is impossible. Drawing on an information-theoretic framework, we formalize safety as the divergence degree from anthropic value distributions. We theoretically demonstrate that isolated self-evolution induces statistical blind spots, leading to the irreversible degradation of the system's safety alignment. Empirical and qualitative results from an open-ended agent community (Moltbook) and two closed self-evolving systems reveal phenomena that align with our theoretical prediction of inevitable safety erosion. We further propose several solution directions to alleviate the identified safety concern. Our work establishes a fundamental limit on the self-evolving AI societies and shifts the discourse from symptom-driven safety patches to a principled understanding of intrinsic dynamical risks, highlighting the need for external oversight or novel safety-preserving mechanisms.