Fine-tuning remains an important step in adapting 3D foundation models to downstream applications, but it is laborious and time-consuming to annotate training datasets in 3D. Existing semi-automatic annotation methods for segmentation masks fail to generalize beyond RGB point cloud data and are difficult to control. While a number of 3D annotation interfaces provide undo/redo functionality to help correct mistakes, they lack a comprehensive data versioning strategy, limiting the potential for user analysis and quality control. To tackle these issues, we present DynamicSAM Annotator, an AI-assisted point segmentation framework with the following contributions: 1) a novel interactive 3D point cloud segmentation model that exclusively uses XYZ coordinates with an adaptive thresholding mechanism in segmentation masks; 2) a 3D annotation platform to enable efficient 3D point cloud annotation for semantic and instance segmentation tasks; and 3) a fine-grained version control system that efficiently persists each annotation operation to disk using Git-like operations. Rigorous evaluations across seven benchmark datasets demonstrate DynamicSAM's effectiveness in both indoor and outdoor environments, achieving an average Intersection over Union (IoU) of 59.6% for a single click, highlighting the model's outstanding generalization capabilities. Meanwhile, our simulated benchmark shows that our version control system maintains sub-second latency even with tens of thousands of snapshots stored.
Can AI Agents be benchmarked within strictly controlled simulated trading environments to investigate how external factors impact their collective trading behaviors? These factors, which frequently influence trading behavior, are critical elements in the quest to maximize investors’ profits. Our work aims to address this problem by utilizing large language model-based agents. We have developed a multi-agent AI system, StockAgent, driven by LLMs and designed to systematically explore simulated trading behaviors in controlled environments. StockAgent enables examination of how external factors might affect agent behavior and profitability in simulations, without empirical validation for real-world use. Additionally, StockAgent avoids the test-set leakage issue present in existing AI-agent-based trading simulation systems. Specifically, it prevents the model from leveraging prior knowledge it may have acquired related to the test data. We evaluate different LLMs within the StockAgent framework, which serves as a benchmark for testing LLM behavioral tendencies with rigorous internal validity checks and non-LLM baselines. The experimental results demonstrate the impact of key external factors on stock market trading, including trading behavior and the rules governing stock price fluctuations. This research examines the phenomenon of agents’ free-trading gaps in the context of no prior knowledge of market data. The patterns identified through StockAgent simulations offer methodological insights into LLM behaviors in simulated financial environments. The code is available at: https://github.com/MingyuJ666/Stockagent .
Recommendation systems have demonstrated success in suggesting relevant items to users by efficiently filtering through a wide range of potential options. However, recommender systems used in physical environments need to find a balance between the relevance of suggestions and potential contextual risks. This paper introduces a novel type of risk-aware recommender system (RARS) designed for smart spaces, where the user movements and the room occupancy influence both the quality and risk of recommendations. Using the Museum of Modern Art (MoMA) as a case study, we integrate a virus contagion model into the recommendation process to minimize exposure risk. A knowledge graph-based architecture is employed to generate content-based recommendations, enriched through semantic embeddings. We propose and evaluate multiple reward functions that integrate both recommendation scores and infection risk using different normalization score strategies. An exhaustive set of simulations prove how various decision-making strategies and risk thresholds affect the recommendation quality, infection risk, and movement efficiency. Our findings reveal that RARS can significantly reduce user exposure while preserving a satisfactory relevance of recommendations, offering a generalizable framework for safety-aware recommender systems in public venues.
Learning effective latent representations for users and items is the cornerstone of recommender systems. Traditional approaches rely on user-item interaction data to map users and items into a shared latent space, but the sparsity of interactions often poses challenges. While leveraging user reviews could mitigate this sparsity, existing review-aware recommendation models often exhibit two key limitations. First, they typically rely on reviews as additional features, but reviews are not universal, with many users and items lacking them. Second, such approaches do not integrate reviews into the user-item space, leading to potential divergence or inconsistency among user, item, and review representations. To overcome these limitations, our work introduces a Review-centric Contrastive Alignment Framework for Recommendation (ReCAFR), which incorporates reviews into the core learning process, ensuring alignment among user, item, and review representations within a unified space. Specifically, we leverage two self-supervised contrastive strategies that not only exploit review-based augmentation to alleviate sparsity, but also align the tripartite representations to enhance robustness. Moreover, we enhance ReCAFR with LLM-based contrastive strategies to not only improve robustness to varying review quality, but also reduce information loss during view aggregation. Empirical studies on public benchmark datasets demonstrate the effectiveness and robustness of ReCAFR.
Few-shot semantic segmentation, which is dedicated to the generalization of models to segment novel classes with scarce annotated samples, has achieved tremendous progress recently due to the significant advancement of deep CNNs. However, existing approaches in medical scenarios, namely Few-Shot Medical Image Segmentation (FSMIS), still encounter two primary obstacles. First, there exists huge appearance discrepancy between support and query images, which hinders the knowledge transferring and adversely affects segmentation performance. Second, almost all current prototype-based methods struggle to learn and optimize limited support prototypes, giving insufficient attention to query information, which makes it challenging to achieve high-quality query segmentation. Consequently, we propose a novel Query-Assisted Multiple Prototypes (QAMP) approach, where in addition to normal support prototypes, query prototypes are additionally mined leveraging high-confidence initial query predictions. Specifically, we design a Query Prior Generation (QPG) module to locate positions where query objects may belong to with high confidence. Subsequently, based on corresponding support mask and query prior, a Mask Guided Support Prototypes (MGSP) module and a Prior Guided Query Prototypes (PGQP) module are employed to generate support and query prototypes respectively, which can effectively capture underlying characteristics of the query targets. Extensive experiments and visualization on three publicly available medical image datasets demonstrate the superiority of our QAMP compared with current methods. Code is available at https://github.com/jcjiang99/QAMP .
Discovering urban functional zones (UFZs) is critical for understanding city spatial structures and supporting effective urban planning. Existing approaches to UFZ discovery typically rely on one of three costly strategies: (1) training large vision models directly on satellite imagery, which demands substantial computational resources; (2) leveraging crowdsourced data such as Points of Interest (POIs) from platforms like OpenStreetMap, which may be incomplete, inconsistent, or unavailable in many regions; or (3) collecting custom labeled data, which requires significant time, expense, and expert effort. Recently, large multi-modal models (LMMs) emerged as a promising alternative, offering strong capabilities in interpreting visual content without requiring extensive data labeling. However, their performance remains limited when applied to the UFZ discovery task, often struggling to capture the complex spatial and functional details and interactions of urban regions. To address this challenge, we propose a new approach that enhances LMMs’ reasoning capability to recognize urban functional zones by keeping LMM encoders frozen while training only lightweight graph-based models, eliminating the need for LMM fine-tuning or additional pre-training. Specifically, our approach first partitions the target area into small regions by road network, where an LMM is used for each region to generate visual and textual embeddings independently using its image and text encoders. Then, two graphs are constructed in which nodes represent regions with features defined by their respective embeddings, and edges encode their spatial adjacency. Message-passing on the two graphs hence captures spatial correlation between the visual and textual modalities. After that, graph clustering will suggest the prototypes representing nearby zones with similar urban functions, where contrastive learning is further leveraged to encourage cross-modal consistency. Evaluation on four city districts, namely Philadelphia (PA, USA), Pudong (Shanghai, China), San Francisco (CA, USA), and Seattle (WA, USA), substantiates the effectiveness of our proposal and that its performance is on par with supervised competitors.
Representation learning is essential for deep-neural-network-based recommender systems to capture user preferences and item features within fixed-dimensional user and item vectors. Unlike existing representation learning methods that either treat each user preference and item feature uniformly or categorize them into discrete clusters, we argue that in the real world, user preferences and item features are naturally expressed and organized in a hierarchical manner, leading to a new direction for representation learning. In this article, we introduce a novel matryoshka representation learning method for recommendation (MRL4Rec) , by which we restructure user and item vectors into matryoshka representations with nested vector spaces to explicitly represent user preferences and item features at different hierarchical layers. We theoretically establish that training with the same triplets for each sliced vector cannot guarantee representation learning with hierarchical structures. Subsequently, we propose the layer- and hardness-adaptive negative sampling (LHANS) mechanism to construct training triplets, which further ensures the soundness of learned matryoshka representations in capturing hierarchical user preferences and item features. The experiments demonstrate that MRL4Rec can consistently and substantially outperform a number of state-of-the-art competitors on several real-life datasets. Our code is publicly available at https://github.com/Riwei-HEU/MRL .
An ontology is a formal, explicit specification of a shared conceptualisation, which can be combined with problem-solving methods and reasoning functionality to develop high-quality technology and application systems efficiently. Ontology engineering typically involves extensive manual effort to elicit intended use cases (user stories) from users for the target ontology-based systems. Recent studies have demonstrated the positive potential of large language model-based conversational agents in supporting user story generation in OE. However, we argue that we are not leveraging LLM to its fullest potential by not supporting users in formulating effective prompts. To address this, we identify the prompt guidance users need during user story generation workflows by conducting a formative study (N = 10) using participatory prompting. We demonstrate its usefulness through the design and development of the OntoChat LLM-based system for OE, as well as a user evaluation with knowledge engineers (N = 24). To our knowledge, this is the first work to design and validate a prompt guidance framework that helps users leverage LLM to its fullest potential to generate effective requirements for ontology development. This advances how we interact with LLM for requirements elicitation.
Federated Learning (FL) enables collaborative model training while preserving data privacy, but the presence of noisy labels in local datasets remains a significant challenge, particularly under heterogeneous noise conditions and class imbalance. In this work, we introduce a novel Multi-Stage Robust Federated Learning (MRFL) framework to address these issues. In the warm-up noise detection stage, MRFL computes per-class average losses on each client and employs a Gaussian mixture model to accurately identify clients with substantial label noise. In the subsequent noise-robust training stage, a robust loss function and noise solver are designed to distinguish clean from noisy samples, while semi-supervised learning is used to recover valuable information from tail classes. Moreover, a robust weighted aggregation strategy is adopted to mitigate the adverse effects of noisy clients. Extensive experiments on CIFAR-10/100-LT and ICH datasets demonstrate that MRFL outperforms state-of-the-art methods in federated noisy label learning scenarios characterized by data heterogeneity and imbalance.
The rapid advancement of Large Vision-Language Models (LVLMs) has enhanced their capabilities from content creation to productivity enhancement. Despite their innovative potential, LVLMs exhibit vulnerabilities, especially in generating potentially toxic or unsafe responses. Malicious actors can exploit these vulnerabilities to propagate toxic content using strategically crafted prompts without fine-tuning or compute-intensive procedures. Despite ongoing red-teaming efforts to identify and mitigate these risks, the exploration of LVLM vulnerabilities remains nascent and yet to be fully addressed in a systematic approach. This study systematically examines the vulnerabilities of open and closed-weight LVLMs, including LLaVA , InstructBLIP , Fuyu , Qwen , DeepSeek , Gemini , GPT , and Grok , using adversarial prompting strategies informed by social theories to simulate real-world social manipulation tactics. Our findings show that (i) toxicity and insult are the most prevalent behaviors, with mean toxicity scores 19.32% and 12.36%, respectively; (ii) Gemini-2.0-Flash , LLaVA-v1.6-Vicuna-13B , and Grok-2-Vision-1212 are the most vulnerable models. Their toxic response rates reach 46.93%, 23.81%, and 17.98%, respectively, while insult response rates reach 47.94%, 14.62%, 12.27%, respectively; (iii) prompting strategies incorporating dark humor and multimodal toxic prompt completion significantly elevate these vulnerabilities. Despite extensive safety alignment efforts, models still generate content with varying degrees of toxicity when prompted with adversarial inputs, highlighting the urgent need for enhanced safety mechanisms and robust guardrails in LVLM development.
Variational autoencoder (VAE) frameworks for multi-interest discovery have demonstrated strong recommendation performance. However, because VAEs model users independently, they fail to capture shared interest patterns among like-minded users, a key limitation of existing approaches. The previous work, Varium [41], addressed this by introducing a flat structured memory with iterative refinement to capture reusable interest patterns across users. In this paper, we extend Varium into Varium+, a novel semantic expert-structured memory-augmented VAE that fundamentally rethinks how shared interests are organized and retrieved. Unlike its predecessor, Varium+ organizes memory as a mixture of semantically-grounded experts, each dedicated to a different modality (textual, visual, collaborative) and composed of multiple prototypes representing fine-grained interest groups. For each user, we first discover multiple intra-user interest factors behind their item adoptions. Intra-user interest factors next attend to these expert memories via cross-attention to retrieve relevant inter-user signals, then pass through self-attention to model interactions among retrieved interests. The resulting shared interest representations are integrated back into each user’s multi-interest profile, enriching personal preferences with collaborative wisdom. Experiments on real-world datasets demonstrate that Varium+ consistently outperforms both its predecessor and a comprehensive set of baselines, with particularly strong gains when leveraging semantic knowledge. Qualitative analysis further reveals that our expert-structured memory organizes the latent interest space into well-separated clusters, validating its ability to discover meaningful shared patterns among users.
The development of mobile web services has brought significant attention to spatial crowdsourcing. The uneven distribution of tasks and workers has led to recent research on Cross-Platform Spatial Crowdsourcing (CPSC), aiming for a multi-win situation for platforms, workers, and task requesters. Previous studies on CPSC problems focused on task assignment and worker selection performance, overlooking the importance of privacy preservation. This article addresses the existing challenges of privacy preservation and service quality by formulating a Privacy-Preserving Cross-Platform Spatial Crowdsourcing (PP-CPSC) problem and proves it to be NP-hard. We propose an Evolutionary Differential Privacy (Evo-DP) approach to optimize PP-CPSC. Evo-DP’s evolutionary framework enables efficient and flexible optimization of privacy budget allocation. Within Evo-DP, each solution to the privacy budget allocation is represented as an individual in the population. To approximate the optimal solution, three evolutionary operations—mutation, crossover, and scaling—are employed for population updates, along with a selection process. A hybrid population model is introduced to balance exploration and exploitation abilities. Experimental results demonstrate Evo-DP’s superiority over previous strategies in terms of solution quality, convergence speed, and scalability.
Breast cancer is a prevalent and life-threatening disease where early and accurate diagnosis is critical for effective treatment. Conventional histopathological analysis, while the standard for diagnosis, can be laborious and is subject to inter-observer variability, highlighting the need for robust automated methods. This article introduces EVC-Net, a novel hybrid deep learning framework designed to automate the classification of breast cancer from histopathological images. EVC-Net synergistically integrates an EfficientNetV2S for fine-grained texture feature extraction, a vision transformer (ViT) for capturing global context, and a capsule network to preserve spatial hierarchies within tissue structures. The proposed model is evaluated on the public BreakHis dataset. Across all four magnification levels, EVC-Net demonstrates robust performance, achieving an average accuracy of 0.985 and an AUC-ROC of 0.994 for binary (benign vs. malignant) classification. For the eight-subtype multi-class task, the model maintains high efficacy, attaining an average accuracy of 0.954 and an AUC-ROC of 0.980. Furthermore, interpretability analysis using Grad-CAM is conducted, generating heatmaps overlay visualization to understand the rational behind model’s predictions. These results demonstrate the potential of the EVC-Net framework to enhance diagnostic accuracy and consistency, offering valuable support for clinical workflows in oncology.
Neuro-symbolic predictive process monitoring aims to leverage symbolic knowledge about processes to improve predictive models developed from event logs. While previous work has primarily focused on crisp declarative process knowledge, many existing neuro-symbolic formulations rely on deterministic background knowledge, where rules and constraints are assumed to hold universally and without exception. In real-world business processes, however, this assumption is often unrealistic, as constraints and domain rules may hold only with varying degrees of compliance. Incorporating probabilistic background knowledge, therefore, provides a more flexible and realistic framework for modeling domain constraints that hold with certain likelihoods rather than as strict deterministic rules. In this article, we propose an approach that leverages a probabilistic Declare model to represent contextual factors that can improve the accuracy of trace suffix prediction. By extending crisp constraints with the probability with which they are satisfied in an event log, a probabilistic Declare model entails a set of possible compliance scenarios whose satisfaction or violation can be monitored as a trace execution unfolds. The proposed approach exploits this compliance-checking information to adjust the posterior probability of an activity occurrence returned by a next-activity predictor when generating the suffix of a given prefix. The experimental evaluation shows that incorporating this type of symbolic process knowledge generally improves the predictive performance of the model.
In recent years, social recommendation systems have emerged as a pivotal technology for enhancing recommendation accuracy by leveraging user social homophily and influence. Although many works have been devoted to this area, existing works still struggle to extract the beneficial structural information from social relationships that is beneficial for recommendations and neglect the inherent popularity bias in the social networks, which leads to suboptimal recommendation performances. To address these challenges, we propose a novel framework termed Causal Disentanglement-Enhanced Diffusion Denoising for Social Recommendation (CaDDiSR). This framework first employs causal graphs to disentangle the complexities of social relationships, generating user representations with high-order structures, which are subsequently used as inputs to a diffusion process to effectively denoise social networks and retain social signals beneficial for recommendation tasks. Furthermore, the framework integrates a bidirectional knowledge distillation mechanism, which balances user representations between social and recommendation contexts, thereby facilitating the effective fusion of their respective advantages while simultaneously mitigating noise interference and enhancing overall system performance. Finally, cross-domain contrastive learning is utilized to optimize user and item representations, ensuring consistency in recommendation performance across diverse scenarios. Experimental results on multiple real-world datasets demonstrate that CaDDiSR significantly outperforms existing baseline models, substantiating its superior performance.
Structural relationships among Knowledge Components (KCs) are essential for adaptive learning systems, as they support accurate cognitive diagnosis, personalized path planning, and targeted resource recommendation. However, existing approaches frequently capture correlations instead of reliable directional dependency signals and tend to converge prematurely or become inefficient as graph dimensionality grows. These limitations weaken the reliable modeling of KC-level structure, which in turn reduces interpretability and limits downstream benefits for diagnosis, planning, and recommendation. To this end, we propose a novel structure learning framework that integrates psychometric modeling with structural search. First, we design the Item Response Theory (IRT)-Based Information Criterion (IRIC) , an interpretable scoring function that combines information entropy with causal effect estimation grounded in IRT, jointly capturing statistical associations and directionality-sensitive signals under latent ability control. Second, we develop Co-Evolutionary Optimization for Structural Search (CEO-SS) , a multi-population evolutionary algorithm with a game-inspired co-evolution mechanism that balances exploration and exploitation, avoiding premature convergence and showing robust search behavior as graph dimensionality increases within the evaluated benchmarks. Extensive experiments on three types of datasets—including benchmark causal discovery datasets, the public educational dataset, and real-world classroom data—demonstrate that our framework consistently outperforms strong baselines in accuracy and stability, with especially clear gains in adjacency recovery and more modest improvements in edge-direction recovery. In addition, expert evaluation suggests that the learned structures are more diagnostically useful, more actionable for remediation, and more pedagogically plausible than those produced by alternative scoring methods. Overall, the proposed framework provides an interpretable and practically valuable approach to learning KC structures for adaptive learning.
Spatio-temporal event forecasting, including emergency event prediction, traffic accident prediction, and mobility demand prediction, plays a vital role in resource allocation and public safety. However, existing prediction models can inadvertently amplify socioeconomic disparities through biased predictions that systematically favor advantaged communities. While recent fairness-aware approaches address global fairness across entire regions, they neglect local fairness that focuses on reducing disparities along the boundaries between neighborhoods with different socioeconomic statuses. We propose LG-Fair, a novel prediction framework that jointly optimizes both global and local fairness while maintaining prediction accuracy. Specifically, we propose novel local fairness metrics based on the corresponding global fairness counterparts for spatio-temporal forecasting. We design a flexible fairness framework that can optimize a hybrid loss function that balances prediction accuracy with fairness at both global and local scales. Our framework is generalizable as it can incorporate different fairness metrics and adapt to different backbone prediction architectures. Experiments on diverse real-world datasets demonstrate that LG-Fair significantly reduces fairness disparities at multiple spatial scales while maintaining competitive prediction performance, highlighting its potential in various application scenarios.
Large-scale multi-agent systems face two core challenges: inefficient policy learning and the explosion of state dimensions. Existing methods often rely on manually designed task sequences to guide agents’ learning in stages, but these designs lack adaptability to agents’ learning abilities, making it difficult to ensure the rationality of task difficulty. Moreover, the representation capability of current network structures is limited, making it challenging to efficiently handle high-dimensional state information and complex interaction relationships. To address these issues, we propose a Progressive Multi-Agent Reinforcement Learning (PMARL) framework. PMARL introduces a task adapter that adaptively selects task difficulty based on agents’ learning abilities, eliminating reliance on manual experience. Additionally, a Dynamic Dimension Adaptive Network (DDAN) is designed, incorporating hypernetwork and self-attention mechanisms to achieve adaptive feature extraction of high-dimensional states and efficient representation of agent interaction relationships. Experimental results demonstrate that PMARL exhibits higher efficiency and better adaptability compared to existing methods when addressing large-scale multi-agent tasks.
Urban traffic congestion remains a persistent challenge, primarily due to outdated road network infrastructures that fail to accommodate evolving transportation demands. Although traditional mitigation strategies such as traffic signal control and lane reconfiguration offer localized and short-term relief, they often lack the systemic capacity necessary for ensuring long-term urban mobility resilience. To support structured and future-oriented road reconstruction, there is an urgent need for intelligent planning tools capable of perceiving network topology, understanding traffic dynamics, and executing effective reasoning. In this study, we introduce RoadGPT , an innovative decision-making framework designed for intelligent road network reconstruction. RoadGPT positions a Large Language Model (LLM) as the central planning agent, integrating multi-source urban data and leveraging carefully constructed prompts to emulate expert-level reasoning. Through a fusion of semantic representation and structure-aware guidance, the LLM effectively identifies congestion bottlenecks and generates coherent and actionable structural intervention strategies. To enhance the LLM's capacity for spatial understanding, RoadGPT incorporates Graph Convolutional Networks (GCNs) to model the topology and connectivity of the urban road network. These GCN-based embeddings provide essential structural context, thereby improving decision quality and enabling system-level coherence. Furthermore, we formalize the decision-making workflow as a Markov Decision Process (MDP) and adopt a Reinforcement Learning (RL) optimization loop driven by continuous feedback from the SUMO microscopic traffic simulator. This interactive loop allows the model to iteratively refine its reconstruction policy based on key performance indicators such as average travel time, queue length, and reconstruction cost. We validate RoadGPT on five real-world urban datasets of varying scales, experimental results indicate that RoadGPT consistently outperforms state-of-the-art baselines.
