Predicting vehicle traffic optimizes transportation management and urban planning. In this article, we combine real-time data from vehicle-detection Internet of Things (IoT) devices with external variables from Google Trends. Integrating such heterogeneous, complex data streams is challenging for traditional machine learning models that struggle to capture the dynamics of traffic patterns, which are influenced by multiple interdependent factors. To effectively model these complex, interdependent factors, we introduce the Granger-Causal Transformer (GCT), a transformer-based architecture for traffic prediction that integrates an LSTM network with a modified multi-head attention mechanism. This mechanism extends Granger causality to the spatio-temporal domain to analyze all causality relations between features consistently, while capturing long-range dependencies and temporal patterns. Before applying GCT, we generate lagged versions of the Google Trends time series to capture lead and lag effects. Tourists usually make searches about their destination weeks before traveling, so peaks in search interest occur earlier than peaks in weekly traffic volume. Using lags aligns the predictors with weekly traffic volume and allows the model to use past searches to predict future traffic. We semantically validate the Google Trends terms by comparing each term with a reference string describing the study area, using a language model aligned with the data’s linguistic context. We then apply a dual filtering process comprising Granger noncausality and correlation tests to minimize noise and redundancy. We evaluate our proposed methodology against classical statistical models, deep learning models, large foundation models, and transformers across two case studies. The results demonstrate consistently superior performance and generalizability, with GCT achieving \(R^{2}\) improvements between 47% and 68% compared to the best-performing baselines across both settings, alongside substantial reductions in MAE and MSE.

Cognitive modeling, as an emerging technology in the field of computer-aided education, aims to explore students’ knowledge levels and learning abilities to achieve various intelligent educational applications. Although some existing work focuses on addressing the problem of student forgetting, it is still a less explored area how to naturally integrate the forgetting effect caused by the time interval between answering exercises into student knowledge state modeling. Additionally, traditional cognitive modeling methods mostly assume that students answer exercises one by one, which often does not align with real answering behavior and cannot be directly extended to diverse learning scenarios. Therefore, in this article, we propose a Continuous Time-based Neural Cognitive (CT-NC) framework and several implemented models (CT-NCM and two extensions) to effectively integrate the dynamic and continuous characteristics of knowledge forgetting into student learning process modeling, making it more natural. Specifically, we adopt a specially designed learning event encoding method to adjust the neural Hawkes process to capture the relationship between knowledge learning and forgetting over continuous time. Furthermore, we propose a customizable learning function to jointly model the changes in different knowledge states and their interaction with each practice moment. In the end, we demonstrate an extension CT-NCM+ that can adapt well to diverse learning scenarios, indicating that CT-NCM can solve real-world problems by flexibly adjusting its structure. Extensive experimental results on real datasets clearly demonstrate that CT-NCM and CT-NCM+ outperform the current state-of-the-art KT methods in student performance prediction, while our work points out a realistic research direction for KT and demonstrates its interpretability in knowledge learning visualization.

This article introduces the Missingness-aware Self-Supervised Graph Neural Network (MissGNN), a novel model for tabular data predictions in scenarios with missing values. Addressing a significant challenge in fields like finance, healthcare, and environmental studies, MissGNN innovatively integrates feature imputation and label prediction within a unified framework. Utilizing a Tabular Bipartite Graph, it distinctively represents data points and features, enabling intricate modeling of their relationships. This approach is pivotal in contexts where missing data can significantly skew model training and predictions. MissGNN employs a Graph Neural Network to generate embeddings, capturing interactions between data points and features for accurate missing value imputation and label prediction. It features a dual-focus mechanism, comprising edge-wise feature imputation and node-wise label prediction. A pre-training phase for feature imputation enhances the model’s predictive ability for both features and labels. Additionally, MissGNN incorporates a multi-view learning strategy through feature subsetting, further enriching its representation learning. Experiments on 14 tabular datasets demonstrate MissGNN’s superiority over existing models in managing missing values across both regression and classification tasks. The results underscore its robustness and practical applicability, establishing it as a versatile tool for tackling tabular data prediction challenges with missing values. MissGNN’s contributions significantly redefine the approach to missing value imputation, aligning feature imputation and label prediction in a cohesive manner and underscoring the potential of Graph Neural Networks in processing tabular data.

The dominant framing of Artificial General Intelligence (AGI) as a discrete breakthrough obscures the more urgent reality: AGI is arriving as a gradual, cumulative erosion of human verification power distributed across institutions and decision-making systems. This article reframes the AGI transition through the lens of absorption capacity; that is, the rate at which human systems can integrate, govern, and maintain meaningful oversight of increasingly autonomous AI. Drawing on empirical observations from deploying enterprise-scale generative AI in a large public university and personal experiences as a long-standing AI researcher and educator, in this article, I identify three critical asymmetries characterizing this transition: (1) governance lag, where policy cycles cannot match technological iteration speed; (2) institutional misalignment, where locally rational AI systems produce collectively irrational societal outcomes; and (3) capability inequality, where uneven access to AI amplifies structural advantage. I argue that the defining challenge is not achieving technical alignment with human values, but maintaining epistemic authority, which is the human capacity to verify, understand, and steer systems reasoning in latent spaces beyond direct audit. The article concludes that the true measure of preparedness for AGI is not computational power or algorithmic sophistication, but adaptive governance: institutional architectures capable of co-evolving with the technologies they must regulate. The frontier is not artificial superintelligence. It is collective human capacity to remain intelligible to ourselves while embedded in AI-mediated decision ecosystems.

The rapid evolution of AI has driven advancements across numerous sectors. In the domain of government affairs, large language models (LLMs) hold significant potential for applications such as policy analysis, data processing, and decision support. However, their adoption in government settings faces considerable challenges, including data accessibility issues, the absence of standardized evaluation criteria, and concerns regarding model accuracy, reliability, and security. To address these challenges, we propose a comprehensive evaluation framework specifically designed for LLMs in government affairs. Built on modular principles, this framework ensures adaptability across various industries. Additionally, we introduce the Multi-Scenario Government Affairs Benchmark (MSGABench 1 ) dataset, a Chinese-language dataset specifically crafted to meet the practical needs of government professionals. Employing the proposed framework and the MSGA dataset, we conducted an empirical evaluation of 15 prominent LLMs, revealing critical insights: (1) Performance: Many models demonstrated low accuracy and reliability, particularly under minor input variations, with some dropping below 35% accuracy, whereas GPT-4 achieved above 95% reliability; (2) Security and Compliance: Significant concerns were identified, including privacy vulnerabilities, legal compliance risks, and persistent biases, which may hinder secure deployments in government contexts; (3) Task Avoidance: Certain models exhibited excessive caution, often avoiding responses to basic tasks like document classification and government-related inquiries, which restricts their usability. These findings highlight essential limitations and opportunities for improvement, contributing to the safe and effective application of LLMs in the government sector.

Spatial-temporal imputation remains a challenging problem in transportation, environment and healthcare, where the missing value is filled based on spatial, temporal, and cross correlations. Previous research mainly focused on feature-level correlation integration and comprehension with the hand-crafted enhancement strategy. Meanwhile, the recently prevalent large language models (LLMs) provide token-level understanding for language linguistics, and whether they could be applied for spatial-temporal correlation enhancement is under exploration. To this end, we proposed an LLM-native framework STOMA to fully utilize the intrinsic relevance. We designed semantic enhancing methods by converting the complex correlations, e.g. spatial correlation in network, temporal correlation with periodicity and cross correlation from human behavior, into the embedded tokens. Specifically, we reform dynamic time warping as an asymmetric correlation constructor for complex dynamics. We adapt the proposed backbone along with the spatial-temporal fine-tuning technique, and the empirical results demonstrate the effectiveness of our methods over recent LLM-inspired methods evaluating on real-world datasets.

Task offloading in distributed computing involves complex trade-offs among delay, scalability, cost, and resource utilization. Cloud platforms face long communication delays, while edge nodes have constrained capacity. Static, rule-based schedulers cannot adapt to fluctuating loads or per-instance heterogeneity. Similarly, Reinforcement Learning (RL) schemes typically address only a single layer or assume homogeneous servers, overlooking the hierarchical and multi-instance nature of deployments. To address these challenges, we introduce a server-aware Proximal Policy Optimization (PPO) framework that performs fine-grained offloading across a three-tier (Edge, Regional, Cloud), multi-instance architecture. We formulate offloading as a Markov Decision Process whose state vector includes per-instance delay, CPU/memory utilization, network congestion, cost, and energy metrics. The PPO agent learns to offload tasks to the best server in real time. Our developed RegionalEdgeSimPy simulation shows that PPO agent makes optimal offloading choices for over 90% of tasks, keeping each server near; however, below a 70% utilization. This optimized decision making drives up to 66.9% delay reduction, 78.6% energy savings, and 47.8% cost reductions relative to cloud-only and edge-only baselines.

Accurate traffic forecasting is crucial for effective urban planning and transportation management, enabling efficient resource allocation and enhanced travel experiences. However, existing models often face limitations in generalization, struggling with zero-shot prediction on unseen regions and cities, as well as diminished long-term accuracy. This is primarily due to the inherent challenges in handling the spatial and temporal heterogeneity of traffic data, coupled with the significant distribution shift across time and space. In this work, we aim to unlock new possibilities for building versatile, resilient and adaptive spatio-temporal foundation models for traffic prediction. We introduce OpenCity, a foundation model that captures underlying spatio-temporal patterns from diverse data, facilitating zero-shot generalization across urban environments. OpenCity integrates Transformers with graph neural networks to capture complex spatio-temporal dependencies in traffic data. By pre-training OpenCity on large-scale, heterogeneous traffic data from web platforms, we enable the model to learn rich, generalizable representations that can be seamlessly applied to a wide range of traffic forecasting scenarios. Experiments show OpenCity excels in zero-shot prediction and exhibits scaling laws, highlighting its potential as a universal one-for-all traffic prediction solution adaptable to new urban contexts with minimal overhead. Source codes are available at: https://github.com/HKUDS/OpenCity

ABSTRACT Currently, depression treatment relies on closely monitoring patients’ response to treatment and adjusting the treatment as needed. Using self-reported or physician-administrated questionnaires to monitor treatment response is, however, subjective, costly and suffers from recall bias. In this paper, we explore using location sensory data collected passively on smartphones to predict treatment outcome. To address heterogeneous data collection on Android and iOS phones, the two predominant smartphone platforms, we explore using domain adaptation techniques to map their data to a common feature space, and then use the data jointly to train machine learning models. We further explore integrating contrastive learning with domain adaptation to augment data and learn feature embeddings. These learned embeddings are then used to train machine learning models to predict depression treatment outcomes. Our evaluation shows that using the embeddings learned by jointly integrating contrastive learning and domain adaptation leads to the best prediction accuracy. In addition, our results show that using location features and baseline self-reported questionnaire score can lead to F1 score up to 0.76. This accuracy is comparable to that obtained using periodic self-reported questionnaires, indicating that using location data is a promising direction for predicting depression treatment outcome. Last, when all location and questionnaire data are used together, the F1 score further increases to 0.79.

In the context of rising demands for marine resource exploitation and scientific research, collaborative trajectory planning for multiple Autonomous Underwater Vehicles (AUVs) in complex underwater environments—marked by obstacles, ocean currents, and low visibility—remains a critical challenge. Although the Gray Wolf Optimization (GWO) algorithm has advanced multi-objective trajectory planning, it faces issues such as poor high-dimensional space adaptability, susceptibility to local optima, and insufficient constraint handling. To address these, this article proposes a multi-AUV trajectory planning algorithm (EA-GWO) based on evolutionary learning to improve GWO. The method optimizes multi-AUV trajectory planning by leveraging hierarchical population hunting behavior, integrating position update equations to prioritize population bootstrapping, and balancing exploration and exploitation via fitness-based population distribution. Experimental validation across general, ocean current, and threat environments compares EA-GWO with the traditional GWO and multiple population GWO (MP-GWO). For sailing time: in the general environment, EA-GWO reduces total time by 90.6% compared to GWO and 90.6% compared to MP-GWO; in the ocean current environment, it cuts time by 0.9% versus GWO and 2.4% versus MP-GWO; in the threat environment, it cuts time by 13.6% versus GWO and 14.9% versus MP-GWO. For sailing distance: in the general environment, EA-GWO shortens total distance by 9.8% compared to GWO and 3.4% compared to MP-GWO; in the ocean current environment, it reduces distance by 2.3% versus GWO and 4.9% versus MP-GWO; in the threat environment, it shortens distance by 5.5% versus GWO and 1.0% versus MP-GWO. In terms of convergence performance reflected by the fitness curve: across the three environments, EA-GWO demonstrates faster convergence speed. These results highlight that EA-GWO outperforms the other two algorithms in sailing time, distance, and convergence efficiency, verifying its effectiveness in real-time dynamic coordination and constraint handling for multi-AUV missions.