Mobile GUI agents are becoming increasingly important for helping users control devices efficiently, with (Multimodal) Large Language Model (MLLM)-based approaches emerging as key contenders. However, most agents are limited to following human instructions, which hinders their effectiveness in scenarios that require foresight and autonomous decision-making. Additionally, These agents fail to leverage the contextual information associated with users during task execution, thereby neglecting the potentially vast differences in user preferences. To address these challenges, We introduce the FingerTip benchmark. It contains unique human demonstrations of multi-step device interactions across a variety of everyday apps, including the screens and actions, and corresponding natural language intents as labels. These demonstrations are not isolated but are continuously acquired from the users' long-term usage simulating real-world conditions, and encompass essential user-related contextual information. This benchmark poses a new challenge: predicting user intents by directly interpreting user interfaces, which enables proactive task suggestions and personalized task execution. We will fine-tune models to compare with zero and few-shot baselines, and organize our benchmark to facilitate robustness analysis, i.e., how well an agent performs in the presence of new intents, new applications, or new users. This is a work in progress, and we welcome any insightful questions or suggestions!

The evolutionary processes of complex systems contain critical information regarding their functional characteristics. The generation time of edges provides insights into the historical evolution of various networked complex systems, such as protein-protein interaction networks, ecosystems, and social networks. Recovering these evolutionary processes holds significant scientific value, including aiding in the interpretation of the evolution of protein-protein interaction networks. However, existing methods are capable of predicting the generation times of remaining edges given a partial temporal network but often perform poorly in cross-network prediction tasks. These methods frequently fail in edge generation time recovery tasks for static networks that lack timestamps. In this work, we adopt a comparative paradigm-based framework that fuses multiple networks for training, enabling cross-network learning of the relationship between network structure and edge generation times. Compared to separate training, this approach yields an average accuracy improvement of 16.98%. Furthermore, given the difficulty in collecting temporal networks, we propose a novel diffusion-model-based generation method to produce a large number of temporal networks. By combining real temporal networks with generated ones for training, we achieve an additional average accuracy improvement of 5.46% through joint training.

There has been extensive research on simulating individual behavior using LLMs, but studies on simulating organizational decision-making with LLMs are scarce, especially in the context of the macroeconomics. In this work, we use LLMs to simulate key decisions of firms, including output, pricing, wages, and organizational arrangements, to replicate stylized facts in the labor and consumption markets within a macroeconomic environment.

The frequency and intensity of natural disasters has significantly increased over the past decades. Facing these possible and unexpected disasters, accurately modeling human mobility become the critical issue for planning effective humanitarian relief, disaster management, and long-term societal reconstruction. In particular, human mobility under big disasters such as rainstorm deviates from its daily routine to a large extent, which makes the task more challenging. However, existing works mainly focus on trajectory generation in normal situations. To tackle this problem, we propose a diffusion model-based method for generating trajectories in disasters to accurately model the impact of disasters on human mobility. Specifically, we integrate weather and geographical information based on a GCN model to capture the spatiotemporal impact of rainfall on mobility. Subsequently, we generate high-fidelity trajectories based on a diffusion model conditioned on disaster information, and constrain the outcomes by modeling mobility perturbation. This work is ongoing, and we welcome your valuable feedback.

While AI has achieved remarkable advances in weather forecasting, climate forecasting presents a much greater challenge due to its long-term nature, complex processes, and multiscale interactions between the atmosphere, oceans, and land. Accurate climate forecasting is essential for addressing climate change, informing policy decisions, and predicting natural disasters. A key aspect of climate prediction lies in understanding teleconnections—climate anomalies or patterns in one region that influence distant areas, typically driven by atmospheric or oceanic circulation patterns. Existing AI solutions often oversimplify climate data, treating it as trivial spatiotemporal data, and primarily rely on models like CNNs and Transformers to predict specific climate modes. In this work, we introduce a unified AI model designed to capture and predict global climate dynamics by modeling global spatio-temporal dynamics and teleconnection patterns. The model is expected to simultaneously forecast multiple climate modes, including the El Niño-Southern Oscillation (ENSO), Indian Ocean Basin (IOB), Indian Ocean Dipole (IOD), and Tropical North Atlantic (TNA). Our goal is to demonstrate the potential of deep foundation models in improving the efficiency and accuracy of climate forecasting.

Witnessing rapid development, large language models (LLMs) are increasingly capable of general problem-solving across various tasks. However, their capabilities in complex reasoning tasks, like math problems, are still limited. On the one hand, recent advances such as OpenAI o1 and o3 significantly enhance the reasoning capabilities of LLMs, but training such models requires massive data and computational resources as well as sophisticated architecture designs. On the other hand, with fixed LLM parameters, inference-time techniques, including Tree-of-Thoughts (ToT) and Graph-of-Thoughts (GoT), guide the reasoning process utilizing particular logical structures. Still, these structures are applied to all tasks without distinction, lacking adaptability. In this work, we design five atom reasoning actions from the cognitive perspective of understanding how humans engage in complex reasoning tasks. We utilize deep reinforcement learning (DRL) algorithms to train a navigator model, which learns to analyze the characteristics of a given task and adaptively select atom actions to compose the most suitable logical structure for reasoning. Extensive experimental results on multiple benchmarks demonstrate the superior performance of our method among well-known inference-time techniques. With only 2K parameters, our navigator model enhances the reasoning capability of various LLMs with <10B parameters to surpass those with >100B.

With the rapid development of the low-altitude economy, drones are being increasingly utilized in urban environments for applications such as search and rescue, and logistics transportation. However, achieving efficient area exploration and path planning in dynamic and open 3D urban environments remains a significant challenge. To address this, we proposes the CitySearch framework, a multi-agent collaborative exploration approach tailored for urban settings. By leveraging cooperative perception and data sharing, drones can achieve efficient urban environment reconstruction, optimize exploration strategies, and enhance the efficiency of search tasks.

Chain-of-Thought (CoT) reasoning has been shown to enhance the performance of large language models (LLMs) on a variety of complex tasks. However, its effectiveness does not apply to all tasks, suggesting that CoT is not always necessary or beneficial. This study aims to systematically investigate the necessity of CoT in LLMs using a decoding-based approach. We select a set of diverse benchmarks and corresponding LLMs to ensure comprehensive evaluation. Our findings show that the token probability distribution patterns during decoding vary significantly across different benchmarks, and the characteristics of these distributions (such as distribution increment and fluctuation) are strongly correlated with the necessity of CoT. We aim to design a new reasoning strategy from a decoding perspective, enabling LLMs to judiciously apply CoT when beneficial, thereby improving their performance across a range of tasks. This is an ongoing work, and we welcome any valuable questions or suggestions!

With the increasing demand for intelligent travel planning, we introduce CityGPT-Travel, a knowledge-enriched large language model (LLM) tailored for generating personalized and practical travel plans. Unlike traditional systems, CityGPT-Travel leverages SFT datasets—TravelQA, TravelReasoning, and TravelNotes—to enhance its understanding of travel-related knowledge, reasoning over spatial and temporal constraints, and learning from real-world travel experiences. These capabilities form the foundation of TravelAgent, a travel planning system that generates plans based on user inputs while incorporating feedback from an LLM-as-a-Judge module to iteratively improve the output. To evaluate travel plans, we propose TravelBench, a multi-dimensional evaluation framework considering common-sense and preference constraints, alongside automated evaluation from the LLM Judge. While our work is ongoing, CityGPT-Travel shows promise in advancing intelligent, user-adaptive travel planning systems. We welcome questions and suggestions to further improve this research.

Effective management of urban traffic can enhance travel efficiency, reduce energy consumption and emissions, and thus contribute to the sustainable development of cities. With the advancement of Internet of Things (IoT) technologies, information generated by various urban traffic entities can be rapidly and accurately acquired and transmitted. However, there remains significant research potential in fully mining and utilizing this vast amount of traffic information. Although some existing studies have introduced reinforcement learning methods to map information to corresponding optimal decisions, these works still lack scalability in decision-making scope and coordination among decision-making entities. Building on this foundation, the present study focuses on V2I scenario, an important decision-making scenario in future urban traffic, and constructs feasible large-scale decision-making strategies for it. Specifically, by dynamically modeling the spatial and temporal correlations of roads, vehicles can make more "forward-looking" choices of turning paths based on information from potentially dependent roads, thereby balancing efficiency and energy consumption. Additionally, by aligning the collaborative correlations among road traffic signal controls, the study achieves efficient optimization of large-scale signal controls that take into account the turning intentions of vehicles.

Embodied agents are rapidly evolving, but current experiments are often restricted to confined indoor environments. To overcome these limitations, we developed Embodied-City, an open and city-scale simulation platform that integrates Unreal Engine 5 (UE5) with AirSim to simulate both drones and autonomous vehicles. This platform enables the testing and deployment of embodied agents in large, dynamic urban environments, providing a more realistic and comprehensive testing ground compared to traditional indoor simulations.Moving forward, we plan to expand the platform to support a broader range of embodied agents and facilitate large-scale, multi-agent simulations. Our goal is to enhance both the complexity and scale of the environment, allowing for more diverse and extensive embodied agent interactions, ultimately driving advancements in real-world deployment scenarios.

Diffusion models have demonstrated powerful performance across various domains, such as image generation, speech synthesis, and time series forecasting. In the field of time series forecasting, numerous researchers have proposed various methods to further enhance the predictive performance of diffusion models. However, these approaches essentially rely on neural networks to construct a Markov chain mapping from a pure Gaussian distribution to a conditional probability distribution. This raises two intriguing questions: Can we achieve mappings between non-Gaussian distributions? And can neural networks autonomously learn diffusion strategies? The Schrödinger Bridge model might offer a promising direction for exploration. By introducing nonlinear terms into the diffusion equation, the Schrödinger Bridge model enables more flexible distribution mappings. In this presentation, I will briefly share the current research on Schrödinger Bridge models, present the architecture of the Schrödinger Bridge model we have developed, and analyze the issues encountered in preliminary experiments along with potential solutions.

Dynamical systems are ubiquitous in various fields such as physics, engineering, and biology, where the prediction of their behavior are crucial for scientific research and practical applications. With the exponential growth of data and advancements in computational power, foundation models have demonstrated remarkable generalization and adaptability in areas like natural language processing and computer vision. This presentation explores the potential and methodologies for developing foundation models tailored to dynamical systems. We begin by reviewing begins by reviewing existing foundation models applied to dynamical systems, highlighting their architectures and achievements. We then identify key modeling challenges such as handling high-dimensional data, ensuring interpretability, and incorporating physical laws. We explore how foundation models can address classical problems in dynamical systems. Finally, we summarize the findings and discuss future research directions aimed at developing foundation models for dynamical systems, fostering advancements in both theoretical research and engineering applications.

Due to the wide range of recommended scenarios in local life recommendation, including food, entertainment, and maintenance, and the close correlation between user needs, location, and time, diverse spatiotemporal patterns are presented. The traditional recommendation system's unified approach or manual design makes it difficult to dynamically capture subtle changes in different scenarios. LLM, with its ability to understand and reason about context and spatiotemporal factors, extracts and summarizes multi-level spatiotemporal patterns from spatiotemporal data, identifying patterns and trends in user preferences that fit the daily rhythm of users over time and place, and enhances local life recommendations.

Polarization has become a widespread global trend, permeating various aspects of our society. Prior studies have mainly centered on traditional social media platforms, e.g., Facebook and Twitter. Their studied platforms have built their information ecosystems mainly around social networks, which are naturally organized into echo chambers that foster the formation of polarization. However, the rapid growth of AI has transformed the landscape of online social media, gradually replacing traditional platforms with emerging short video platforms. On these short video platforms, people are no longer directly connected, but instead, gathered through algorithmic recommendations, challenging the very existence of echo chambers, i.e., the traditionally recognized foundation of polarization. Currently, our understandings of whether and how polarization emerges on these platforms remain largely limited. Therefore, we perform a large-scale empirical study on one of China's most prominent short video platforms. By analyzing the opinions of over 15 million users on five major social issues, we highlight the role of AI-driven recommendation algorithms in fueling the polarization of public opinions. Specifically, the AI-driven recommendation algorithm contributes to two primary drivers of this polarization: (i) echo chambers, where the algorithmic gathering of like-minded users reinforces their shared similar opinions, and (ii) battlefields, where opposing users conflict, strengthening pre-existing opinions. Our study not only uncovers the polarization of public opinions on the short video platform and its algorithmic drivers, but also offers implications for reducing polarization on increasingly AI-dependent social media.

In recent years, foundational models have revolutionized the fields of language and vision, demonstrating remarkable abilities in understanding and generating complex data; however, similar advances in user behavior modeling have been limited, largely due to the complexity of behavioral data and the challenges involved in capturing intricate temporal and contextual relationships inherent in user activities. To address this, we propose BehaviorGPT, a foundational model designed specifically for large-scale user behavior prediction. Leveraging transformer-based architecture and a novel pre-training paradigm, BehaviorGPT is trained on vast user behavior datasets, allowing it to learn complex behavior patterns and support a range of downstream tasks, including next behavior prediction, long-term generation, and cross-domain adaptation. Our approach introduces a pre-training paradigm tailored for user behavior data, which improves model generalization and transferability across tasks by equitably modeling both head and tail behaviors. Extensive experiments on real-world datasets demonstrate that BehaviorGPT outperforms state-of-the-art baselines in key metrics, showcasing its ability to effectively capture and predict user behavior. Furthermore, we preliminary explore the scaling phenomenon in user behavior modeling, providing insights into how model performance scales with increased data and parameter sizes.

In developing our large-scale social simulator, we aim to create an efficient system for multi-domain simulation. Our system is built in a modular way, which provides the important flexibility for responding to the changing needs of cities and integrating future technologies. Additionally, our parallel distributed design based on open source framework Ray is capable of handling a large number of simulated entities. Our current work focuses on combining traffic and economic simulations within a distributed system design, thus providing a better understanding of how traffic patterns and economic behavior interact. We are working to provide a useful tool for social science researchers that can help them see how different aspects of urban environments are connected, thus providing new tools for their research paradigm.

Electric Vehicles (EVs) have become increasingly popular as a sustainable mode of transportation, making crucial contributions to emission reduction and climate goals. California has set a mandated EV transition goal that all new cars sold in the state be electric by 2035, meaning 12 million more EVs on California’s roads. The vehicle electrification process in California has rapidly proceeded in recent years but facing strong income inequity. Both EV possession and charging infrastructure are highly concentrated in high-income areas. To assess how charging infrastructure impact vehicle electrification, especially in fulfilling unmet demands, we conduct a state-wide analysis to quantify the effect of building new public charging stations on future EV sales and adoption rate. Our findings reveal income disparities where public charging stations highly improve EV sales in high-income areas while improving EV adoption rates in low-income areas.

An increasing reliance on recommender systems has led to concerns about the creation of filter bubbles on social media, especially on short video platforms like TikTok. However, their formation is still not entirely understood due to the complex dynamics between recommendation algorithms and user feedback. In this paper, we aim to shed light on these dynamics using a large language model-based simulation framework. Our work employs real-world short-video data containing rich video content information and detailed user agents to realistically simulate the recommendation-feedback cycle. Through extensive simulation scenarios, our research shows that LLMs are capable of replicating real-world dynamics between users and the recommender system, shedding light on the mechanisms that give rise to the filter bubble. Furthermore, we conduct several ablation studies investigating the specific factors that give rise to the filter bubble, highlighting the impact of various user characteristics and feedback types. Our proposed simulation framework makes it possible to rapidly test new recommender system designs by using LLM agents for user feedback, allowing for more rapid iteration and validation of strategies prior to deployment. Finally, our work takes a step towards a healthier social media environment by proposing interventions that can be employed in real-world systems to increase content diversity

Universal Basic Income (UBI) remains a controversial policy, with various real-world experiments yielding mixed results on its effectiveness. While some studies highlight its potential to alleviate poverty, others suggest that UBI may reduce work participation, raising concerns about productivity and long-term economic sustainability. To address these inconsistencies and explore the broader good of UBI, we propose an innovative agent-based simulation methodology that uses Large Language Models (LLMs) to role-play individuals within a virtual economy. By combining LLMs with real-world data, our approach simulates complex, human-like behavior and decision-making within the context of UBI policies. It also enables the measurement of non-economic metrics, such as mental health, which are central to the broader social goals of UBI. Through a series of experiments, we demonstrate that this framework can replicate a wide range of real-world UBI trial results, providing valuable insights into the factors that contribute to UBI’s effectiveness. Ultimately, our work shows how LLM-powered simulations can serve as a powerful tool for understanding and optimizing UBI, helping policymakers better design and assess its economic and social impact.

With the advancement of technology, personalized intelligent smartphone assistants are playing an increasingly vital role in people's lives, significantly enhancing user convenience, efficiency, and interactive experiences. Understanding user behavior is a key factor in enabling these assistants to meet personalized needs. By leveraging the language comprehension and reasoning abilities of large language models (LLMs), along with the capability of behavioral sequence models to capture user behavior patterns, we align behavioral language models and behavioral sequence models based on existing research. This enables us to fine-tune the LLM into a multi-functional intelligent assistant that can understand and interpret user behavior.In practice, we draw inspiration from visual-language alignment techniques and propose a two-stage alignment approach. In the first stage, feature alignment is achieved through a sequence reconstruction task, enabling the LLM to understand user behavior sequences. In the second stage, we fine-tune the model for various downstream tasks, including next-intent prediction, explanatory counterfactual reasoning, and user profile summarization, ultimately creating a versatile, behavior-aware, and personalized intelligent assistant.

Traffic congestion is a major issue that impacts urban mobility and efficiency. Developing accurate dynamic models of traffic jams is essential for optimizing traffic management and urban planning. However, existing approaches often focus on constructing models that describe specific traffic phenomena rather than their dynamics, or they derive dynamics models based on assumed mechanisms rather than empirical data, leading to inaccuracies. In this study, we apply a neural-symbolic regression method to urban-scale empirical traffic data to uncover network dynamics formulas that characterize the spread of traffic jams. These discovered dynamics provide new insights into traffic congestion and enhance scientific understanding of boarder flow-based networks. This is a work that has just begun, and any suggestions are warmly welcomed.

The frequency and intensity of natural disasters has significantly increased over the past decades. Facing these possible and unexpected disasters, accurately modeling human mobility become the critical issue for planning effective humanitarian relief, disaster management, and long-term societal reconstruction. In particular, human mobility under big disasters such as rainstorm deviates from its daily routine to a large extent, which makes the task more challenging. However, existing works mainly focus on trajectory generation in normal situations. To tackle this problem, we propose a diffusion model-based method for generating trajectories in disasters to accurately model the impact of disasters on human mobility. Specifically, we utilize spatial interpolation techniques to infer fine-grained rainfall in space from sparse station observations. Then, we employ diffusion models to generate high-fidelity trajectories while integrating detailed disaster information to guide the generation outcomes. This work is ongoing, and we welcome your valuable feedback.

The increasing frequency of extreme events highlights the urgent need to enhance urban resilience, particularly in urban mobility systems, which are critical to maintaining city functions under shocks. Despite its importance, the resilience of urban mobility remains underexplored, with limited empirical and theoretical understanding of the mechanisms behind cities' varied responses to such events. In this study, based on large-scale empirical data from 30 cities, we propose a stochastic dynamics framework to characterize urban mobility resilience through resistance and recovery. Our model predicts distinct recoverable and collapse phases, with a critical point governed by a resistance index. Combining our theory and empirical data, we reveal an inverse relationship between resistance and recovery rates. We discover that this antagonism between resistance and recovery rates is driven by the diversity of urban mobility systems. Our discovery not only has implication on policy making and disaster prevention in urban planning but also extends random graph theory to urban mobility dynamics.

The past decade has witnessed China’s remarkable progress in poverty reduction: in 2021, it announced its success in eradicating extreme poverty. One significant contributor to this success is the continuous construction of roads: as the Chinese saying goes, “the first step to prosperity is to build the road”. However, it remains largely unclear the extent to which road construction would benefit economic growth and thus poverty reduction. Through delving into the policies and actual circumstances of road construction in the poorest counties in China, this study seeks to empirically quantify the relationship between road construction and economic growth and figure out the effectiveness of roads on poverty reduction.

Accurate prediction of mobile traffic, which refers to the network traffic from cellular base stations, is crucial for optimizing network performance and enhancing user experience on web-based ser- vices. However, mobile traffic data is highly non-stationary with fluctuations driven by human activity and environmental changes, resulting in both regular patterns and abrupt variations. Diffu- sion models have shown great promise in modeling such complex temporal dynamics due to their ability to capture the inherent uncertainties. Most existing approaches focus on designing novel denoising networks with conditions but overlook the critical role of noise itself, which may result in sub-optimal performance. In this paper, we introduce a novel perspective by emphasizing the role of noise in the denoising process. Through an in-depth anal- ysis of the noise, we demonstrate that it fundamentally shapes the prediction outcomes of mobile traffic, exhibiting clear and con- sistent patterns. To leverage this, we propose NPDiff, a general framework that incorporates the intrinsic dynamics of mobile traf- fic data as noise priors. By decomposing the noise into prior and residual parts, NPDiff enhances the model’s ability to capture both regular patterns and complex, abrupt variations. Our approach is a general framework that can seamlessly integrate with various state- of-the-art diffusion-based prediction models, boosting prediction accuracy and accelerating training convergence. Extensive experi- ments demonstrate that NPDiff achieves superior performance with an improvement over 30%, offering a new perspective on leverag- ing diffusion models in this domain.

Recent research focuses on advancing the reasoning capabilities and optimization strategies of large language models (LLMs), addressing key challenges in improving their performance and adaptability. This report highlights three significant works in this domain.The first study, Quiet-STaR [1], generalizes the Self-Taught Reasoner (STaR) approach by enabling LLMs to autonomously generate rationales across diverse text. Through token-wise sampling and reinforcement-based learning, this method significantly enhances zero-shot reasoning accuracy on tasks like GSM8K and CommonsenseQA, providing scalable improvements in reasoning. The second work, Scaling Test-Time Compute Optimally [2], explores the trade-offs between model size and inference-time computation. By employing adaptive compute allocation strategies, the study demonstrates that smaller models can outperform larger ones on challenging tasks, offering a resource-efficient pathway for improving LLM performance. The third research, ReST-MCTS [3], introduces a novel tree search-based reinforcement learning framework that integrates per-step process rewards for training. This approach effectively identifies high-quality reasoning traces, outperforming traditional self-training methods in benchmarks like MATH and SciBench.

With the rapid advancement of large language models, they have demonstrated broad applicability across diverse domains. Similarly, these models are increasingly being leveraged in urban computing and research, encompassing tasks such as spatial knowledge learning, trajectory modeling, time series analysis, and other urban challenges. This presentation will explore the topic through the lens of urban challenges, highlighting the critical need for researching and applying large language models. We will then review and synthesize existing methodological paradigms. Finally, we will discuss potential pathways toward achieving advanced urban intelligence as we progress toward general artificial intelligence.

With the rapid development of wireless networks, the availability of diverse channel knowledge data has become crucial for optimizing network performance. Channel knowledge map (CKM) is a site-specific database, which provides locations-specific channel knowledge. Existing CKM dataset primarily focus on simulating large-scale parameter--path loss in outdoor scenarios via ray tracing approach. However, few works have constructed and released small-scale parameter--channel state information (CSI) dataset, which characterizes the knowledge of current conditions of communication channel between the transmitter and receiver in a wireless communication system. Inspired by existing pathloss datasets, we aim to construct a user-centric dynamic channel state information dataset with consideration of multiband and material information. The dataset is designed to support downstream wireless communication channel research, providing high-fidelity channel knowledge in different scenarios.

Policy exploration is critical in training reinforcement learning (RL) agents. Ex- isting RL algorithms employ various approaches for policy exploration, includ- ing the ϵ-greedy method in deep Q-learning, the Gaussian process in DDPG, etc. However, these approaches are designed based on prefixed stochastic processes and are indiscriminately applied in all kinds of RL tasks without considering any environment-specific features that influence policy exploration. Moreover, dur- ing the training process, the evolution of such stochastic process is rigid, which typically only incorporates a decay of the variance. This makes the policy explo- ration unable to adjust flexibly according to the agent’s real-time learning status, limiting the performance. In this paper, we design LLM-Exp, which improves policy exploration in RL training with large language models (LLMs). During the RL training in a given environment, we sample a recent action-reward tra- jectory of the agent and prompt the LLM to analyze the agent’s current policy learning status based on the trajectory and then generate a probability distribu- tion for future policy exploration. We update the probability distribution period- ically and derive a stochastic process that is specialized for the particular envi- ronment and can be dynamically adjusted to adapt to the learning process. Our approach is a simple plug-in design, which is compatible with DQN and any of its variants or improvements. Through extensive experiments on the Atari bench- mark, we demonstrate the capability of LLM-Exp to enhance the performance of RL.

This sharing will provide an overview and summary of the latest research in the field of Large Language Model (LLM) Agents for social simulation. We will explore the state-of-the-art advancements, highlighting key studies, methodologies, and breakthroughs in applying LLMs to model and simulate social behaviors, decision-making, and interactions. In addition to this, we will present the efforts and work we have undertaken in this domain, focusing on the development and application of LLM Agents. Our contributions aim to address critical challenges, such as scalability, the modeling of complex human interactions, and ethical considerations. By sharing our experiences and insights, we hope to demonstrate the potential of LLM Agents to offer valuable perspectives for solving real-world societal issues and advancing the field of social simulation.

Optimization-based energy-saving methods for mobile networks presents significant challenges in real-world deployment. Through collaboration with China's largest mobile network operator, we identified three key challenges: (1) network heterogeneity, (2) constraints related to cell status control and traffic migration, and (3) the need for alignment with parameter-based strategies. To address these, we developed a simulation environment that closely approximates real-world networks. This includes a coverage model to simplify network heterogeneity and an energy consumption model to evaluate the effectiveness of energy-saving strategies. Leveraging detailed energy-saving constraints for cells, provided by the operator, we incorporated these into our simulation environment to develop deployable energy-saving strategies. We also implemented a hyperplane-based classification approach to identify optimal parameters, aligning advanced optimization-based methods with parameter-based strategies for real-world applications. Our strategy was tested in a pilot network in Hangzhou, and the results demonstrated a significant reduction in network energy consumption. Comprehensive analysis further validated the effectiveness of our approach across spatiotemporal patterns of multiple performance metrics.

With the rapid development of the Fifth Generation Mobile Communication Technology (5G) networks, network planning and optimization have become increasingly crucial. Generating high-fidelity network traffic data can preemptively estimate the network demands of mobile users, which holds potential for network operators to improve network performance. However, the data required by existing generation methods is predominantly inaccessible to the public, resulting in a lack of reproducibility for the models and high deployment costs in practice. In this article, we propose an Open data-based Diffusion model for mobile traffic generation (OpenDiff), where a multi-positive contrastive learning algorithm is designed to construct conditional information for the diffusion model using entirely publicly available satellite remote sensing images, Point of Interest (POI), and population data. The conditional information contains relevant human activities in geographical areas, which can effectively guide the generation of network traffic data. We further design an attention-based fusion mechanism to capture the implicit correlations between network traffic and human activity features, enhancing the model’s controllable generation capability.

The fast development of location-based social networks (LBSNs) has led to significant changes in society, resulting in popular studies of using LBSN data for socioeconomic prediction, e.g., regional population and commercial activity estimation. Existing studies design various graphs to model heterogeneous LBSN data, and further apply graph representation learning methods for socioeconomic prediction. However, these approaches heavily rely on heuristic ideas and expertise to extract task-relevant knowledge from diverse data, which may not be optimal for specific tasks. Additionally, they tend to overlook the inherent relationships between different indicators, limiting the prediction accuracy. Motivated by the remarkable abilities of large language models (LLMs) in commonsense reasoning, embedding, and multi-agent collaboration, in this work, we synergize LLM agents and knowledge graph for socioeconomic prediction. We first construct a location-based knowledge graph (LBKG) to integrate multi-sourced LBSN data. Then we leverage the reasoning power of LLM agent to identify relevant meta-paths in the LBKG for each type of socioeconomic prediction task, and design a semantic-guided attention module for knowledge fusion with meta-paths. Moreover, we introduce a cross-task communication mechanism to further enhance performance by enabling knowledge sharing across tasks at both LLM agent and KG levels. On the one hand, the LLM agents for different tasks collaborate to generate more diverse and comprehensive meta-paths. On the other hand, the embeddings from different tasks are adaptively merged for better socioeconomic prediction. Experiments on two datasets demonstrate the effectiveness of the synergistic design between LLM and KG, providing insights for information sharing across socioeconomic prediction tasks.

Our city simulator has had some success in microscopic simulations of the physical domain of cities. Based on the success of LLMs in recent years, we will further realize the simulation of human society in the city based on LLM agent technology. By combining short-term simulation with long-term evolution, we hope to achieve the simulation of economic behavior, social behavior, cognition, psychology, etc. of urban participants such as citizens, enterprises, and governments. In this proposal, we will design a framework for programming LLM agents, a distributed parallel system for massive LLM agents, a multi-domain simulation system, and a set of ideas for combining short-term simulation and long-term evolution, thus providing new tools for social science research.

State representation plays a critical role in reinforcement learning. Beyond the reward signal, many existing works incorporate auxiliary self-supervised losses that leverage additional information to aid in learning accurate state representations. However, current self-supervised losses are entirely dependent on expert knowledge and the understanding of the target environment. Once designed, these losses are fixed and cannot be dynamically adjusted based on feedback from the environment. This hand-crafted design of losses is both inflexible and suboptimal. Recently, the emergence of large language model (LLM) has provided promising alternatives to prior knowledge injection, requiring minimal human intervention. Motivated by this, we propose LLM for Automatic Self-supervised representation in Reinforcement Learning (LASR), a novel approach that utilizes LLMs to autonomously design self-supervised losses, thereby enhancing state representation and improving training efficiency. We conduct experiments on discrete-action benchmarks such as Atari and continuous control tasks, such as the DMC benchmark. This work is ongoing, and we welcome your valuable feedback.

In the era of information overload, recommendation algorithms play a critical role in filtering content for users, yet they often lead to the formation of information cocoons, where users are exposed to increasingly homogeneous content, resulting in negative consequences such as addiction and opinion polarization, therefore, early prediction of information cocoons and timely implementation of intervention measures is crucial. In this work, we labeled the users' information cocoons status based on their viewing data from a short-video platform, and designed a model to predict the labels based on the early viewing data of users who have just entered the platform, utilizing feature interactions and temporal changes simultaneously. Our approach, validated on real-world industrial datasets, outperforms baseline models and proves to be an effective practical tool for mitigating the risks associated with information cocoons.

Automating agent design using Large Language Models (LLMs) is at the forefront of current AI research, offering new methodologies to enhance flexibility, adaptability, and task efficiency. This research sharing will explore five recent innovations in the automated agent domain: OPRO, Promptbreeder, GPTSwarm, ADAS, and AgentSquare. Each of these approaches highlights distinct strategies, from gradient-free optimization with natural language prompts to evolutionary self-improvement of prompts, computational graph-based multi-agent collaboration, automated agent generation in code space, and modular frameworks for task-specific agent assembly. By comparing their principles, unique design features, and applications, this sharing will provide a comprehensive overview of the possibilities and advancements in automated LLM agent design.

While urban residents appear to navigate freely in cities, they encounter ``invisible walls''—subtle cultural, socioeconomic, and racial-ethnic barriers that impede access to specific locales, leading to experienced social segregation that extends beyond mere physical distances. The abundance of visual and textual data on location-based social networks (LBSN) encapsulates nuanced perceptions and feelings associated with different urban places, potentially helpful for predicting segregation. Considering the extensive and intricate data and our goal to generate insightful explanations for segregation, we've developed a large language model (LLM) workflow that includes Open-Coding, Rate-Generalization, Finetuning and MLP-Reinforcement. Our study marks an important step toward understanding implicit social inequalities through web data analysis, paving the way for more inclusive urban planning and community policy-making. This work is nearly complete, so any suggestions and discussions are warmly welcomed!

Living needs refer to the needs people generate in their daily lives for survival and well-being. On life service platform like Meituan, user purchases are driven by living needs, making accurate living needs predictions crucial for personalized service recommendations. Traditional approaches treat living needs prediction as a closed-set classification problem, limiting their ability to handle the complexity of human living needs. In this work, we redefine living needs prediction as an open set problem, where needs are not restricted to predefined categories. We introduce PIGEON, a novel system that leverages large language models for open set living needs prediction. Our system first employs graph neural networks (GNNs) to encode user behaviors and retrieve relevant historical data, guiding LLMs to generate need predictions. Next, LLMs incorporate Maslow's hierarchy of needs to structure and refine initial predictions, aligning them more closely with real-world human needs. Finally, our system improves service retrieval by using a fine-tuned text embedding model to manage diverse LLM-generated need descriptions. Experimental results demonstrate that PIGEON outperforms closed-set classification models and LLM-based baselines. More broadly, we extend open set classification from methods still constrained by limited categories to support limitless class predictions.

The increasing saturation of social media, particularly short-video platforms, has raised concerns about excessive usage and potential addiction. As short-form videos become an ingrained part of daily routines, concerns about their impact on people have intensified. Prior research has often relied on psychological questionnaires or device monitoring to study digital addiction, yet it falls short in assessing or predicting addiction using solely online behavioral data and lacks a comprehensive understanding of how natural in-app behaviors can be leveraged for this purpose. This study addresses that gap by combining psychometric evaluations with online user activity data to identify key indicators of short-video addiction and gain a deep understanding of its evolving trajectory. We collect data from 483 users, categorizing them into three addiction levels—severely addicted, moderately addicted, and non-addicted—using a tailored psychometric survey. By integrating this data with users’ in-app activities, we create an addiction-labeled dataset, identify critical behavioral indicators, and use a classification model to expand the dataset. With a platform-wide addiction-labeled user population, we conduct temporal and feature analyses to further explore the dynamic patterns of addiction. Our preliminary findings reveal the difficulty in distinguishing between addicted and non-addicted users solely based on online behavioral patterns. However, clearer distinctions emerge when differentiating between severely addicted and the rest of the users. We aim to contribute to the development of standardized criteria for assessing online addiction by linking digital footprints with psychological markers. Our proposed framework will provide actionable insights for developers and policymakers, allowing them to proactively address addiction and promote healthier, more sustainable digital engagement.

Complex systems with nonlinear dynamics pose significant challenges for finite time optimal control, especially when accurate system models are unavailable. This paper introduces DIFOCON (DIffusion Finite-time Optimal CONtrol), a novel data-driven framework for finite-time optimal control that operates without prior knowledge of system parameters or dynamics. DIFOCON reformulates the control problem as a generative task, optimizing control signal trajectories to guide systems to target states within a finite time. Our approach utilizes a diffusion model with a dual-Unet architecture to capture nonlinear system dynamics and generate entire control sequences in a single step. Additionally, an inverse dynamics module is integrated to ensure that the generated control signals are appropriate for complex systems. To further enhance performance, we propose a retraining strategy that improves out-of-distribution generalization. Experiments on two nonlinear complex systems demonstrate DIFOCON’s superior performance.

Multimodal large language models (MLLMs) have exhibited remarkable capabilities for performing complex vision-language tasks in various domains. Currently, MLLMs in urban studies are only developed focusing on remote sensing imagery. However, except for the macroscopic information from remote sensing imagery, effective urban understanding also requires detailed appearance information of urban zones from street-view imagery, which is largely overlooked by existing MLLMs. The primary challenges of developing such a versatile urban MLLM are twofold. Firstly, it needs a large-scale corpus with well-organized, cross-view urban imagery paired with corresponding text for cross-modal training. Secondly, traditional MLLMs typically learn image-text pairs independently, hard to support joint modeling of cross-view urban imagery. To address these challenges, in this work, we propose UrbanMLLM, a novel MLLM that jointly learns from remote sensing and street-view imagery to harness their complementary information. We first collect a large-scale dataset containing satellite-view and street-view imagery along with their geotags and annotated texts. Technically, we propose a brand MLLM architecture with a cross-view perceiver to explicitly connect visual information of cross-view urban imagery. We also introduce a novel pre-training paradigm based on structural interleaved urban image-text documents integrating satellite-view, street-view imagery and related textual descriptions. This approach encourages the model to implicitly learn the relationships between different types of urban imagery, enhancing the understanding in each domain. We evaluate our model on a comprehensive benchmark including 13 types of urban understanding tasks across satellite-view, street-view and cross-view domains. Extensive experiments demonstrate that UrbanMLLM achieves an average of 27.3% and 25.5% performance improvement compared with the best open-sourced and closed-sourced MLLMs, respectively. Moreover, we thoroughly study the impact of different pre-training data choices and model scales on performance, offering practical insights for effective MLLM design. The proposed UrbanMLLM offers a scalable and versatile solution for understanding urban environments.

Data-driven time series analysis and physics-driven dynamical system modeling have always been the focus of research in the field of engineering science. This report focuses on the connection between the two, using the theory of dynamical system embedding theory to construct physics-guided time series modeling, thereby understanding the underlying dynamics of time series in the most fundamental way. Compared to traditional data-driven modeling methods, physics-guided approaches significantly improve downstream task accuracy, generalization, interpretability, and efficiency. This report include the following papers: 1. Toward Physics-guided Time Series Embedding 2. Time-SSM: Simplifying and Unifying State Space Models for Time Series Forecasting 3. Attractor memory for long-term time series forecasting: A chaos perspective (NIPS2024) BIO Hu Jiaxi is a first year Ph.D. student at the Hong Kong University of Science and Technology (Guangzhou). His main research interests include time series analysis, state space models, and physics-guided machine learning. He has served as a reviewer of major AI/ML/DM conferences.

Generating origin-destination (OD) data for cities without existing OD information is a critical task for urban planning and transportation analysis. Traditional data-driven OD generation methods typically attempt to model OD flows in a single step. However, due to the high variance and inherent instability of OD flows, this approach makes learning highly challenging. In contrast, the classical four-step transportation model first generates the outflow and then distributes it to various destinations, breaking down the complex task into two steps. While this method can produce more stable results, it introduces cascading errors due to the separation of the steps. To address these issues, we propose an end-to-end graph diffusion-based model that integrates both trip generation and distribution into a unified, data-driven framework. This approach mitigates the challenges of instability and error accumulation, offering a more robust solution for OD generation.

In an increasingly urbanized world, cities are experiencing growing segregation and inequality. In particular, segregation affects access to essential resources like housing, community facilities, healthcare, and a clean environment. To study urban segregation, we plan to use simulation methods. For over 50 years, agent-based models (ABMs) have been used to explore how individual behaviors lead to complex societal patterns in urban spaces. Recently, the rise of large language models (LLMs) has enabled the development of LLM agents that can simulate urban activities with remarkable realism. However, the high computational cost of LLMs makes scaling up these simulations challenging. Fortunately, we've developed OpenCity, a scalable platform for simulating urban activities with large numbers of LLM agents. With this platform, we can create realistic large-scale simulations of urban segregation. In this talk, I will discuss my upcoming research plan, which includes collecting segregation datasets from various contexts, designing LLM agents for simulations, and exploring the intrinsic biases in language models.

The optimization of energy efficiency (EE) is crucial in large-scale mobile networks, where additional metrics such as network coverage and service quality also need to be considered. In practical network scenarios, different performance metrics often require a balance that aligns with various preferences, and traditional methods fall short in providing a unified solution for these conflicting objectives. We address this challenge using a multi-agent reinforcement learning approach based on graph neural networks to achieve a pareto front for the multi-objective optimization problem of energy efficiency and network performance. By simultaneously optimizing both transmit power and antenna configuration, our method dynamically adjusts network parameters, aiming to reduce energy consumption while enhancing coverage quality. We specifically tackle the challenge of large-scale base station optimization using multi-agent reinforcement learning; we employ RGCN to manage the complex state space of large-scale mobile networks, and utilize the Envelop exploration method to balance multiple objectives effectively. Experimental results demonstrate that, compared to traditional multi-objective optimization methods, our approach achieves better exploration of the pareto front, providing superior trade-offs between energy efficiency and service quality, thereby enhancing overall network sustainability and adaptability.

Recently, the concept of next-token-prediction has marked a new era in general artificial intelligence, excelling in processing data across multiple scenarios, domains, and even modalities. The most successful examples are the large language models (LLMs) and multimodal large languange models (MLLMs) . Furthermore, the concept has also been applied to decision-making tasks directly. For instance, Reed et al. (2022) developed a generalist agent to handle different control environments simultaneously, such as Atari games and robot benchmarks. We summarize recent works for generic decision making in such a scheme. We also introduce our recent work, solving diverse combinatorial optimization problems via a unified model.

User mobile traffic is crucial for network simulation, including user behavior analysis, network performance optimization, and smart city construction. However, the application usage patterns of different users have not received sufficient attention in related research. In this study, we aim to analyze the impact of user behavior and the surrounding environment on the application-level traffic generated by different users. Additionally, we have developed a structure that simultaneously establishes user profiles and the surrounding environment, which is used as conditional input to a conditional diffusion model. Preliminary results indicate that considering these two factors can significantly improve the quality of the generated user mobile traffic.

During extreme disasters like hurricanes, cold waves, and wildfires, human mobility patterns in urban areas undergo significant changes due to factors like inaccessible transportation, infrastructure damage, and shortages of essential supplies. Different disaster types have varying impacts on human mobility. This study aims to explore common characteristics in the changes of human mobility patterns under different disaster types. A dynamic model of human mobility networks was constructed, describing both node status updates and the evolution of network edges. By identifying functional patterns from empirical data, the dynamic model was refined to simulate the evolution of human mobility networks during disasters, capturing the variations in human mobility patterns before, during, and after disasters. This model not only helps reveal the mechanisms by which disasters affect human mobility patterns but also provides insights into the intrinsic changes in human mobility under extreme disasters.

Location-based knowledge graph, a mainstream model in LSBN with multiple types of nodes and links, however, has limitation in the feasibility of conventional network embedding techniques. In this seminar, two works about dealing with heterogeneous networks in LSBN will be shared.The first one [1] explores the performance differences of GNNs when dealing with nodes with mixture structural patterns. It showed GNNs' effectiveness in capturing structural patterns on both homophilic and certain heterophilic graphs while struggling on the opposite node set, eg, homophilic nodes in heterophilic graphs.The second one [2] investigate the embedding function in social networks. It accommodate the fact that (1) homophily results in cycles and (2) influence results in hierarchies in networks. This study models the link as a mixture outcome of both node homophily and social influence, rather than a simple function of the corresponding node embedding.

Mobile traffic data is crucial for operators to effectively plan and deploy network resources. As they are hard to access in practice, generating mobile user traffic serves as an effective method to assist in those areas. Mobile traffic patterns vary according to urban contexts, especially across different functional areas. As users’ mobility dynamically shifts within these areas, it is important to explore the relationships between user mobility and mobile traffic volume. However, current research primarily focuses on city-scale traffic generation and often relies on a static understanding of urban contexts. Additionally, there has been limited exploration of how users' trajectories and urban contexts jointly influence mobile traffic. To address this gap, we propose a conditional diffusion network that collaboratively generates user trajectories and traffic patterns, enhancing the understanding of mobile users’ behaviour and improving the performance of traffic generation in urban environments. This is an ongoing work.

Urban mobility data play a crucial role in various fields such as urban planning and traffic management. However, the collection of real mobility data is expensive and raises significant privacy concerns, making the generation of mobility data essential. In recent years, diffusion models have gained great attention in the data generation domain, with multiple works applying them to urban mobility generation tasks, achieving promising results in individual trajectory modeling. However, existing methods often sample noise in an i.i.d manner and generate data at the level of single trajectories, leading to generated data that fails to accurately reflect the collective flow characteristics of real-world data. By analyzing the process through which diffusion models generate individual trajectories from noise, we find a strong spatiotemporal correlation between the initial sampled noise and the generated trajectories. Thus, unlike traditional independent sampling methods, we propose a spatiotemporal collaborative noise sampler, which models collective flow patterns into individual trajectory generation through a noise prior approach.

The phase transitions of complex systems are related to many major issues of human concern, such as drastic global climate changes, ecosystem collapses, emergence of pandemics, and systemic risks in financial markets. Predicting the critical points of complex systems is of common interests across many scientific fields. However, theoretically solving the exact critical points of nonlinear complex systems is extremely difficult. Recently, we have rigorously derived the universal complex system dynamics of arbitrary cooperation networks, and found that the critical point of network collapse lies between the network’s maximum k-coreness and maximum eigenvalue, providing a theoretical basis for regulating the resilience of complex systems. We analyzed the trade network resilience of Belt and Road countries using global trade network data and found that the economic ecological resilience of these countries shows a trend of oscillating upward trend. Additionally, we provide specific recommendations for enhancing the economic ecological resilience among Belt and Road countries.

Many dynamic processes across various disciplines exhibit multi-timescale characteristics, such as ecological evolution and protein folding. Modeling the fundamental principles behind these multi-timescale phenomena is crucial for understanding and simulating these systems. Despite extensive efforts to develop efficient reduced-order models for accurate simulation, they often overlook the fact that dynamic processes occur on the system's inherent energy landscape. The low-energy states, acting as attractors, dominate the slow dynamics of long-term system evolution. The gap between theoretical methods and real-world mechanisms results in deficiencies in the interpretability and accuracy of existing approaches. In this work, we learn the latent energy landscape to guide the accurate simulation of multi-timescale dynamic processes. Specifically, we first predict the state energy, and then incorporate the energy as inductive bias by designing a graph neural diffusion equation to model the Fokker-Planck flow of the state probability distribution. Experimental results demonstrate that the proposed method can reliably predict system energy and improve simulation accuracy by about 30% compared to SOTA methods.

Mobile traffic data is crucial for operators to effectively plan and deploy network resources. Users exhibit highly intricate mobile traffic patterns driven by their app usage habits, making it important to explore the underlying intentions behind mobile traffic. Recently, large language models (LLMs) have been increasingly used to predict human behavior due to their strong reasoning capabilities. However, their potential to assist in generating individual-level mobile traffic remains unexplored. In this work, we propose a framework to generate individual-level daily mobile traffic with the assistance of LLM agent. First, the LLM agent generates physical and cyber intention sequences for each user. Second, to ensure the accuracy of the LLM-generated intentions, we incorporate two validation mechanisms: a Traffic-Cyber and a Cyber-Physical Retrieval-Based Validation Mechanism. Finally, the physical and cyber intention sequences are encoded and input into a self-designed spectral-temporal conditional diffusion model, which generates both the general patterns and residuals of mobile traffic. This is ongoing work.

With the emergence of edge AI applications such as AI smartphones and personal AI assistants, enhancing the reasoning performance of large language models (LLMs) on edge devices has become an increasingly important topic. Due to the limited capacity of edge devices, cloud collaboration is a promising solution. It involves running smaller-scale LLMs locally and calling larger-scale LLMs on the cloud in real-time, achieving a balance between efficiency and performance through their collaboration. We propose a novel collaborative reasoning framework that first decomposes the overall task into multiple subtasks and constructs a subtask dependency graph based on the logical dependencies between them. Subsequently, based on the difficulty of the subtasks and the dependency relationships within the graph structure, the framework designates either cloud or local models to perform subtask reasoning. This is an ongoing work.

In recent years, with the frequent occurrence of natural disasters, predicting human mobility under disasters has become more and more important. This is of great significance for post-disaster resource allocation, individual-level mobility warning and guiding the designated emergency response of the impending disaster city. However, because the human mobility pattern in disaster scenarios is different from that in normal scenarios, it is difficult for the existing normal human mobility prediction models to handle it in disaster scenarios. In addition, this task still faces the problem that the target city has little mobility data available under a particular disaster situation beyond experience. To address these issues, we designed a method that combines knowledge across cities and disasters at the intention level and uses LLMs to predict the mobility of an individual. This is an ongoing work.

The downlink throughput is a crucial indicator of communication quality and user experience in wireless communication systems. Accurately predicting this rate helps optimize resource allocation and enhance user experience in dynamic and complex environments. Currently, channel estimation methods are commonly used in the wireless communication field to obtain channel state information, which enables the modeling of the information transmission process and estimates the theoretical upper limit of downlink throughput. However, channel estimation often relies on real-time carrier frequency detection, making it difficult to achieve long-term prediction and generation. Additionally, these methods often lack accuracy and generalization capabilities when dealing with highly dynamic scenarios and complex urban environments. To address these issues, we prepare to propose physics-informed diffusion generation models that accurately generates downlink throughput over long time spans based on user trajectories and base station information in urban environments. The model employs a modular design and end-to-end training, sequentially generating latent variables for RSRP and SINR, and ultimately producing the downlink throughput. We utilize a teacher-student joint training architecture, where the diffusion model is pre-trained using results calculated from physical formulas and then fine-tuned on real datasets. This approach integrates expert knowledge into the diffusion model's generation process, improving generation accuracy while accelerating model parameter convergence, and allows for adaptive performance in various urban scenarios. Additionally, this method overcomes the limitations of traditional channel estimation methods, which struggle to achieve long-term predictions, thereby providing guidance for optimal base station deployment and resource allocation.

This paper presents a novel approach to enhancing the performance of large models across a diverse range of downstream tasks by utilizing an interleaved pretraining strategy on joint satellite, street-view, and text data. By integrating these three complementary data modalities, we enable models to capture both the detailed visual information from street-level images and the broad spatial context provided by satellite imagery, while also incorporating the rich semantic knowledge from textual data. This unified pretraining framework significantly improves model performance in a various of tasks such as object detection, geographic location inference, spatial relationship reasoning, socioeconomic indicator prediction, and scene understanding. To support this approach, we have constructed an extensive and diverse dataset, tailored to the specific challenges of these tasks, paving the way for more accurate and context-aware applications in urban planning, environmental monitoring, and socioeconomic analysis.

Human mobility prediction plays a crucial role in various real-world applications. Although deep learning based models have shown promising results over the past decade, their reliance on extensive private mobility data for training and their inability to perform zero-shot predictions, have hindered further advancements. Recently, attempts have been made to apply large language models (LLMs) to mobility prediction task. However, their performance has been constrained by the absence of a systematic design of workflow, as they directly generate the final output by LLM. In this paper, we introduce AgentMove, a systematic agentic prediction framework to achieve generalized mobility prediction for any cities worldwide. In AgentMove, we first decompose the mobility prediction task into three subtasks and then design corresponding modules to complete these subtasks, including spatial-temporal memory for individual mobility pattern mining, world knowledge generator for modeling the effects of urban structure and collective knowledge extractor for capturing the shared patterns among population. Finally, we combine the results of three modules and conduct a reasoning step to generate the final predictions. Extensive experiments on mobility data from two sources in 12 cities demonstrate that AgentMove outperforms the best baseline more than 8% in various metrics and it shows robust predictions with various LLMs as base and also less geographical bias across cities。

Reinforcement learning (RL), an important paradigm for urban decision-making, typically requires simulation to evaluate its effects, such as residential mobility in spatial planning and disease spread in epidemic control. Due to the large number of simulated urban entities and their complex dynamics, simulation time often takes several minutes, creating a significant bottleneck since RL agents generate solutions in milliseconds. We propose an efficient urban decision-making framework which incorporates a reward model to bridge the gap between the fast RL agent and the slow urban simulation. By approximating simulated results with neural networks, our framework allows the RL agent to interact with the reward model in real time without delays while the slow simulation operates offline, addressing the simulation time bottleneck. We further introduce adaptive reward modeling that continuously fine-tunes the reward model with the latest simulation data to align with the agent's policy, ensuring accurate reward approximation and avoiding outdated decision evaluation. Extensive experiments on urban spatial planning and epidemic control demonstrate that our framework not only accelerates the training process by over 10,000 times but also achieves state-of-the-art decision performance with over 12.6% improvements compared to existing approaches.

Identifying urban land-use types is crucial for effective resource management, urban planning, and sustainable development. However, classifying land use is complex due to the complexity of the city and the poor data available in undeveloped areas. In this work, we present the Multi-modal Model for Land-use Classification (M3LUC). Our model is the first to leverage the advanced Vision-Language Model (VLM) to better capture urban functionality through remote sensing data and Points of Interest (POI). We have also designed specific mechanisms to robustly and extensively tackle the modality missing and conflict to enhance transferability. Experiments conducted in four major cities in China demonstrate our model’s superior performance in both transfer and non-transfer tasks, revealing its potential for broader applications.

The rapid evolution of LLM-based agents has led to a growing number of diverse architectures as well as novel applications. However, current research focuses on proposing and evaluating specific architecture designs of LLM agents, while lacking a general design principle for practice. Additionally, LLM agent designs are often dependent on targeted tasks, yet few efforts have been made to understand how to quickly find the best agent configuration for a novel task. In this work, we first define a unified design space for LLM agents consisting of a Cartesian product of four design dimensions including the module of planning, reasoning, memory and tool use, producing hundreds of different designs. Based on this, we further develop a hierarchical architecture search framework for automatic agent architecture optimization given specific tasks. Experiments on several representative tasks from different domains demonstrate the superiority of discovered agent architectures. Overall, our work offers a principled and scalable approach to transition from studying individual LLM agent designs for specific tasks, to systematically studying the modularized LLM agent design space, together with a practical way for automatic agent architecture discovering.

LLMs have become a crucial tool for many aspects of day-to-day life. However, their ability to fully capture the intricacies of human communication is limited due to the bias present in their training corpora. In light of the global usage of LLM chatbots, it is important to quantify whether these tools can effectively adapt to users' needs by providing appropriate responses according to different cultures. While a prominent, universal Western bias has been discovered within many models, our work investigates the presence of a relative intra-language misalignment, or the dominance of one culture within a given language representation. We evaluate culture-language misalignment using a six-dimension survey to assess the cultural state of two types of agents: culturally displaced agents and regional language agents, and then conduct a comprehensive analysis of the alignment within the two types. Through extensive experiments, we find that some language representations have a dominant culture that transcends agent nationality for both displaced and regionalized agents, and that certain cultural dimensions are more responsible for language-culture misalignment than others.