Origin-Destination (OD) flows are essential for modeling urban mobility and traffic planning. However, existing OD generation methods typically overlook the influence of external events, limiting their ability to support scenario-driven or conditional simulation tasks. To this end, we propose EventOD, a conditional diffusion framework that enables controllable OD flow generation based on natural language descriptions of events. The framework uses a large language model to extract structured control conditions, which are then injected into a diffusion model to guide OD generation. To overcome the lack of real event-OD pairs, we synthesize training data by perturbing existing OD matrices and apply a mixed training strategy combining regular and event-conditioned samples. The generated flows can be evaluated using a traffic simulator to verify their consistency with intended event impacts. This is an ongoing study, with core components implemented and full evaluation planned in future work.

Urban land use is closely tied to carbon emissions, accounting for one-third of human-induced emissions. However, current urban planning practices primarily focus on functional layout metrics while overlooking carbon mitigation goals, leaving substantial potential for promoting low-carbon cities untapped. In this work, we propose a deep reinforcement learning-based urban planning framework that integrates carbon mitigation considerations. By combining fine-grained carbon accounting with resident trajectory simulation, the model assesses indoor and transportation emissions as well as carbon sink from green spaces. A multi-objective reward mechanism balances traditional planning goals with emission reduction. Experimental results demonstrate superior performance in both emission metrics and urban functionality, highlighting the framework's potential for advancing sustainable urban development. Notably, we find that planning-focused decisions tend to favor dispersed layouts, while emissions-focused optimization leads to centralized, high-density distributions of non-residential buildings. Our method effectively reconciles these trade-offs, achieving a balanced solution that performs well on both critical objectives.

Since the explosive success of GPT-3, research into the internal mechanisms of large language models (LLMs) has never ceased. Researchers have approached this question from multiple angles, trying to reveal how these models learn linguistic patterns and knowledge from massive-scale data: some studies focus on the role of the attention mechanism, exploring how different attention heads specialize in syntax, semantics, and even reasoning tasks; others analyze the intermediate representations to understand what abstract features are captured at different layers; meanwhile, some scholars investigate the training dynamics under various fine-tuning strategies, aiming to uncover how different training paradigms shape the models’ behavior patterns. At the same time, a significant body of work seeks to connect model behavior to the training data itself, examining how specific training samples influence the model’s outputs. From theoretical analyses to empirical studies, these efforts continue to converge on a central question: how do LLMs accumulate simple gradient updates into complex reasoning abilities and world knowledge? In this presentation, I will briefly review recent advances in understanding LLM mechanisms. Building on existing work, and based on observations of decoder-only LLM architectures, we propose a new method for investigating internal mechanisms, designed to reveal how different training paradigms (RL vs. non-RL) impact the internals of LLMs and to identify common patterns across different LLM families. As this research is still at an early stage, we warmly welcome any questions, critiques, or suggestions regarding our approach.

In this seminar, I will present my research related to RL and VLMs. Regarding VLMs for RL, model ensemble is a useful approach in reinforcement learning (RL) for training effective agents. Despite the wide success of RL, training effective agents remains difficult due to the multitude of factors requiring careful tuning, such as algorithm selection, hyperparameter settings, and even random seed choices, all of which can significantly influence an agent’s performance. Model ensemble helps overcome this challenge by combining multiple weak agents into a single, more powerful one, enhancing overall performance. However, existing ensemble methods, such as majority voting and Boltzmann addition, are designed as fixed strategies and lack a semantic understanding of specific tasks, limiting their adaptability and effectiveness. To address this, we propose LLM-Ens, a novel approach that enhances RL model ensemble with task-specific semantic understandings driven by large language models (LLMs). Given a task, we first design an LLM to categorize states in this task into distinct “situations”, incorporating high-level descriptions of the task conditions. Then, we statistically analyze the strengths and weaknesses of each individual agent to be used in the ensemble in each situation. During the inference time, LLM-Ens dynamically identifies the changing task situation and switches to the agent that performs best in the current situation, ensuring dynamic model selection in the evolving task condition. Our approach is designed to be compatible with agents trained with different random seeds, hyperparameter settings, and various RL algorithms. Extensive experiments on the Atari benchmark show that LLM-Ens significantly improves the RL model ensemble, surpassing well-known baselines by up to 20.9%. In terms of RL for VLMs, I plan to use RL fine-tuning methods to mitigate the hallucination problem in VLMs.

Efficient optimization of cellular network configurations is important for satisfying increasingly demanding performance requirements in dynamic wireless environments. Reinforcement learning-based methods have shown promise in automating network configurations; however, their practical adoption is limited by poor transferability and weak generalization across diverse real-world scenarios. To overcome these limitations, we propose a novel optimization framework that employs domain randomization during pre-training of a diffusion-based generative policy, significantly enhancing adaptability and robustness. Specifically, we introduce an environment-aware embedding prediction network that encodes meaningful representations from randomized environmental parameters. Additionally, our framework models the network state as an attributed bipartite graph, effectively capturing intricate interactions between base stations and mobile users. Leveraging reinforcement learning-guided exploration strategies, our enhanced diffusion model efficiently explores high-dimensional, previously unseen configuration spaces. This framework thus offers a scalable, highly generalizable solution, advancing cellular network optimization across diverse deployment contexts. Extensive experiments conducted in diverse simulated network scenarios demonstrate that our proposed method achieves superior performance.

World models have shown great promise in modeling visual dynamics and generating furtue frames in complex environments. However, maintaining consistency over long-range video generation remains a significant challenge due to the lack of stable spatial memory and structural understanding. In this work, we propose to enhance the temporal and spatial consistency of world models by integrating features derived from 3D Gaussian Splatting (3DGS). As a neural-free, high-fidelity 3D representation, 3DGS captures fine-grained geometric and appearance information across views. We leverage its reconstructed features as spatial memory to guide the rendering process of world models.

The development of a comprehensive social agent that accurately replicates human behavior across various domains is crucial for advancing simulations of social life. In this paper, we propose a novel framework for constructing social agents that are driven by a combination of behavioral science theories. This approach enables modeling human behavior across multiple dimensions, including mobility, social interaction, economic decision-making, and cognition, thereby creating an integrated and coherent system. To validate the effectiveness of this approach, four extensive experiments were conducted, and the behavior of the constructed social agents was compared with real-world data across various domains. The experimental results demonstrate the model's capacity to accurately replicate human behavior, thereby validating its effectiveness. This work underscores the potential of LLM-based agents to generate realistic simulations of human behavior in complex, dynamic environments.

The evolutionary processes of complex systems encode critical information about their underlying functional mechanisms. The generation time of edges reflects the historical trajectory of network evolution, offering valuable insights into the formation and development of systems such as protein-protein interaction networks, ecosystems, and social networks. However, existing methods predominantly rely on access to the full structural evolution of networks for modeling and prediction, which is often difficult to obtain in real-world scenarios. This significantly limits their applicability and generalizability. To address this challenge, we propose a novel paradigm for network evolution inference—Inferring Dynamic Network Evolution from Node Steady States. Our approach leverages only the steady-state observations of nodes before and after structural perturbations, enabling accurate identification of newly added edges without requiring prior structural information. Compared to existing structure-based methods, the proposed model demonstrates greater adaptability and practicality, achieving superior predictive performance across multiple real and synthetic networks. This study introduces a new direction for modeling network evolution and holds promise for advancing network inference tasks under data-scarce conditions.

Driven by advances in Large Language Models (LLMs), integrating them into recommendation tasks has gained interest due to their strong semantic understanding and prompt flexibility. Prior work encoded user-item interactions or metadata into prompts for recommendations. In parallel, LLM reasoning, boosted by test-time scaling and reinforcement learning, has excelled in fields like mathematics and code, where reasoning traces and correctness signals are clear, enabling high performance and interpretability. However, directly applying these reasoning methods to recommendation is ineffective because user feedback is implicit and lacks reasoning supervision. To address this, we propose R2Rec, a reasoning-enhanced recommendation framework that samples interaction chains from the user-item graph and converts them into structured interaction-of-thoughts via a progressive masked prompting strategy, with each thought representing stepwise reasoning grounded in interaction context. This allows LLMs to simulate step-by-step decision-making based on implicit patterns. We design a two-stage training pipeline: supervised fine-tuning teaches basic reasoning from high-quality traces, and reinforcement learning refines reasoning via reward signals, alleviating sparse explicit supervision. Experiments on three real-world datasets show R2Rec outperforms classical and LLM-based baselines with an average 10.48% improvement in HitRatio@1 and 131.81% gain over the original LLM. Furthermore, the explicit reasoning chains enhance interpretability by revealing the decision process.

Reinforcement Learning (RL) has been a subject of extensive research for decades, leading to impactful applications in optimizing large-scale decision-making across sectors like transportation, energy, and public health. Following the advent of Large Language Models (LLMs), a natural first attempt at combining RL and LLMs is to use the semantic capabilities of LLMs to enhance traditional RL frameworks, for instance by generating reward functions or improving state representations. However, it soon became apparent that a more valuable and transformative research direction lies in tackling the inverse challenge: using RL to fundamentally train and align the large models themselves. This endeavor presents significant challenges, including navigating immense parameter sizes and action spaces and overcoming the persistent difficulties in reward evaluation. Our ongoing work is focused on this frontier, actively developing methods to address these challenges and enhance the reasoning abilities of LLMs, VLMs, and world models through RL, aiming to unlock more sophisticated planning and inference.

Coverage optimization is a fundamental problem in wireless communication systems, directly impacting network quality, capacity, and user experience. Traditional heuristic and model-based methods often struggle to adapt to the dynamic and complex nature of real-world wireless environments. In this work, we propose a reinforcement learning (RL)-based framework to optimize coverage in simulated wireless communication environments. By interacting with a high-fidelity communication simulator, the RL agent learns to make intelligent decisions to adjust transmission parameters and positioning strategies, aiming to maximize coverage performance under various environmental and network conditions. The proposed method is model-free, adaptable to heterogeneous scenarios, and does not require prior knowledge of channel dynamics. This work highlights the potential of RL in autonomous network management and lays a foundation for future Sim2Real transfer in intelligent wireless systems.

World models have shown great promise in modeling visual dynamics and generating furtue frames in complex environments. However, maintaining consistency over long-range video generation remains a significant challenge due to the lack of stable spatial memory and structural understanding. In this work, we propose to enhance the temporal and spatial consistency of world models by integrating features derived from 3D Gaussian Splatting (3DGS). As a neural-free, high-fidelity 3D representation, 3DGS captures fine-grained geometric and appearance information across views. We leverage its reconstructed features as spatial memory to guide the rendering process of world models.

With the growing interest in how LLM agents can simulate human society, research is increasingly focused on their ability to understand and navigate social structures and interactions. However, to truly simulate human behavior, LLM agents should, like humans, gradually develop and refine their abilities over time. In our study, we investigate the development of social intelligence in LLM agents, specifically exploring how their social capabilities can be enhanced within a simulated society through environmental and social influences. We aim to understand how these agents can develop world knowledge and social cognition over time, learning dynamic social relationships and navigating complex urban environments. We envision a scenario where a new agent enters an unfamiliar virtual society, interacts with others, and adapts to the environment by acquiring knowledge about the city's layout and social norms and dynamics. Overall, this presentation will provide a literature review on the topic and offer a brief introduction to the research design.

Forecasting mobile data enables operators to proactively perceive network status, allowing them to schedule resources and formulate optimization strategies in advance to improve service quality. However, in large-scale heterogeneous access networks, effective optimization and planning often require the prediction of multiple types of mobile data across different spatio-temporal scales to meet diverse objectives such as coverage, energy efficiency, and throughput. Traditional prediction methods with tailored models for exclusive data types undoubtedly increase the design complexity and deployment costs. In this paper, we propose a Mobile Foundation Model (MobiFM) for data forecasting, which adopts a unified diffusion model framework to forecast mobile data with diverse types (mobile traffic and users), various time granularities (hourly and minute-level), and multiple spatial scales (cell-level and grid level). MobiFM utilizes mix-of-expert networks to explore the features of different types of mobile data. Moreover, MobiFM employs a spatial knowledge graph to represent the complex spatial environments and uses a memory network to store the differentiated temporal features of the data. We train the MobiFM using 8 real-world datasets with over 80 million time-series points, totaling more than 1 billion tokens.

AgentSociety offers a versatile platform for building and experimenting with intelligent agents in interactive virtual environments. As agent behaviors grow in complexity and research applications diversify, users increasingly demand more intuitive and expressive ways to compose, manage, and adapt agent workflows. We introduce two major enhancements to address this need: a node-based programming interface and a Model Context Protocol (MCP) framework. The node-based system enables users to visually construct agent logic by connecting modular components via a drag-and-drop interface, lowering the barrier for non-programmers and supporting rapid prototyping. Meanwhile, the MCP provides a standardized mechanism for managing model states and contextual information during execution, facilitating more consistent agent behavior across complex scenarios. Together, these additions empower users to build, debug, and deploy sophisticated agents with greater flexibility and control.

The rapid proliferation of wireless applications requires accurate path loss estimation for efficient radio resource management. Radio maps provide a location-aware solution by representing received signal strength distribution over a geographical area, enabling intelligent and adaptive network management. However, constructing high-resolution radio maps is hindered by costly sensor deployments and time-consuming driving-test data collection. To overcome these challenges, we propose a foundation model DiffRM for multi-band radio map construction. DiffRM leverages diffusion models to capture the probabilistic characteristics exhibited by received signal strength. Additionally, DiffRM adopts a unified structure that incorporates masking strategies to jointly perform prediction and imputation, enabling the model to generalize across diverse radio map construction tasks. Furthermore, modality-specific encoders are applied in DiffRM to embed multimodal contextual information-including environmental layouts, numerical metadata, and semantic descriptions-into latent space, enhancing the model's adaptability across diverse scenarios, including varying terrains, sampling heights, and frequency bands. Extensive experiments on a comprehensive radio map dataset validate its superior performance and generalization, paving the way for enhanced wireless network design and optimization.

With the increasing frequency of natural disasters these years, the assessment of disasters based on remote sensing images has become increasingly important. The existing disaster assessment methods can mainly be divided into those based on semantic segmentation and those based on change detection. However, in complex disaster scenarios, both of these methods have encountered challenges. Therefore, we proposed UoDA, a Unified model of Disaster Assessment, to achieve disaster assessment under given single and multiple image conditions. In addition, we have also constructed a dataset UoDA-VQA, to evaluate the performance of both semantic segmentation and change detection models under different given conditions.

With the rapid expansion of urban areas and the increasing integration of distributed energy resources (DERs) and flexible loads into the urban power grid, traditional grid scheduling methods are becoming inadequate. The complexity of managing these resources necessitates a more intelligent and adaptive approach. This work proposes an optimization method for urban-level Virtual Power Plant (VPP) scheduling based on hierarchical multi-agent reinforcement learning (HMARL). The method employs a two-layered architecture: the upper layer consists of regional VPP agents that manage inter-VPP electricity transactions, while the lower layer comprises device-level agents that control individual DERs and loads. Currently, this method is being validated through simulations on a city-scale energy simulator based on Beijing's urban power grid. As an ongoing work, we welcome suggestions to further refine and enhance the proposed approach.

Large language model (LLM) agents have demonstrated strong capabilities across diverse domains. However, designing high-performing agentic systems remains challenging. Existing agent search methods suffer from three major limitations: (1) an emphasis on optimizing agentic workflows while under-utilizing proven human-designed components such as memory, planning, and tool use; (2) high evaluation costs, as each newly generated agent must be fully evaluated on benchmarks; and (3) inefficient search in large search space. In this work, we introduce a comprehensive framework to address these challenges. First, We propose a hierarchical search space that jointly models agentic workflow and composable functional components, enabling richer agentic system designs. Building on this structured design space, we introduce a predictive value model that estimates agent performance given agentic system and task description, allowing for efficient, low-cost evaluation during the search process. Finally, we present a hierarchical Monte Carlo Tree Search (MCTS) strategy informed by uncertainty to guide the search. Experiments on seven benchmarks, covering embodied, math, web, tool, and game, show that our method achieves an average performance gain of 8.34\% over state-of-the-art baselines and exhibits faster search progress with steeper improvement trajectories.

Large-scale traffic simulation has been a topic of interest in urban science. Existing simulation frameworks such as SUMO [1], CityFlow [2], and CBLab [3] are capable of modeling traffic flow and interactions between vehicles and pedestrians. Additionally, datasets like CityFlow dataset [2] and CBData [3] provide valuable input for traffic simulations. However, works till now lack an integrated workflow combining the generation of realistic, migratable, and reusable datasets and afterwards large-scale traffic simulation. To address this gap, we are working on a high-resolution urban mobility dataset generation workflow at the city scale based on our lab's previous simulation framework MOSS [4]. Our simulation is initially implemented in Beijing, with a traffic flow of over 2 million individuals, capturing both indoor and outdoor mobility patterns, along with corresponding activity behaviors. Outdoor mobility includes mode-specific trajectories for private vehicles, walking, metro, and buses, assigned through a calibrated transportation mode choice model. To ensure realism, we support customizable population profiles for generating indoor activities and apply a regression-based generative algorithm to model individual-level activity intentions and movement patterns.

Lying at the intersection of computational science and social science, the field of the science of science and human-computer interactions are often indicated as major research directions of information schools. Despite their high relevance, the two fields have largely evolved in parallel, with relatively limited integration. We argue that bridging the science of science and human-computer interaction research can cultivate impressive research for both communities. We illustrate these by introducing two lines of research that we have been investigating recently, i.e., science of science for HCI and HCI for science. Through delineating these possibilities, we outline a roadmap to foster more reflective, effective, and beneficial scientific research.

In light of recent changes in international perceptions derived from surveys of Americans' attitudes towards China, this work aims to leverage large language models to simulate the progression of attitudes-evolution on the macro scale. Motivated by works that highlight such models' tendency to reproduce human biases and the echo chamber effect, we first propose a social simulator framework where agents in a society are exposed to media and then propagate sentiments through agent-agent interaction. Through large-scale experiments, we aim to first reproduce the real-world macro-trend of US perceptions toward China over a 20 year period, and then explore the influencing factors of such trend. Through our social simulator we are able to disentangle the influences of media vs. social interaction on attitude propagation, and further determine which forms of media are most likely to engender negative or positive attitudes. Finally, we propose to conduct interventions to reduce the macro-echo chamber effect by introducing contradictory information via heterogenous agents, with the goal of reproducing the recent thaw in negative attitudes of Americans towards China as a result of the mass RedNote migration. Overall, our work aims to shed light on the mechanisms of macro-views evolution on the global scale while contributing suggestions as to how to diffuse bias-driven negative perceptions. 

Information dynamics in complex systems govern how signals, influences, and perturbations propagate, interact, and self-organize across scales. In this presentation, we synthesize recent advances at the intersection of complex systems theory and statistical physics to characterize the mechanisms driving emergent information flows. We begin by formalizing information-theoretic measures—entropy rates, mutual information, and transfer entropy—and their extensions to time‐varying and multilayer networks. Next, we investigate statistical‐mechanical models of information spreading, from percolation and epidemic‐like processes on static graphs to non‐equilibrium spin systems and random‐field models on adaptive networks. We then discuss computational frameworks that combine dynamical inference, maximum‐entropy approaches, and path‐integral methods for reconstructing causal interactions in high‐dimensional data. Special attention is devoted to the role of criticality and phase transitions in optimizing information storage and transmission, highlighting universal scaling laws and finite‐size effects. Finally, we present previous applications ranging from neural and genetic regulatory networks to socio‐technical systems, illustrating how information‐driven feedback can induce multistability, synchronization, and pattern formation. By unifying theoretical insights and empirical methodologies, this presentation aims to chart a cohesive roadmap for future research on information dynamics in complex adaptive systems.

Effective model training and updating strategies are essential for enabling intelligent end-cloud collaborative systems, particularly in user-centric applications like intent prediction. Traditional cloud-based approaches face significant limitations in personalization due to privacy concerns and the distributional gap between device and cloud data. To overcome these challenges, we propose FDCCL, a novel Federated Device-Cloud Collaborative Learning framework designed for efficient and privacy-preserving model training and updating. FDCCL leverages large foundation models in the cloud for global guidance while enabling on-device personalization through LoRA-based lightweight fine-tuning. The framework incorporates a cloud-guided training process using non-sensitive features, federated parameter aggregation to preserve user privacy, and a two-stage data selection strategy to reduce communication overhead. Extensive experiments on large-scale real-world datasets show that FDCCL achieves superior model accuracy and adaptability compared to state-of-the-art baselines, while maintaining strong privacy guarantees and efficient communication. An ablation study validates the contribution of each component to the effectiveness of the end-cloud collaborative learning strategy.

The 15-minute city is a pivotal concept in contemporary urban planning, designed to enhance residents’ quality of life. Conventional planning approaches often depend on human expertise or simplistic topological metrics, which inadequately account for the complexity and variability of residents’ mobility behaviors. To address this limitation, we propose an adaptive mobility simulation algorithm that models macro-level group mobility patterns, generating realistic travel trajectories for large populations across diverse urban layouts. Additionally, we introduce a reinforcement learning framework with reward modeling to bridge the efficiency gap between computationally intensive mobility simulations and rapid, intelligent decision-making. Experiments on synthetic urban environments demonstrate that our approach yields more efficient 15-minute city layouts. Furthermore, tests in real-world urban renewal scenarios confirm that our generated plans align closely with those of human experts in their core principles.

AgentSociety provides an authentic interactive environment for developing agents in virtual societies, enabling agents to perform movement behaviors, social interactions, economic activities, and more. As the community scale and potential application scenarios expand, the default social agents in AgentSociety can no longer meet diverse demands. Therefore, designing a more flexible agent construction mechanism to foster the growth of the community ecosystem has become a crucial task. To support community contributions and knowledge sharing, as well as enable GUI and prompt-based agent customization, we propose a three-tier Agent-Block-Function architecture in AgentSociety V1.5, allowing free combination of various behavioral modules. Based on this framework, researchers can: 1) Develop their own Blocks to implement specific agent behaviors (via rules or large model-driven Function calls). 2) Dynamically add or remove Blocks in an Agent to modify its available behavioral capabilities. The GUI application further facilitates these operations, offering flexible customization options—from profile settings to prompt engineering and parameter adjustments—to support tailored agent design. Building upon this framework, we established agentsociety_community as a hub for community contributions. Initially, it will host specialized agents for the BDSC2025 competition as a preliminary test case.

Langevin dynamics sampling suffers from extremely low generation speed, fundamentally limited by numerous fine-grained iterations to converge to target distribution. We introduce PID-controlled Langevin Dynamics (PIDLD), a novel sampling acceleration algorithm that reinterprets the sampling process through control theory principles. By treating energy gradients as feedback signals, PIDLD combines historical gradients (integral term) and gradient trends (derivative term) to efficiently traverse energy landscapes and adaptatively achieve stablization, significantly reducing required iterations to produce quality samples. Our approach requires no additional training, datasets, or prior information, making it immediately integrable with any Langevin-based method. Extensive experiments across image generation and reasoning tasks demonstrate that PIDLD achieves higher quality with fewer steps, making Langevin-based generative models more practical for efficiency-critical applications.

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 two new tracks: proactive task suggestions by directly interpreting user interfaces, and personalized task execution by catering to users' action preferences. We collected unique human demonstrations of multi-step Android device interactions across a variety of everyday apps. These demonstrations are not isolated but are continuously acquired from the users' long-term usage in their real lives, and encompass essential user-related contextual information. Our experiments reveal challenges of the tasks we propose. The model fine-tuned with the data we collected effectively utilized user information and achieved good results, highlighting the potential of our approach in building more user-oriented mobile GUI agents.

Multi-agent systems (MAS) based on large language models (LLMs), where multiple agents collaborate to solve complex tasks, have demonstrated remarkable performance in areas such as software engineering, operations research, and mathematical problem solving. However, the resilience of such systems—defined as their ability to maintain functionality under perturbations like agent failures—has received limited systematic investigation, especially regarding the influence of network topology and system scale. In this work, inspired by classical studies of network robustness, we propose a comprehensive framework for evaluating the resilience of LLM-based MAS. We define resilience as the area under the performance curve measured across varying agent error rates, and apply this metric to MAS configurations with diverse sizes and topologies. Our experiments reveal a strong scaling law between system size and resilience—a correlation even stronger than that between size and performance. Additionally, we find that network topology significantly affects resilience. Based on these findings, we propose a topology generation model that aims to enhance MAS resilience under cost constraints. Our work provides new insights into the design of robust MAS and lays the groundwork for building more reliable multi-agent AI frameworks. 

User mobility and mobile traffic are inherently coupled, reflecting both the physical movement and cyber behavior of users. Modeling their joint distribution is essential for a wide range of downstream applications. However, existing studies often treat them as separate tasks, with limited exploration of their mutual influence. Capturing their relationship remains challenging due to differences in data modalities. Moreover, their interactions are shaped by urban functional areas that influence both user mobility intentions and app usage patterns, highlighting the importance of incorporating urban knowledge into the generation process. In this work, we propose a diffusion-based framework that consists of a multi-scale Transformer to jointly capture the dependencies between user mobility and mobile traffic. Firstly, we apply a wavelet transform to the mobile traffic signals to extract temporal features at multiple resolutions. Secondly, we incorporate an urban knowledge graph into the trajectory representations to encode spatial semantics. Finally, a cross-attention mechanism is then applied across different temporal scales, enabling the model to capture traffic and mobility features in a unified and context-aware manner.

Base stations account for the majority of energy consumption in mobile networks, even under low traffic loads due to the need to maintain basic coverage. However, due to overlapping coverage designed for seamless service and capacity enhancement, a significant portion of this energy becomes redundant, leading to high operational costs and environmental concerns. Existing methods attempt to mitigate this by identifying low-load cells with overlapping coverage and offloading their traffic to neighboring cells, allowing them to enter energy-saving sleep modes. Yet, determining optimal sleep configurations and traffic migration plans is challenging, as modern mobile network is complex and inappropriate decisions can result in coverage loss and network congestion. Traditional approaches often rely on per-timestep optimization under static assumptions, which limits their applicability to real-world networks with complex temporal and spatial dynamics. In this work, we propose a novel framework that integrates a world model with deep reinforcement learning to optimize base station energy consumption. The world model captures user mobility and traffic patterns over time, enabling accurate simulation of network responses to energy-saving actions. By leveraging the predictive capabilities of the world model, our method learns realistic and deployable energy-saving strategies with long-term planning abilities, improving energy efficiency without compromising quality of service.

The rapid pace of urbanization has led to unequal benefits for residents, creating significant inequality and segregation issues and discussions around Sustainable Development Goals 10 and 11. Accurate measurement of inequality within urban areas is essential for effective mitigation strategies. Traditional methods rely on survey-based census data, which are time-consuming and delayed, while some studies use coarse proxies like nighttime lights. However, these methods are limited by resolution and fail to capture fine-grained disparities within communities. To address this, we aim to leverage accessible urban imagery, which offers detailed visual features. Two key challenges must be addressed: 1) accurately perceiving micro-level inequalities within neighborhoods, and 2) ensuring that this perception is interpretable for policy guidance. To address these gaps, we propose UI-CoT, a framework that leverages the power of urban imagery-based visual language models in urban inequality and segregation perceiving, enhanced by Chain-of-Thought prompting to improve reasoning capabilities. We fine-tune a visual language model to predict three essential neighborhood inequality indicators: the income Gini coefficient, dominant race, and racial income ratio. Extensive experiments show that our model can effectively perceive micro-level inequalities, with the incorporation of Chain-of-Thought reasoning further improving the model's performance by 17.2%. This research offers valuable insights into addressing inequalities within urban environments and demonstrates the potential of web resources in empowering urban sustainable development.

As artificial intelligence systems, particularly large language models (LLMs), increasingly interact in multi-agent and decentralized settings, questions arise about how coordination, cooperation, and behavioral regulation might naturally emerge. This work explores the emergence of social norms in LLM societies, drawing upon foundational theories from sociology, game theory, and computational modeling. Social norms—informal, collectively enforced behavioral expectations—have historically governed human interactions, enabling stability and coordination without formal laws. The first part of this study revisits key definitions and classical examples of social norms, grounded in the work of Cristina Bicchieri, Robert Axelrod, and Elinor Ostrom. Building on this, the second part examines mechanisms behind the emergence of norms, identifying three critical drivers in artificial societies: (1) the scale and density of agent interactions, (2) social learning and imitation, and (3) the presence of meta-norms—second-order norms that regulate the enforcement of primary norms. By contextualizing these mechanisms within LLM-based agent environments, this research proposes a conceptual framework for simulating and analyzing norm emergence among LLMs. This includes theorizing how LLMs may enforce compliance, adapt behaviors, and even self-regulate through emergent higher-order rules. These insights lay the groundwork for future studies in AI alignment, multi-agent governance, and AI-driven institutions.

Travel surveys are essential for urban mobility studies, providing critical data on individual movement patterns within cities. These surveys play a vital role in city planning, traffic management, and environmental assessments. However, traditional travel surveys face significant challenges, including low response rates, high costs, and privacy concerns, which hinder their effectiveness and scalability. Additionally, the data collected often lacks the depth and contextual understanding necessary for more nuanced urban mobility analysis. To address these limitations, this study proposes a novel approach to enhance travel survey simulations using a large reasoning model. We aim to enable the model to simulate the thought process of urban residents, allowing it to reason about key aspects of travel—such as whether, why, when, where, and how—to enhance its "human-like" simulation capabilities. Specifically, we constructed a "chain of trips" dataset for travel surveys, using this dataset, we performed few-shot learning experiment and fine-tuned the model. Our results show that this method significantly improves the model's ability to simulate more realistic travel patterns. In the future, we want to further stimulate the ability of the reasoning model through fine-tuning and make more attempts using RLHF. Any questions or suggestions are welcome.

The recent development of large language models (LLMs) has not only advanced machine capabilities in traditional natural language processing tasks, but also unlocked new possibilities for designing agents with human-level intelligence in communication, reasoning, and decision-making. In our previous work, we proposed AgentSociety 1.0, a large-scale social simulation platform that integrates LLM-driven generative social agents, realistic societal environments, and a powerful simulation engine to study complex social dynamics. By simulating over 10k individual agents and five million interactions, we explored topics such as political polarization, the spread of inflammatory content, the effects of universal basic income, and responses to external shocks. Furthermore, real-world societies are not only composed of individuals, but also of diverse institutional collectives such as firms, governments, and banks, which together form intricate social and economic structures. Motivated by this broader perspective, in this talk I will review some classical works on simulating economic and financial systems, discuss emerging attempts to integrate LLMs into such domains, and share some of our recent thoughts and ongoing explorations in extending the AgentSociety framework towards richer modelling.

High-quality human mobility data is crucial for applications such as urban planning, transportation management, and public health, yet its collection is often hindered by privacy concerns and data scarcity, particularly in less-developed regions. To address this challenge, we introduce WorldMove, a large-scale synthetic mobility dataset covering over 1,600 cities across 179 countries and 6 continents. Our method leverages publicly available multi-source data, including gridded population distribution, point-of-interest (POI) maps, and commuting origin-destination (OD) flows, to generate realistic city-scale mobility trajectories using a diffusion-based generative model. The generation process involves defining city boundaries, collecting multi-source input features, and simulating individual-level movements that reflect plausible daily mobility behavior. Comprehensive validation demonstrates that the generated data closely aligns with real-world observations, both in terms of fine-grained individual mobility behavior and city-scale population flows. Alongside the pre-generated datasets, we release the trained model and a complete open-source pipeline, enabling researchers and practitioners to generate custom synthetic mobility data for any city worldwide. WorldMove not only fills critical data gaps, but also lays a global foundation for scalable, privacy-preserving, and inclusive mobility research, empowering data-scarce regions and enabling universal access to human mobility insights.

World models have emerged as powerful world simulators, offering the potential to enhance limited real-world data and enable closed-loop agent learning. However, existing research primarily focuses on gaming and autonomous driving, with limited exploration of world models tailored to robotic learning due to the scarcity of real-world data and the complexity of robotic motions. In this paper, we present PhysicsRobo, a controllable video generation model designed to predict future visual observations based on past experiences and robotic actions. To effectively capture the intricate dynamics of robotic motion, PhysicsRobo utilizes an action-controllable autoregressive transformer model, trained on a carefully curated high-quality egocentric videos of robotic agents. The model is highly scalable, capable of expanding to over 1 billion parameters. By leveraging the depth maps and trajectory points of continuous robotic action videos, we have significantly enhanced the world model’s implicit perception of the laws of physics. Experiments have shown that this approach significantly enhances visual performance.  

The Chain of Thought (CoT) methodology has demonstrated significant potential in enhancing the reasoning capabilities of large language models, establishing multi-step thinking as a paradigm for their reasoning processes. But the expensive API costs dramatically limit its application, which could be particularly difficult to tackle for realisitic complicated tasks. Meanwhile, collaborative reasoning among multiple models has also shown promise at the respect of reducing costs. Some previous work has already proved that the combination of multi-model collaborative reasoning and multi-step reasoning could reduce API costs while maintaining relatively powerful reasoning abilities. But there is still some space for further enhancing the granularity and accuracy of collaborative allocation. Our ongoing work further investigates the integration of multi-model collaborative reasoning into multi-step reasoning as a means to maintain reasoning capabilities and reduce costs. We propose a framework incorporating two modules for task decomposition and model allocation, along with an integrated reinforcement learning framework, to explore the feasibility of this approach. It aims to validate the effectiveness of this methodology in boosting the reasoning performance of multi large language models.

Public transportation systems play a vital role in promoting environmental sustainability and improving urban efficiency. These systems aim to meet the daily mobility needs of people across diverse urban settings. However, top-down design approaches can often lead to unequal access, leaving certain populations underserved or excluded from reliable transit options. In this study, we explore how mobility patterns can be used to assess inequality within urban transit systems. Leveraging large-scale origin-destination (OD) flow data, we aim to evaluate public transportation performance across a wide range of cities. To support this analysis, we have collected transit data from over 1,300 cities worldwide and are now preparing for a comprehensive assessment of mobility-based transit equity.

The frequency and intensity of natural disasters has significantly increased over the past decades. Facing these sudden or potential disasters, accurately modeling human mobility become the critical issue for evacuation route planning, rescue resource allocation, and long-term social reconstruction. During disasters such as heavy rainstorm and wildfires, human mobility patterns often deviates from their daily routine to a large extent, which makes the task more challenging. However, existing works primarily focus on trajectory generation under normal scenarios or on models tailored to specific disaster scenarios. In this work, we propose a unified model for human mobility generation in natural disasters, capable of cross-disaster and cross-city generalization. Specifically, inspired by a spatiotemporal decay model that uncovers the underlying uniformity in mobility perturbations caused by disasters, we design physics-informed dynamic condition mechanisms and loss functions to model the impact of disasters on mobility, thereby enabling cross-disaster transfer. Furthermore, our model architecture incorporates shared modules to capture city-universal mobility patterns and private modules to model city-specific characteristics, enhancing generalization across different cities.

Classical urban sensing approaches often rely on street view or remote sensing imagery to predict socioeconomic indicators, typically using contrastive learning or relatively simple computer vision models. However, with the advent of large language models, urban sensing is entering a new era. Trained on Internet-scale multimodal data—including massive amounts of images and their corresponding textual descriptions—LLMs have acquired a rich knowledge base and remarkable capability in understanding urban environments. Recently, a growing number of studies have explored leveraging LLMs to predict socioeconomic indicators, with LLMs playing roles broadly categorized as either enhancers or predictors. Nevertheless, existing studies have yet to fully explore the potential of LLMs in urban spatial understanding. In this work, we propose to leverage LLMs to predict urban socioeconomic indicators from street view imagery, aiming to overcome previous models’ limitations in geographic generalization, spatial reasoning capabilities, and the cultural and subjective perception of places. 

The statistics of wealth status is crucial for the evaluation and development of every region. However, in African countries, this kind of data is usually missing due to the high cost of data collection and the indifference of local governments. Considering the long-term impact and benefits of wealth status understanding, we present an economic yet convenient way to update wealth status in Africa: Using multi-modal Large Language Model method to predict the wealth index of African countries via frequently updating satellite and street view imagery. Based on UrbanMLLM, which is pre-trained on satellite and street view imagery from the United States, we use Chain-of-Thought fine-tuning approach to transfer the model's ability on urban imagery analysis to African regions. Learning through the urban imagery and wealth index from thousands of administratives in four African countries, the model understands the relationship between images and wealth. Our wealth prediction results are further validated on these countries, including Nigeria, Uganda, Kenya, and South Africa, achieving high accuracy and R-squared values, which shows great potential of Multimodal LLM's transfer learning ability on economic prediction tasks.

AI empowerment has become a key focus in the development of sixth-generation (6G) mobile communication technologies. With the increasingly outstanding capabilities of generative AI in areas such as network traffic modeling, channel estimation and prediction, and user behavior forecasting and simulation, building digital twin systems for mobile wireless networks centered on generative AI is emerging as a powerful tool for network optimization and planning. The development of universal generative models for wireless networks and semantic information-oriented network simulation represents major directions in wireless digital twins. A pressing topic among researchers is how to effectively apply large language models (LLMs), which integrate both generative and reasoning capabilities, to mobile networks. This survey explores foundational research on generative AI for enabling wireless digital twins in both sequential and image modalities. It further introduces ongoing efforts and future directions involving large-scale models, particularly LLMs, with the aim of building integrated, general-purpose models that unify generation, simulation, and decision-making.

Urban transportation resilience characterizes a city's road network’s ability to withstand attacks while maintaining efficiency, serving as a crucial property of urban mobility systems. Traditional studies assess resilience through an interventional approach, where specific road segments are deliberately disrupted, followed by traffic assignment on the perturbed network to measure efficiency changes. However, accurately quantifying resilience requires repeated interventions and reassignments under varying attack proportions, making the process computationally inefficient and limiting its ability to reveal knowledge about key influencing factors. In this work, we shift from an interventional to an observational indicator, identifying critical determinants of road network resilience. We train a graph neural network (GNN) that directly predicts resilience from observed traffic assignment results given an origin-destination (OD) demand, bypassing the need for disruptions and reassignment. Furthermore, we conduct interpretability analyses on the trained model, uncovering key roads and features influencing resilience and deriving a symbolic formula for resilience computation. Our findings provide insights into fundamental questions such as what makes a transportation network resilient and how resilience can be improved, offering practical guidance for urban transportation planning and infrastructure design.

Neonatal mortality rate (NMR), defined as the number of infant deaths within the first 28 days of life, remains a critical public health challenge, particularly in regions with limited access to essential healthcare services. To achieve the Sustainable Development Goal (SDG) of reducing NMR to below 12 per 1,000 live births by 2030, it is essential to assess the impact of various factors and identify effective intervention strategies.In this study, we focus on electricity access as a crucial factor influencing neonatal survival, given its impact on healthcare quality and living conditions. To accurately and comprehensively assess its effect across Africa, we rely on publicly available data to address the limitations of small-scale and incomplete survey records. Specifically, we employ a remote-sensing model to estimate NMR across African countries. Additionally, we evaluate electricity access using publicly available nighttime light data. Preliminary findings reveal a correlation between electricity access and lower NMR, underscoring the need to prioritize electrification initiatives to improve neonatal health outcomes. 

K-Anonymity is a classic work in privacy protection. It reveals a fact to us: Even if you post information anonymously, attackers can still correlate and link the data back to its owner through user linkage attacks. With the rapid development of LLM, its profound language comprehension ability raise concern in privacy protection. In this work, we want to explore that whether LLM can link data back to its owner given some history data released by each user. Furthermore, we want to build an agentic workflow with a smaller LLM or fine-tune an LLM to help protect the data from being linked to its owner without compromising linguistic usability.

World models have emerged as powerful world simulators, offering the potential to enhance limited real-world data and enable closed-loop agent learning. However, existing research primarily focuses on gaming and autonomous driving, with limited exploration of world models tailored to robotic learning due to the scarcity of real-world data and the complexity of robotic motions. In this paper, we present RoboScape, a controllable video generation model designed to predict future visual observations based on past experiences and robotic actions. To effectively capture the intricate dynamics of robotic motion, RoboScape utilizes an action-controllable autoregressive transformer model, trained on a carefully curated high-quality egocentric videos of robotic agents. The model is highly scalable, capable of expanding to over 1 billion parameters. RoboScape functions as a versatile world simulator for various downstream tasks, enabling policy evaluation from low-level action inputs, generating synthetic data for robotic policy training, and supporting online reinforcement learning for robotic agents.

Transfer learning has long been a core topic in the field of machine learning. In the context of physical systems, this issue translates into predicting the dynamic behavior of the same system under different environmental conditions. This is a common yet challenging task in practice, requiring models to make zero-shot predictions on unseen out-domain data based on scarce in-domain data. Current approaches mainly fall into three categories: large foundation models, in-context learning, and meta-learning. However, these methods either rely on large amounts of data, use few-shot samples for warm-starting, or constrain the model's weight space. Considering their inherent limitations, I have turned to a novel paradigm to address the generalization prediction problem for physical systems by learning the data-weight landscape (joint distribution). In this talk, I will formalize the generalization prediction task for physical systems, review related work, and present my preliminary approach along with experimental results.

Trajectory generation has extensive applications in areas such as traffic congestion, energy modeling, and public health. Numerous works have been conducted using physical models, machine learning models, and agent-based approaches. However, physical models fail to capture the complex patterns of human mobility, while machine learning models rely heavily on large amounts of training data and perform poorly in scenarios where trajectory data is scarce. Agent-based methods, on the other hand, offer good interpretability and transferability, making it easier to model aspects such as movement distance, types of trajectory points, and historical transition patterns in trajectory data. They also perform well in zero-shot scenarios. Nevertheless, traditional agent-based methods have two main limitations: ① they overly rely on historical data, leading to poor performance for new users; and ② they cannot be directly applied to real urban spaces, instead requiring indirect methods such as introducing external data or retrieving information from historical trajectories to map to urban spaces. This study employs semantic modeling to capture high-level mobility patterns, enabling generalization from existing users to new users. Additionally, by modeling individual mobility patterns and leveraging CityGPT, a foundational model with urban spatial knowledge, we aim to align trajectory generation knowledge with the capabilities of large models, thereby achieving direct user trajectory generation in real urban spaces.

This sharing will provide an overview and summary of the latest research in the field of internal function zone of Large Language Models (LLMs).Studying the internal function zone of the LLM provides us with new insights into the internal representation and processing mechanisms of the LLM.We refer to a taxonomy that classifies methods by what part of the network they are based on.We will explore latest advances, focusing on methods,key research and breakthroughs for the three research objects: independent neuron, subnetwork and latent representation of LLM. In addition to this, we will analyze possible future directions in this field, focusing on the further understanding and application of functional zone within the LLM.

As technology advances, personalized intelligent smartphone assistants are becoming more essential, greatly improving user convenience, efficiency, and interaction. Understanding user behavior is crucial for these assistants to meet personalized needs. Although behavior sequence models can capture basic behavioral knowledge, they are limited in functionality and cannot satisfy diverse user requirements. By leveraging the language comprehension and reasoning abilities of large language models (LLMs), we have completed the modality alignment between the language model and the behavior sequence model. This alignment allows us to fine-tune the LLM to understand behavior sequences and develop a multi-functional intelligent assistant that can interpret user behavior. In practice, we have adopted a three-stage alignment approach inspired by visual-language alignment techniques. In the first stage, we fine-tune only the projection layer parameters to enable the LLM to initially understand the behavior sequence embeddings. In the second stage, we integrate subtasks such as behavior prediction, generation, and profile creation through multi-turn dialogues and use a balanced loss strategy to enhance the model's multi-task performance. In the third stage, we introduce self-reflection and improvement training for user profiling to enhance profile quality and further boost the model's performance in prediction tasks. Through this multi-stage alignment training, we have developed a user behavior intelligent assistant that can understand behavior sequences and perform multiple tasks effectively.

Chaotic systems, defined by their exquisite sensitivity to initial conditions and intrinsic long-term unpredictability, present formidable challenges that reverberate across scientific and engineering domains. This presentation illuminates the transformative role of modern state-space models, specifically Mamba-based architectures, in forecasting such systems, offering an in-depth exploration of their methodological frameworks and empirical efficacy. By conceptualizing system dynamics as the evolution of latent state variables, these models masterfully encapsulate the nonlinear complexities inherent in chaotic behavior, seamlessly integrating observational data to deliver robust and precise predictions. Our approach distinguishes itself by outperforming a spectrum of established physics-informed and data-driven paradigms, including reservoir computing, NeuralODEs, LSTMs, Transformers, etc, achieving accurate predictions extending as far as four Lyapunov times in the Lorenz system. This leap in performance heralds a significant advancement in chaotic system forecasting, redefining the boundaries of predictive modeling in complex systems. Since it is still an ongoing work, we will appreciate it if you can give some valuable suggestions.

Large Language Models (LLMs) have shown strong potential in planning and decision-making tasks, building on their success in natural language processing. They can act as standalone planners or enhance traditional planning systems, offering benefits like commonsense reasoning and transforming unstructured text into structured data. However, challenges remain, particularly in long-horizon and multi-step planning, where systematic exploration and deliberate reasoning are critical. One practical and promising application of LLMs for planning is in the travel domain. By generating detailed itineraries, suggesting accommodations and transportation, and adapting to changing conditions, LLMs demonstrate their ability to handle complex, real-world scenarios. This use case also highlights the importance of integrating LLMs with external tools and structured algorithms to improve planning efficiency, accuracy, and user satisfaction.

Accurate prediction of urban spatiotemporal dynamics is essential for enhancing urban management and decision-making. Existing spatiotemporal prediction models are predominantly deterministic, focusing on primary spatiotemporal patterns. However, those dynamics are highly complex, exhibiting multi-modal distributions that are challenging for deterministic models to capture. In this paper, we highlight the critical role of probabilistic prediction in capturing the uncertainties and complexities inherent in spatiotemporal data. While mainstream probabilistic models can capture uncertainty, they struggle with accurately learning primary patterns and often suffer from computational inefficiency. To address these challenges, we propose CoST, which collaborates deterministic and probabilistic models to improve both predictive accuracy and the ability to handle uncertainty. To achieve this, we design a mean-residual decomposition framework, where the mean value is modeled by a deterministic model, and the residual variations are learned by a probabilistic model, specifically diffusion models. Moreover, we introduce a scale-aware diffusion process, which better accounts for spatially heterogeneous dynamics across different regions. Extensive experiments on eight real-world datasets demonstrate that CoST significantly outperforms existing methods in both deterministic and probabilistic metrics, achieving a 20\% improvement with low computational cost. CoST bridges the gap between deterministic precision and probabilistic uncertainty, making a significant advancement in the field of urban spatiotemporal prediction. 

The ability to attribute others' mental states, known as Theory-of-Mind (ToM), is a cornerstone of social intelligence. While large language models (LLMs) have exhibited impressive performance at various tasks including role-playing, their ToM reasoning capabilities remain limited and unreliable compared to humans. Meanwhile, the potential of leveraging role-playing for enhancing social cognition remains largely unexplored. To bridge this gap, we pioneer the investigation into how role-playing influences LLMs' Theory-of-Mind capabilities. We introduce RoleToM, a novel approach that integrates step-by-step reasoning with role-playing, demonstrating superior ToM abilities compared to perspective-taking and chain-of-thought methods alone. Additional experiments including ablation studies of role-playing and fine-tuning Llama 3.1-8B-Instruct on RoleToM-generated data, show that structured first-person simulation can effectively improve LLMs' social intelligence capabilities and generalize across different scenarios. We hope the role-playing methodology opens potential avenues for further applications and research in LLMs' social cognition.

In recent years, large language models (LLMs) have witnessed remarkable advancements, with the test-time scaling law consistently pushing the limits of reasoning capabilities. Through systematic evaluation and exploration of a diverse spectrum of intermediate ``thoughts'', LLMs demonstrate the potential to generate deliberate reasoning steps, thereby substantially enhancing reasoning accuracy. However, LLMs' autoregressive generation paradigm results in reasoning performance scaling sub-optimally with test-time computation, often requiring excessive computational overhead to propose thoughts while yielding only marginal performance gains. In contrast, diffusion language models (DLMs) can efficiently produce diverse samples through parallel denoising in a single forward pass, inspiring us to leverage them for proposing intermediate thoughts, thereby alleviating the computational burden associated with autoregressive generation while maintaining diversity and quality. In this work, we propose an efficient collaborative reasoning framework, leveraging DLMs to generate candidate thoughts and LLMs to evaluate their quality. Extensive experiments across diverse benchmarks demonstrate that our framework not only achieves strong performance in complex reasoning tasks with remarkable generalization capabilities, but also explores a novel and more efficient test-time scaling law, offering a promising direction for future research.

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. In this research, we aim to investigate the ability of LLMs to simulate real-like segregation behavior and explore the patterns of urban segregation. We propose a simple segregation model where agents living in the urban environment based on certain characteristic. We use a LLM to simulate the behavior of agents in this model and evaluate its performance in reproducing segregation patterns. Our preliminary results show that the LLM is able to simulate segregation behavior across cities. By varying the agent profiles and the distribution of points of interest (POIs), we observe different segregation patterns, indicating the influence of these factors on segregation behavior. Furthermore, we find that the LLM can easily perform counterfactual simulations, allowing us to explore how segregation behavior would change under different scenarios, such as during a pandemic.  

Large language models (LLMs) possess extensive world knowledge, including geospatial knowledge, which has been successfully applied to various geospatial tasks such as mobility prediction and social indicator prediction. However, LLMs often generate inaccurate geospatial knowledge, leading to geospatial hallucinations—incorrect or inconsistent representations of geospatial information—that compromise their reliability. While the phenomenon of general knowledge hallucination in LLMs has been widely studied, the systematic evaluation and mitigation of geospatial hallucinations remain largely unexplored. To address this gap, we propose a comprehensive evaluation framework for geospatial hallucinations, leveraging structured geospatial knowledge graphs for controlled assessment. Through extensive evaluation across 20 advanced LLMs, we uncover the hallucinations in their geospatial knowledge. Building on these insights, we introduce a dynamic factuality aligning method based on Kahneman-Tversky Optimization (KTO) to mitigate geospatial hallucinations in LLMs, leading to a performance improvement of over 29.6% on the proposed benchmark. Extensive experimental results demonstrate the effectiveness of our benchmark and learning algorithm in enhancing the trustworthiness of LLMs in geospatial knowledge and reasoning tasks.

Large language models (LLMs) have exhibited remarkable capabilities and achieved significant breakthroughs across various domains, leading to their widespread adoption in recent years. Building on this progress, we investigate their potential in the realm of local life services. In this study, we establish a comprehensive benchmark and systematically evaluate the performance of diverse LLMs across a wide range of tasks relevant to local life services. To further enhance their effectiveness, we explore two key approaches: model fine-tuning and agent-based workflows. Our findings reveal that even a relatively compact 7B model can attain performance levels comparable to a much larger 72B model, effectively balancing inference cost and model capability. This optimization greatly enhances the feasibility and efficiency of deploying LLMs in real-world online services, making them more practical and accessible for local life applications. We will release our benchmark, code, and fine-tuned LLMs upon acceptance.

Understanding experienced segregation in urban daily life is crucial for addressing societal inequalities and fostering inclusivity. The abundance of user-generated reviews on social media encapsulates nuanced perceptions and feelings associated with different places, offering rich insights into segregation. However, leveraging this data poses significant challenges due to its vast volume, ambiguity, and confluence of diverse perspectives. To tackle these challenges, we introduce using Large Language Models (LLMs) to automate online review mining for segregation prediction. In this work, we design a Reflective LLM Coder to digest social media content into insights consistent with real-world feedback, and eventually produce a codebook capturing key dimensions that signal segregation experience. Moreover, we design a REasoning-and-EMbedding (RE'EM) framework, which combines the reasoning and embedding capabilities of language models to integrate multi-channel features for segregation prediction. Experiments on real-world data demonstrate that our framework greatly improves prediction accuracy. The derived codebook is generalizable across three different cities, consistently improving prediction accuracy. Moreover, our user study confirms that the codebook-guided summaries provide cognitive gains for human participants in perceiving POIs' social inclusiveness.

Human mobility prediction plays a crucial role in urban planning, transportation optimization, and personalized services. However, the inherent randomness, non-uniform time intervals, and complex patterns of human mobility, compounded by the heterogeneity introduced by varying city structures, infrastructure, and population densities, present significant challenges in modeling.Existing solutions often require training separate models for each city due to the unique spatial representations and geographic coverage of urban areas, limiting their generalizability. We propose UniMove, a unified model to predict human mobility across multiple cities. Specifically, we propose a trajectory-location dual-tower architecture, where the location tower generates universal spatial representations and the trajectory tower captures mobility patterns from sequential trajectories. We design MoE Transformer blocks to adaptively select expert networks based on data characteristics, effectively handling diverse movement patterns. Extensive experiments across multiple datasets from diverse cities demonstrate that UniMove truly embodies the essence of a unified model. By enabling joint training on multi-city data and facilitating mutual data enhancement, it significantly improves mobility prediction across different urban environments by over 10.2%. UniMove represents a key advancement toward realizing a true foundational model with a unified architecture for human mobility.

Human mobility prediction is essential for applications like urban planning and transportation management, yet it remains challenging due to the complex, often implicit, intentions behind human behavior. Existing models predominantly focus on spatiotemporal patterns, paying less attention to the underlying intentions that govern movements. Recent advancements in large language models (LLMs) offer a promising alternative research angle for integrating commonsense reasoning into mobility prediction. However, it is a non-trivial problem because LLMs are not natively built for mobility intention inference, and they also face scalability issues and integration difficulties with spatiotemporal models. To address these challenges, we propose a novel LMP (LLMs for Mobility Prediction) framework. Specifically, LMP introduces an Schema Learning-based Agentic Workflow to unleash LLM’s commonsense reasoning power for mobility intention inference. Besides, we design a data-efficient intent fine-tuning scheme to effectively distilling reasoning power from proprietary LLM to smaller-scale, open-source language model, ensuring LMP’s scalability to millions of mobility records. Moreover, we propose a transformer-based intent-aware mobility prediction model to effectively harness the intent inference ability of LLM. Evaluated on three real-world datasets, LMP outperforms state-of-the-art baselines on 7 out of 12 metrics and ranks as the second-best method on the remaining ones, demonstrating improved accuracy in next-location prediction and effective intention inference. The interpretability of intention-aware mobility prediction highlights our LMP framework’s potential for real-world applications.