Jie Feng 冯杰 中文
← Research
RESEARCH SERIES

Multi-Agent
& Social Intelligence.

Collaborative problem solving and behaviorally grounded simulation.

We study how agents coordinate models and tools to solve tasks, and how computational agents can represent human behavior in social environments.

AGENTIC COLLABORATION

Coordinate expertise.

Task organization, model collaboration, and feedback-driven optimization.

SOCIAL SIMULATION

Understand collective behavior.

Spatially grounded individual behavior, population dynamics, and simulation reliability.

01
AGENTIC COLLABORATION

TrajAgent

TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration

NeurIPS 2025

Yuwei Du*, Jie Feng*†, Jie Zhao, Yong Li†

Coordinate models to solve trajectory tasks.

TrajAgent coordinates LLM-based agents with specialized models through a unified execution environment. Its workflow organizes task interpretation, model selection, execution, and result analysis.

Feedback guides subsequent optimization. The focus is task-oriented orchestration and large-and-small model collaboration.

PaperCodearXivProject
TrajAgent — original framework figure: a unified environment and collaborative workflow for trajectory modeling.
TrajAgent — original framework figure: a unified environment and collaborative workflow for trajectory modeling. SourceClick the figure to view full size.
02
HUMAN MOBILITY

AgentMove

AgentMove: A Large Language Model based Agentic Framework for Zero-shot Next Location Prediction

NAACL 2025

Jie Feng, Yuwei Du, Jie Zhao, Yong Li

Connect individual behavior with urban context.

AgentMove combines individual mobility memory, urban spatial knowledge, and population-level mobility patterns to predict a person’s next location in a zero-shot setting.

Structured reasoning links individual behavior with collective regularities across cities.

CodePDFPaper
AgentMove combines spatial-temporal memory, world knowledge, and collective mobility knowledge for next-location prediction. Original framework figure from the authors’ repository.
AgentMove combines spatial-temporal memory, world knowledge, and collective mobility knowledge for next-location prediction. Original framework figure from the authors’ repository. SourceClick the figure to view full size.
03
BEHAVIORAL GROUNDING

GAMS

GAMS: Geospatial Knowledge Enhanced Agentic Framework for Urban Human Mobility Behavior Simulation

EMNLP 2026 Findings

Yuwei Du, Jie Feng†, Jian Yuan, Yong Li†

Connect individual behavior with spatial context.

GAMS investigates how geospatial knowledge and human mobility patterns can inform agent-based behavior generation. It links individual behavioral preferences with urban spatial context and collective mobility constraints.

This research examines how spatial knowledge and behavioral evidence improve the realism of urban mobility simulation.

RepositoryEarlier preprint - CAMSPaper
GAMS — Figure 1. MobExtractor, GeoGenerator, and TrajEnhancer. From the Findings of EMNLP 2026 paper.
GAMS — Figure 1. MobExtractor, GeoGenerator, and TrajEnhancer. From the Findings of EMNLP 2026 paper. SourceClick the figure to view full size.
04
COLLECTIVE SOCIAL PATTERNS

OpenCity

Designing A Scalable LLM Agent Framework for Large-scale Urban Segregation Simulation

KDD 2026 AI4S

Qingbin Zeng, Yuwei Yan, Zhiheng Zheng, Jingzhe Yuan, Jun Zhang, Jie Feng, Fengli Xu, James Evans, Yong Li

Study urban segregation through individual activity.

OpenCity provides a scalable framework for examining urban segregation through LLM-based agent simulation. It supports investigation of how heterogeneous individual activities relate to collective spatial patterns.

Larger populations become tractable while agent-specific dynamic information is retained. Explore the inference systems perspective

PaperCodeEarlier preprint
Fig. 2OpenCity Figure 2: the Group-and-Distill Meta-Prompt Optimizer groups agents by static attributes while retaining individual dynamic properties Fig. 3OpenCity Figure 3: the earlier preprint's scalability experiment plots time per agent against population size, with baseline and average curves
OpenCity Figures 2 and 3 from the earlier preprint (arXiv:2410.21286v1): the Group-and-Distill Meta-Prompt Optimizer and the reported scalability experiment. SourceClick the figure to view full size.
05
POPULATION RELIABILITY

APS

APS: Bias-Controlled Adaptive Prototype Simulation for Population-Scale LLM Agents

NeurIPS 2026 Oral

Quan Zheng, Yan Gao, Shaobin He, Haoxiang Guan, Yuanhe Tian, Jie Feng, Ming Wang, Shuxin Zheng, Zhen Liu

Preserve variation under approximation.

APS examines how population-scale simulations can remain informative when only a subset of agents receives direct LLM queries. Adaptive prototypes are combined with individual treatment of selected heterogeneous agents and shadow-audit correction.

The goal is to control deviations from full-LLM simulation while preserving important population variation. Agreement with this reference does not by itself establish fidelity to real human behavior.

Paper
APS — Figure 1. Adaptive prototype simulation framework (arXiv version).
APS — Figure 1. Adaptive prototype simulation framework (arXiv version). SourceClick the figure to view full size.
RESEARCH OUTLOOK

Questions we are exploring next.

We are interested in coordination under changing tasks, heterogeneous agents, and dynamic environments. Evaluation should distinguish task success, behavioral plausibility, and the reliability of population-level conclusions.

Back to the research program