Generative World Models
Build interactive 3D worlds from real-world observations.
UrbanWorld2.0ACM MM 2026
SPATIAL INTELLIGENCE × AGENTIC INTELLIGENCE
Intelligence for
an open world.
Jie Feng 冯杰Assistant Professor · Zhongguancun AcademyTsinghua University · B.S., Ph.D. & postdoctoral research
I study how AI understands and builds spatial worlds, and how agents reason, collaborate, and improve the systems they run on. My research connects spatial reasoning, world models, efficient intelligence, and human behavior. My work includes 30+ papers at CCF-A / CORE A* venues and has received 7,000+ citations.
OUR RESEARCH LANDSCAPESpatial × Agentic Intelligence
Open positions
Learn more PhD students · Postdoctoral researchers · Research interns
01
Research overview Two connected research lines
01 / SPATIAL Understand the world.
Build capable agents.
Understand the world.
Make it interactive.
Spatial foundation models, grounded reasoning, and generative worlds for open environments.
Spatial Intelligence ↗02 / AGENTICBuild capable agents.
Improve their systems.
AI4AI, embodied interaction, and multi-agent collaboration with measurable feedback.
Agentic Intelligence ↗GROUNDED IN URBAN SCIENCE Real spaces, human behavior, and collective dynamics.
02
All publications Selected research
Grounded Spatial Reasoning
Learn and evaluate how multimodal models reason about real-world places from visual evidence.
Efficient LLM & Agent Systems
Accelerate LLM inference and coordinate complementary models for capable agent systems.
Social & Behavioral Simulation
Model heterogeneous human behavior and simulate population dynamics at scale.
03
Recent updates
CacheSpec selected for an Oral presentation at EMNLP 2026
CacheSpec and GAMS accepted by EMNLP 2026
Congratulations to Jingquan Chen and Yuwei Du.
CityRiSE and UrbanWorld2.0 accepted by ACM MM 2026
Congratulations to Tianhui Liu, Shengyuan Wang, Zhiheng Zheng, and collaborators.
Earlier updates 23 updates · 2023–2026
UrbanWell and OpenCity accepted by KDD 2026
UrbanWell in the D&B Track; OpenCity in the AI4S Track. Congratulations to Yanxin Xi, Qingbin Zeng, and collaborators.
Embodied world models survey released with paper list
A regularly updated paper list is maintained online.
AutoQual and DORIS accepted by EMNLP 2025 Industry Track
Congratulations to Xiaochong Lan, Zhiguang Han, and collaborators.
MoveGCL and UniMove accepted by SIGSPATIAL 2025
Both papers were accepted by SIGSPATIAL 2025.
World models survey published in ACM Computing Surveys
A regularly updated paper list is maintained online.
CityGPT, CityBench, and LocalGPT accepted by KDD 2025
CityGPT in the Research Track; CityBench in the D&B Track; LocalGPT in the ADS Track.
Basic Spatial Abilities and PIGEON accepted by ACL 2025
Basic Spatial Abilities accepted by ACL; PIGEON accepted by ACL Findings.
Selected as a 2025 Spatial Data Intelligence Rising Star
Recognized by SIGSPATIAL China Chapter.
CityGPT and CityBench code released
Code repositories for both urban LLM benchmarks were released.
Selected as a Stanford/Elsevier Top 2% Scientist 2024
Selected in the 2024 Stanford/Elsevier Top 2% Scientists list.
High Efficiency Delivery Network and UniST accepted by KDD 2024
Both works were accepted by KDD 2024; UniST was later recognized as a highly influential KDD paper.
Urban Generative Intelligence preprint released
A preprint on urban generative intelligence was released.
Meituan delivery network selected as INFORMS Edelman finalist
Real-time intelligent dispatching algorithms for minute-level delivery networks.