research
Selected research projects
UrbanLLMind
COLM 2026 · ongoing
A simulated San Francisco where every resident is an open-weight LLM agent that plans its day, decides where to go, remembers, and reflects. It scales to 10,000 agents on local GPUs, matches real travel diaries, and reacts to events like a storm warning.
Urban Mobility Synthesis
SIGSPATIAL 2024 (Best Paper) · ACM TSAS 2026
Can an LLM fill in a travel diary like a survey participant? Fine-tuned on 10,000 real diaries, open models produce synthetic travel surveys that match real ones across U.S. cities, age groups and seasons.
Geospatial Knowledge of LLMs
SIGSPATIAL 2023
Do language models know where cities are, whether "near" means nearby, and enough about distances to redraw a map? Probing open models with prompts alone shows real geospatial knowledge, awareness and reasoning that grow with model size and quality.
From Text to Maps
NLP4PI @ EMNLP 2024 (Outstanding Paper)
A two-step LLM pipeline that reads outbreak reports, extracts case counts as structured records and geotags them to administrative regions with reasoning a person can check. On Rift Valley fever it finds every record in a human-built dataset and more than twice as many correct ones.