Human behavior,
agentic AI, and mobility systems.

01. LLM agents for travel behavior

I design LLM-agent simulations of human travel behavior at the city scale. The core problem is how to extend a small paired stated-preference survey into population-scale social-response experiments for emerging mobility services, including autonomous shuttles and demand-responsive transit.

This work continues my Yonsei capstone, where I built an RP–SP integrated simulator for individual-level demand prediction. The pipeline combines personal travel state, retrieval of analogous SP cases, and behavioral alignment to constrain LLM choice generation.

• Persona agents
• RP–SP integration
• Choice-set expansion
• Synthetic respondents
• Human alignment

02. Cooperative world models

A second research thread studies vehicle–infrastructure cooperative world models for autonomous driving at signalized intersections. The goal is to connect physical scene understanding, infrastructure information, and sequential decision-making in settings where a vehicle cannot safely reason from its own sensors alone.

This work is linked to a National Research Foundation of Korea project on Physical AI and cooperative autonomous driving.

• Physical AI
• World models
• Signalized intersections
• Cooperative perception
• Model-based RL

03. Trustworthy evaluation

Generating plausible text is not the same as recovering human travel decisions. I therefore evaluate mobility agents against held-out human responses and distinguish individual choice accuracy, aggregate share fit, calibration, and distributional behavior. I am interested in where domain constraints and retrieval improve alignment—and where they create new failure modes.

• Held-out evaluation
• Calibration
• Distribution fit
• Robustness
• Reproducibility

Selected applied systems