projects

My research can be organized into four connected directions. They all study how learning systems behave when transportation data are sparse, shifted, safety-critical, or generated by interacting agents.

Safe Autonomous Driving

I study how to generate, search, and use realistic safety-critical traffic scenarios before autonomous driving systems are deployed. The goal is not only to find failures, but also to convert adversarial cases into training signals that improve long-tail robustness.

• SAGE: preference-aligned adversarial scenario generation for steerable testing.
• EvoDrive: Pareto evolution with self-improving LLM agents.
• ADV-0 and AlignADV: closed-loop adversarial training and learnability-guided curricula.
• LLM-attacker and S2C: LLM-based adversarial testing agents.

LLM Agents for Mobility

I am interested in language models as a new computational layer for transportation: they can encode geospatial semantics, synthesize realistic mobility records, support demand modeling, and coordinate agentic search in complex simulators.

• LLM4TR: a survey, framework, and roadmap for LLMs in transportation systems.
• MobEvolve: self-evolving heuristic agents for interpretable human mobility generation.
• LLMSynthor: macro-aligned micro-record synthesis with LLMs.
• LLM2GeoVec and LLM4Delivery: geospatial representations and LLM-empowered demand modeling.

Spatiotemporal Learning

This direction focuses on learning from sparse, noisy, and incomplete observations in city-scale systems. I develop models that combine low-rank structure, graph inductive bias, Transformers, and implicit neural representations for estimation, imputation, and forecasting.

• ImputeFormer: low-rankness-induced Transformers for generalizable spatiotemporal imputation.
• STINR and MetaINR: implicit neural representations for continuous and cross-city urban time series learning.
• Tensor4Kriging, GNN4Flow, and LRTC: sparse sensing and traffic state estimation.

Urban AI and Scalable Forecasting

I also work on efficient and interpretable modeling for large-scale urban dynamics. This line emphasizes scalable architectures, physical or structural priors, and transferable models that remain practical under real city data constraints.

• Energy-informed graph neural diffusion: interpretable neural diffusion for large-scale urban networks.
• ST-MLP-Mixer: contextualized MLP-Mixers for large-scale traffic forecasting.
• C-LoRA: channel-aware low-rank adaptation for time series forecasting.
• LLM4Delivery: joint estimation and prediction of city-wide delivery demand.