Work spans spatiotemporal modeling, meta-learning and graph learning with emphasis on generalization under distribution shift. Current focus includes robust LLM/vision-language agents, tool-integrated reasoning, calibration, and memory systems for reliable long‑horizon interaction.