research

My research centers on spatial intelligence for multimodal models. I am interested in systems that do more than recognize objects in individual frames: they should maintain spatial state over time, distinguish object motion from observer motion, reason across viewpoints, and preserve relevant evidence throughout a long video.

Dynamic spatial intelligence

I study how spatial intelligence should be defined and evaluated in dynamic scenes, and how video MLLMs can use explicit geometric trajectories rather than relying only on appearance. This line of work spans capability formulation, physically verifiable benchmarks, large-scale training data, and 2D–4D representation alignment.

Long-video spatial memory

Long videos challenge both context length and representation quality. I am exploring vision-side fast-weight memory that accumulates spatial evidence before visual tokens lose their grid structure, allowing geometry and long-term memory to interact before language decoding.

Interactive multimodal agents

I also build agents that convert ambiguous user intent into scalable data systems. Instead of requiring a complete specification upfront, these agents refine both user requirements and their own policies through lightweight interaction, then transition to automatic collection.

Multi-view and 3D vision

My earlier work focused on large-scale real-world multi-view data and accurate 3D reconstruction. This background informs my current interest in connecting 2D appearance, 3D geometry, temporal trajectories, and multimodal reasoning.