My research interests focus on:

• Multimodal-in and Multimodal-out Agentic Systems: I study agentic systems that can understand heterogeneous multimodal inputs and produce multimodal outputs, integrating understanding, generation, reasoning, planning, and action within a unified framework. I am particularly interested in native multimodal foundation models that support long-horizon interaction and the creation of structured, editable artifacts.
• Recursive Self-improvement Agentic Systems: I study agentic systems that improve recursively through rollout trajectories and accumulated experience. In these systems, model parameters and the surrounding agentic infrastructure—including harnesses, tools, memory, orchestration, evaluation, and feedback loops—are joint optimization targets that co-evolve, rather than treating either the model or its scaffold as fixed.