Multimodal Reasoning and Agents

Our research investigates how to elevate machine capabilities on complex tasks through advanced reasoning and environment interaction. We explore the fundamental mechanisms of artificial cognition, studying whether models reason most effectively through natural language, visual modalities across long time horizons, or external tool integration like code execution and search. Beyond internal deliberation, we focus on grounding these models to interact autonomously with physical systems like robotics and digital spaces like web navigation, bridging high-level reasoning with real-world execution.