Research
Four pillars of real-world intelligence: Reasoning, Embodied, Agentic, and Lifelong-learning AI. We study the fundamental capabilities that connect them, not just each in isolation, toward AI that thinks, acts, and grows in the real world.
Reasoning
We study how large models solve problems that require multi-step inference, from mathematical and code reasoning to long-chain and multimodal reasoning. Our goal is to make reasoning reliable, efficient, and verifiable, so that large models can think and decide as deeply as (or beyond) humans.
- Long-chain and infinite-horizon reasoning
- Mathematical and code reasoning
- Chain-of-thought tuning, process verification, and reward modeling
- Multimodal and spatial reasoning in vision-language models
- Social reasoning and theory of mind for LLMs
Agentic AI
We build AI agents that plan, use tools, and carry long-horizon tasks through to completion. On the application side we focus on deep research agents, code agents, computer-using agents, and embodied agents, moving from scripted workflows to genuinely autonomous behavior across both digital and physical environments.
- Deep research agents: long-horizon search, evidence grounding, and report synthesis
- Code agents: repository-level coding and adversarial code/test co-evolution
- Computer-using agents: GUI grounding and mobile / desktop automation
- Embodied agents: perception, tool use, and action in physical and simulated environments
- Agent memory, multi-agent collaboration, and policy-level reflection
Lifelong-learning AI
Deployment is not the end of training: it is where lifelong learning begins. We study how models accumulate experience through continual interaction with their environment and bootstrap their own capabilities, through reinforcement learning, self-improvement, and feedback-driven tuning, so that intelligence keeps growing throughout its lifetime rather than plateauing after release.
- Reinforcement learning for reasoning and agents
- Self-evolving agents and self-improvement loops
- Generator/verifier co-evolution and policy/reward co-optimization
- LLM steering and efficient adaptation after deployment
Embodied AI
We study how AI systems perceive, reason about, and act in environments: closing the loop from understanding to reasoning to action. Our current focus is on spatial intelligence, embodied reasoning, and benchmarks that evaluate agents in interactive physical scenes.
- Spatial reasoning and multi-perspective localization
- Embodied reasoning that synergizes search, planning, and action
- Benchmarks for agent reasoning in embodied tasks
- World models for physical and interactive environments