Research
My research aims to advance hybrid intelligence by enabling humans and AI to better understand, collaborate with, and adapt to one another. Toward this vision, I pursue two complementary research directions:
- Novel Mechanisms for Human–AI Collaboration: I develop new paradigms for human–AI collaboration, including human–AI deliberation and trust calibration, to unlock the complementary strengths of human and machine intelligence. I also translate these advances into real-world AI applications, such as DBox and Glancee.
- User Modeling for Human-Aware AI: I develop computational models of human goals, preferences, and behaviors to help AI better understand the people it works with. I further investigate how these models can inform AI post-training, enabling more personalized, adaptive, and human-aware intelligence.
Recent Projects
- Understanding Intent in Agent-Assisted Coding
- When developers revise AI-generated code, can agents learn what they actually intended? Drawing on theory of mind, I investigate how humans and coding agents infer, communicate, and refine each other's intentions, moving toward more intuitive and adaptive coding collaboration.
- Data-Efficient User Modeling
- Can AI understand a person without needing extensive interaction data? I explore how to jointly model high-level goals and beliefs alongside observable behaviors, such as edits and corrections. My goal is to develop user models that learn from limited data and inform AI post-training, enabling systems to adapt to individual users.
- Human Oversight of AI Agents
- How can humans stay meaningfully in control without constantly monitoring AI? I investigate how agents can recognize when human intervention is needed, communicate their progress and uncertainty, and support timely, effective oversight while minimizing the burden on users.
- Long-Term Human–AI Alignment and Memory
- How can AI learn from past interactions without letting outdated assumptions shape future decisions? I study how agents can maintain alignment with evolving human goals and preferences, including what to remember, when to retrieve and apply past information, and when to update or forget it. Explore MemoSync.
Background
Before joining ISCAS, I was a postdoctoral researcher at the Computational Behavior Lab at Aalto University and the Finnish AI Center, working with Antti Oulasvirta. I received my PhD from the HCI Lab at HKUST, advised by Xiaojuan Ma. I earned my bachelor’s degree at HIT and my master’s degree at the Chinese Academy of Sciences.
I lead the Human-Aware AI (HAAI) Research Group within the ISCAS HCI Lab.
Group Vision
By human-aware AI, we mean AI that models the person it works with and uses that understanding to guide its behavior. Our vision is for AI to reason about a person’s intentions and knowledge, adapt to their preferences, and recognize when its assumptions about them need to change.
Our current projects connect high-level cognitive models with observable actions, such as edits and corrections, to learn about users from limited interaction. We investigate how these models can inform AI post-training and guide collaboration, including when to explain, offer help, or seek human review. An agent’s memory should also evolve as the person’s goals and circumstances change.
This understanding should work in both directions. People should be able to understand an AI’s reasoning, correct its assumptions about them, and decide when to rely on or redirect it. We bring user modeling together with new mechanisms for human–AI collaboration to pursue hybrid intelligence in which people and AI learn from one another while people retain agency over their decisions.
Join the group
We’re recruiting research assistants, interns, and postdoctoral researchers. Send your CV to mashuai@iscas.ac.cn with the subject [Master/RA/Postdoc Application] + Your Name.