We build AI systems that go beyond intelligence fixed at deployment. They learn through interaction, adapt to human preferences, understand their environments, and plan actions to achieve complex goals.
How do we move AI from intelligence that is merely trained to intelligence that can continuously interact, learn, and adapt?
Most modern AI follows a train-then-deploy paradigm. We study what changes when intelligence is treated instead as an ongoing process of interaction, adaptation, and decision-making.
How can AI systems continually improve through interaction with humans? We develop sample-efficient optimization methods that enable AI systems to adapt their behavior to individual preferences, goals, and feedback.
How can AI understand design intent and turn it into effective actions? We develop intelligent agents that can evaluate, generate, and iteratively improve designs through sequences of decisions.
How can intelligent agents understand unfamiliar environments and anticipate the consequences of their actions? We develop world models that enable agents to learn environment dynamics, predict future outcomes, and plan over long horizons.
Four tiny games, each a working version of a method behind our work: Bayesian optimization with you in the loop, and a reinforcement-learning agent you can try to outwit.
Our work spans adaptive interfaces, human-in-the-loop optimization, mixed reality, computational models of interaction, and learning-based intelligent systems.
Learning priors from simulated users to enable efficient human-in-the-loop optimization.
LLM-based agents are employed to configure optimization for real-time VR interface adaptation.
Learning population-level structure while adapting optimization to each new user.
Meta-learning prior preferences to accelerate personalized visual optimization.
Turning sparse preference signals into personalized multi-objective interface adaptation.
Fast calibration of wrist-based interactions with only a handful of user trials.
Hands-on courses spanning human-centered design, modern AI models, agents, and the evaluation of AI-native systems.
Computer Science & Information Engineering, NTU
Interactive Intelligence · Human–AI Interaction · Bayesian Optimization
Computer Science, NTU
Administrative Assistant
Computer Science, NTU
Representation learning · World model
Computer Science, NTU
Preferential learning · Model steering
Computer Science, NTU
Design agent · LLM benchmark
Computer Science, NTU
Test-time optimization · VLM + optimization
Computer Science, NTU
Design agent · LLM-as-judge
Computer Science, NTU
High-dimensional optimization · Graph-structured optimization
Computer Science, NTU
Interactive optimization library
A new opportunity appeared!
We are recruiting students across levels who are excited by machine learning, HCI, agents, optimization, or computational approaches to intelligent interactive systems.
See recruiting informationWe welcome researchers, students, and friends to visit us and chat about research. If you would like to stop by, please get in touch with Yi-Chi by email.
Department of Computer Science & Information Engineering
National Taiwan University
No. 170, Section 2, Xinhai Rd,
Da’an District, Taipei City 106,
Taiwan