Abstract
Modern artificial intelligence has largely followed a train-then-deploy workflow: models acquire intelligence from large-scale datasets and extensive training pipelines before deployment, while interaction is primarily treated as a way to access that intelligence. Such models excel at producing one-time responses, but usually struggle with complex, evolving, and unfamiliar interaction environments. Many real-world tasks, however, require intelligent systems to learn efficiently from ongoing interactions and continuously adapt their decisions and behaviors accordingly.
In this talk, I introduce Interactive Intelligence, a research vision for moving AI from trained intelligence toward systems that continuously “learn through interaction to become better at interaction”. I will present our work along three complementary pillars. First, I discuss learning to optimize under limited interaction budgets, where Bayesian optimization provides a principled framework for adapting systems to human performance, behaviors, and preferences with minimal interactions. Second, I present learning for long-horizon planning, where reinforcement learning and self-supervised approaches enable agents to acquire increasingly complex behaviors without relying on manually specified rewards. Finally, I discuss learning to interact under unfamiliar dynamics, where agents learn models of previously unknown users and environments from interaction, allowing them to anticipate how the world responds to their actions and adapt accordingly.
Together, these directions (learning to interact optimally, learning for long-horizon planning, and learning the dynamics of interaction) point toward a different view of AI: not as a static model whose intelligence is fixed at deployment, but as an interactive system that continually learns from acting, observing, and collaborating with humans and the world.
In this talk, I introduce Interactive Intelligence, a research vision for moving AI from trained intelligence toward systems that continuously “learn through interaction to become better at interaction”. I will present our work along three complementary pillars. First, I discuss learning to optimize under limited interaction budgets, where Bayesian optimization provides a principled framework for adapting systems to human performance, behaviors, and preferences with minimal interactions. Second, I present learning for long-horizon planning, where reinforcement learning and self-supervised approaches enable agents to acquire increasingly complex behaviors without relying on manually specified rewards. Finally, I discuss learning to interact under unfamiliar dynamics, where agents learn models of previously unknown users and environments from interaction, allowing them to anticipate how the world responds to their actions and adapt accordingly.
Together, these directions (learning to interact optimally, learning for long-horizon planning, and learning the dynamics of interaction) point toward a different view of AI: not as a static model whose intelligence is fixed at deployment, but as an interactive system that continually learns from acting, observing, and collaborating with humans and the world.
Bio
Yi-Chi Liao is an Assistant Professor in the Department of Computer Science and Information Engineering (CSIE) at National Taiwan University (NTU), where he leads the Interactive Intelligence Lab (IXILab). Before joining NTU, he was a postdoctoral researcher at ETH Zürich and Saarland University. He received his Ph.D. from Aalto University. His research lies at the intersection of human-computer interaction and machine learning, with a focus on building intelligent systems that learn and adapt through interaction. More information about his research is available on his personal website (yichiliao.com).