Jiayuan Mao

Assistant Professor

Computer and Information Science

Personal Site

Jiayuan Mao is an Assistant Professor in Computer and Information Science at the University of Pennsylvania and a primary faculty member of the GRASP Laboratory. Her research seeks to understand the principles that enable intelligent machines to continually acquire concepts from experience and use them to reason and plan in open-ended physical worlds. Her long-term goal is to build agents that can generalize far beyond their training experience—reasoning across tasks, sensory modalities, and environments, understanding human activities and intentions, and carrying out complex, long-horizon behaviors. Robotics serves as a central testbed for this research, which spans machine learning, computer vision, robotics, and natural language processing.

A central theme of Mao’s research is neuro-symbolic intelligence: building systems that combine structured abstractions with the flexibility of neural learning. Rather than treating symbolic reasoning simply as logic layered on top of neural networks, her work asks what minimal structure an intelligent agent needs in order to learn and reason effectively. These structures may take the form of objects and relations, geometric and physical abstractions, compositional skills, planning representations, or programs. Neural models ground these abstractions in perception and physical interaction, while algorithms based on search, optimization, and probabilistic reasoning manipulate them to solve new problems. Importantly, the representations and reasoning mechanisms can themselves be learned and adapted from data.

Her work develops both representations of concepts and the algorithms for learning, reasoning, and planning with them. On the representation side, Mao studies compositional structures spanning language, visual concepts, 3D scenes, human activities, and robot actions. Her research on robotic manipulation, for example, has explored representations based on object parts, keypoints, contact mechanisms, and hybrid declarative-imperative behaviors. On the algorithmic side, she studies how the structure of tasks, the composition of training data, and inference procedures jointly determine generalization, with approaches ranging from differentiable logical reasoning to neuro-symbolic planning, constraint solving, and optimization-based inference.

More recently, Mao’s research has investigated how these ideas interact with foundation models and large-scale multimodal learning. Across these projects, the goal is not simply to scale models, but to understand what representations, data, and inference mechanisms allow scaling to produce robust and systematic generalization.