As artificial intelligence continues to shape everything from autonomous cars to home robotics, one Boston University PhD candidate is working to ensure that these systems not only function but also work safely.
Zijian Guo, a second-year PhD candidate in Systems Engineering advised by Associate Professor Wenchao Li (ECE, SE, CS), was recently named a Hariri Institute Graduate Research Fellow for his innovative work in neuro-symbolic AI. Guo’s research was also recognized with the Best Paper Award in CISE’s 2025 Best Student Paper Competition for his paper, “Constraint-Conditioned Actor-Critics for Offline Safe Reinforcement Learning.” This prestigious award recognizes the scientific quality and impact of research conducted by CISE students.
Guo is developing AI systems that can reason, adapt, and avoid failure in dynamic, real-world environments.
“My research focuses on safe and generalizable decision-making,” Guo said. “I integrate formal methods, such as linear temporal logic, with reinforcement learning to enable agents to achieve complex tasks safely and adapt effectively to new ones.”
Traditional reinforcement learning methods often struggle to handle tasks with complex temporal and logical structure and tend to generalize poorly to novel scenarios. Guo’s approach combines RL with temporal logic to create systems that can generalize behaviors across a wide range of tasks and safety needs in dynamic settings, which is especially important for safety-critical applications such as autonomous vehicles.
Through the fellowship, Guo plans to scale up his research.
“Attending conferences means I can meet other researchers and learn from their work. In this area, it’s about collaborations and communications.”
He also credited his advisor, Li, for helping guide the direction of his work. “Professor Li is an expert in formal methods and trustworthy AI,” Guo said. “That background helped me see how we could use logic to guide learning and make AI more reliable.”
Guo’s long-term goal is to make AI systems more adaptable and resilient in the face of change.
“Most RL models are trained for one specific task or a set of tasks,” he said. “But in the real world, things are always changing. My goal is to make these models more flexible so they can keep up without starting from scratch every time.”
Guo said the fellowship is both a milestone and a launchpad. “It feels like a real commitment to my research,” he said. “It’s motivating me to keep pushing forward.”