image: Nicola Bezzo, associate professor of electrical and computer engineering at the University of Virginia’s School of Engineering and Applied Science, has received a 2025 Young Faculty Award from DARPA.
Credit: Matt Cosner, University of Virginia School of Engineering and Applied Science
Nicola Bezzo, associate professor of electrical and computer engineering at the University of Virginia’s School of Engineering and Applied Science, has received a 2025 Young Faculty Award from the Defense Advanced Research Projects Agency, or DARPA, for his project titled, “LLM Embodiment for Collaborative Autonomy.” The prestigious award recognizes rising research leaders whose work shows exceptional promise for national security and societal impact.
An LLM, an abbreviation of “large language model,” is a type of artificial intelligence trained on huge amounts of text so it can understand and generate human-like language. LLM programs use patterns in how words and ideas connect, allowing them to answer questions, write stories, explain concepts or carry on a conversation in a natural way.
Bezzo’s project will enable embodied cognition capabilities in large language models deployed in human-robot applications, equipping robots with the ability to understand and anticipate human intent — much like a teammate who can “read the room.” The work has wide-ranging potential, from assistive technologies in homes and hospitals to collaborative robots on factory floors and defense missions.
“As humans we can predict the behavior of others, empathize, adapt our own plans accordingly, and coordinate effectively in complex situations, even in the absence of explicit communication,” Bezzo said. “Our goal is to bring that same capability to human-robot partnerships, so machines can become more reliable and more autonomous collaborators in everyday and mission-critical settings.”
To achieve this outcome, Bezzo’s lab is developing an active epistemic planning framework — a strategic pursuit of understanding — that integrates LLMs with “theory-of-mind” reasoning in real-world dynamic settings, enabling robots to infer what people are trying to do and adjust their support in real time.
This framework will allow Bezzo’s robots not just to react to physical cues, but also to anticipate human knowledge, beliefs and expectations in the moment itself. Robots will gather information by observing the surrounding environment as well as human actions. This involves a need for “second order” reasoning — requiring the robot to model not just the task itself, but also the mental state of the human partner. The robots then use LLMs and inference techniques to update their theory-of-mind model.
DARPA’s Young Faculty Award accelerates the impact of early-career researchers like Bezzo, providing up to $1 million over three years to help them shape the future of emerging technologies.