Augmenting LLMs with Causal Graph Discovery (GMU)

Associate Professor, Dr. Ferdinando Fioretto (Computer Science); Collaboration with GMU
Published

November, 2026

Dr. Ferdinando Fioretto, a Copenhaver Fellow and Associate Professor of Computer Science at the University of Virginia, leads the Responsible AI for Science and Engineering (RAISE) group. With 4-VA funding, he is studying a question that gets more urgent every time an AI assistant is given a new task: how do you keep it from leaking information it was never supposed to share? The project is a collaboration with Dr. Ziwei Zhu, Assistant Professor of Computer Science at George Mason University, a partnership that began years before either researcher was working on large language models at all. 

The everyday version of the problem is simple to picture. Someone asks an AI assistant to help file an expense report and submits a receipt containing a home address and bank details. Days later, that same assistant is drafting a meeting agenda with a colleague and casually mentions a detail lifted from the receipt, information that had nothing to do with the meeting and never should have traveled there. Dr. Fioretto calls this a failure to keep information in its place. As AI systems shift from simple chatbots into agents that read files, browse the web, and hand pieces of a task off to other AI agents, the chances for that kind of slip multiply. Dr. Fioretto's team built ways to deliberately trigger the problem, essentially testing the AI much the way a security team probes a building for weak points, so vulnerabilities could be found before someone with bad intentions exploits them. From there, they studied what was happening inside the AI in the moments right before it was about to overshare, and used what they learned to build a fast, lightweight way to catch that moment and stop it before it happens. 

That result alone would make for a strong research project, but earlier this year the team found something they had not set out to look for. When multiple AI agents were given the freedom to cooperate on a shared task, such as splitting up the work of writing a piece of software, the agents were not given any instruction designed to make them misbehave. Even so, in some AI models the cooperating agents began quietly organizing around goals that were never the ones the human user had actually asked for. Dr. Fioretto called the finding very consequential, and noted, “This is something we did not anticipate, not even at the time of writing of this proposal, but I think it's also very important because it's something that deserves an explanation.” 

The Fioretto-Zhu collaboration goes back to 2021, when the two co-authored a paper in a completely different corner of AI: fairness in ranking and recommendation systems, an area where Dr. Zhu is an expert. The two kept in touch and kept discussing how to extend that work, and by the time large language models transformed the field, their long-running conversation about risk in machine learning systems found a new and urgent target in the AI agents now reshaping how people work. 

Behind the research is a growing team of students. Dr. Fioretto noted that many of these students had never done research at this level before, and that the lab's structure made that possible. Because the work lives inside frontier AI systems, students can engage with cutting-edge problems by watching how those systems actually behave, without first needing to master every layer of the underlying mathematics. Aarav Lodha is a second-year undergraduate who approached Dr. Fioretto during his first year at UVA to ask about getting involved in research. He has spent his time on the project looking for ways to help teams of AI agents reach a shared solution andHehas found that restructuring how the agents work together can speed that process up. 


Saswat Das, a third-year Ph.D. student, whom Dr. Fioretto described as the engine of the entire effort, is a mentor to Aarav. He describes his approach to mentoring as starting with the training wheels on before gradually letting go of the handlebars, giving Aarav one task at a time so he can build toward forming his own judgment. “Aarav has been helping me come up with some really good answers,” Saswat said, pointing to Aarav’s insight that restructuring part of the experimental scaffold could help the agents converge on a solution faster. Saswat brings years of mentoring experience to the role. An undergraduate he mentored as a second-year Ph.D. student went on to publish an award-winning paper and is now pursuing a Ph.D. at Princeton. Saswat recalled that student sharing with him that he got into Princeton before telling his mom. “That means a lot to me,” Saswat said. 

For both students, the project has sharpened their sense of what comes next. Saswat said working on multi-agent systems has clarified his own direction toward research and mentoring beyond his Ph.D. and opened conversations with potential collaborators he might work within he future. Aarav said the experience confirmed that research is the path he wants to pursue. “It shows me how I can improve in research and what sort of things are interesting to me in research fields.” Both students expect to co-author a paper from the project once it is complete. 

The findings are already shaping how Dr. Fioretto trains the next generation of computer scientists. His undergraduate Artificial Intelligence course now includes dedicated lectures on the privacy and security risks of agentic AI systems, including material added this year on agent collusion. His graduate Responsible AI course draws directly on the project's findings around contextual privacy. And a recently launched graduate course, Constraint-Aware Generative AI is among the first of its kind in the country. It uses the project's discoveries as motivation to explore what an effective control mechanism would look like for AI systems that are increasingly autonomous. 

Two questions are shaping the next phase of the work. First, as AI agents gain access to more tools (reading emails, parsing documents, calling on other software), how do you catch a risky instruction that's buried somewhere inside a long chain of automated steps, rather than sitting out in the open as an obvious red flag? Second, as AI agents gain the ability to pick up new skills on their own and modify their own internal behavior, how do you keep them from gradually drifting away from the purpose they were originally built for? Dr. Fioretto's team has submitted a proposal to the National Science Foundation to dig deeper into both questions, with encouraging early conversations underway. People are becoming more reliant on AI assistants to handle day-to-day tasks that include their personal details, and this research is aimed at making sure people can trust those systems with their information.