Feizi Receives $1M Award to Advance the Foundations of Reasoning AI Models
Soheil Feizi, an associate professor of computer science with an appointment in the University of Maryland Institute for Advanced Computer Studies (UMIACS), recently received a $1 million federal grant to advance foundational research that is focused on reasoning in artificial intelligence models.
The award—$200,000 per year for five years—is the result of Feizi receiving a Presidential Early Career Award for Scientists and Engineers (PECASE) earlier this year.
In January, he was one of 400 scientists and engineers nationwide honored by President Biden for their exceptional potential for leadership and novel research undertaken early in their scientific careers.
The second phase of the PECASE recognition is federal funding that supports scientific exploration that is far-reaching and innovative in developing connections between research and its impacts in society.
Feizi says the PECASE project has four research goals: to defend large language models (LLMs) against adversarial or jailbreaking attempts by strengthening their ability to recognize such attacks; to find weaknesses in how these models learn by systematically characterizing their chains of thought; to encourage these AI-driven models to rely upon transparent, step-by-step logic, via reinforcement learning techniques that reward such behavior; and to reveal more about the “reasoning tokens” that help these LLM technologies make decisions.
He says the PECASE funding gives him an opportunity to take risks and explore innovative approaches that include live activation probing and novel reinforcement-learning designs. He is particularly pleased that the PECASE-funded work could result in real-world usages, transforming theory into practical applications.
“This award empowers us to pursue visionary and fundamental research—such as fine-grained audits of transformer activations and custom reinforcement-learning objectives that directly shape reasoning behaviors—without cumbersome funding restrictions,” Feizi explains. “Already, it’s drawing in top-tier postdoctoral fellows and doctoral students [to my lab] who are passionate about tackling the toughest challenges in AI reasoning.”
Feizi is also a member of the Institute for Trustworthy AI in Law & Society (TRAILS), where his work involves building trust in new AI systems.
He is the founder and CEO of RELAI, a startup whose mission is to make reliable AI accessible for everyone. He was the recipient of the Army Research Office Early Career Program Award in 2023, as well as a National Science Foundation Faculty Early Career Development (CAREER) award in 2020.
Feizi says he is deeply grateful for this latest round of federal funding, noting that much of the award will support his students and postdocs who share the same passion that he does: discovering new ideas and technologies to make AI more trustworthy and accessible.
“I’m truly blessed to have a stellar team of students in the Reliable AI Lab at the University of Maryland whose expertise and drive will be central to every aspect of this project,” he says. “Their contributions are already shaping our path forward.”
—Story by Kathryn Klett, UMIACS communications group
