Soheil Feizi — Reliable AI Lab, University of Maryland

Short Bio

Dr. Soheil Feizi is an Associate Professor of Computer Science at the University of Maryland, College Park. He is also the Founder of RELAI.ai, a startup focused on continual learning for AI agents. His work focuses on the reliability, safety, and optimization of AI systems. He received his Ph.D. from MIT and was a postdoctoral researcher at Stanford University.

Dr. Feizi is a recipient of the 2025 Presidential Early Career Award for Scientists and Engineers (PECASE), the highest honor granted by the U.S. government to early-career scientists and engineers. His honors also include the ONR Young Investigator Award, the NSF CAREER Award, the ARO Early Career Program Award, two best paper awards, the Ernst Guillemin Thesis Award, a teaching award, and more than twenty research awards from federal agencies and industry partners.

His research has been featured in The New York Times, The Washington Post, BBC, MIT Technology Review, Bloomberg, and The Wire. In 2024, he testified before the U.S. House Bipartisan Task Force on AI on issues related to AI safety and reliability.

Research

Toward AI we can rely on.

01 / Robustness

Robustness

Provable and practical defenses against natural and adversarial perturbations across vision, language and multi-agent systems.

02 / Interpretability

Interpretability

Understanding, localizing and editing knowledge in foundation and generative models — and exposing spurious cues and hidden biases.

03 / Provenance

Provenance & Detection

The fundamental limits and real-world fragility of AI-text and AI-image detection, watermarking, and content authentication.

04 / Unlearning

Machine Unlearning

Recovering, removing or verifying knowledge in trained models — from data unlearning to test-set contamination flags.

05 / Agents

Reliable Agents

Why agentic LLMs misjudge tools and time, and how to make their reasoning consistent, faithful and dependable.

06 / Foundations

Foundation Models

Capabilities, limitations and generalization of reasoning models, studied with information-theoretic and statistical tools.

Highlights

Recent & notable.

      2025 · White HousePresidential Early Career Award (PECASE)Awarded the highest U.S. government honor for early-career scientists and engineers, recognizing work on reasoning AI models: capabilities, limitations and generalization.

      In the press · FeaturedOur work, covered widelyResearch featured in The New York Times, The Washington Post, MIT Technology Review and Bloomberg — plus a TV interview with CBS and coverage on BBC and Voice of America on AI reliability and generative AI.

      Jan 2026Google Research AwardA new gift and Gemini cloud credits supporting reliable foundation-model research.

      2024 — presentFounded RELAI.aiA platform to design, train, inspect and improve reliable AI systems. Visit →

      NSF AI InstituteTRAILS — Trustworthy AI in Law & SocietyCo-Investigator in the NSF AI Institute bridging AI reliability and society. Read more →

      2024 · Washington, D.C.Testimony to the U.S. HouseTestified before the U.S. House Bipartisan Task Force on AI on AI safety and reliability.

Prospective students & post-docs

I'm looking for students and post-docs interested in the theoretical and practical foundations of reliable AI/ML. Information for prospective graduate students is available through the CS@UMD catalog.

Google Scholar DBLP LinkedIn Twitter / X

Meet the lab