Soheil Feizi is the founder and chief scientific officer of RELAI and an associate professor of computer science at the University of Maryland, College Park. He develops methods for making artificial intelligence resilient to manipulation and capable of learning from experience without compromising previously reliable behavior.
Feizi earned his bachelor’s degree from Sharif University of Technology in 2008, his master’s degree from MIT in 2010, and his MIT doctorate in electrical engineering and computer science in 2016. After postdoctoral research at Stanford, he joined Maryland in 2018, where his Reliable AI Lab studies adversarial robustness, generative models, interpretability, hallucinations, machine unlearning, and synthetic-content provenance.
His research has exposed important weaknesses in efforts to identify AI-generated material. Work on synthetic-text detection demonstrated that repeated paraphrasing can evade detectors and that spoofed signatures can falsely implicate human-written text. Complementary research on image watermarks and deepfake detection identified vulnerabilities involving watermark removal, spoofing, and false attribution. The practical implication is that provenance systems must withstand deliberate attacks, not merely perform well under controlled conditions.
Feizi contributed to the U.S. House Bipartisan Task Force on Artificial Intelligence and is a co-investigator in the Institute for Trustworthy AI in Law & Society. He also helped establish Maryland’s Rising Stars in Machine Learning program for researchers from underrepresented backgrounds. His honors include an NSF CAREER Award in 2020, an Office of Naval Research Young Investigator Award in 2022, and the Presidential Early Career Award for Scientists and Engineers in 2025.
He founded RELAI in 2024 to bring reliability research into the operational lives of AI agents. His approach centers on several connected ideas:
- Verifiable continual learning: Agent improvements should be executable, measurable, and checked against earlier successful behavior; a production failure becomes a durable test case, not an isolated patch. His AI Engineer World’s Fair talk identifies replayability, whole-agent analysis, lifelong learning, and efficiency as essential requirements.
- Replayable learning environments: Production logs and expert feedback become testable simulations combining synthetic users, real or mocked tools, and explicit evaluators.
- Model, harness, and memory: Corrections can target model weights, prompts and workflows, or persistent memory; the appropriate intervention is the smallest durable change that addresses the actual failure.
- Regression-aware optimization: Previous capabilities become constraints within optimization itself. His Terminal-Bench continual-learning analysis also evaluates transfer to unfamiliar tasks and whether improvements survive repeated optimization cycles.
