Osbert Bastani’s research lies at the intersection of machine learning and programming languages, with a particular focus on building trustworthy artificial intelligence systems. He develops methods for creating software that can safely and reliably incorporate machine learning models, even when those models are inherently imperfect. His broader goal is to make it easier to design dependable software in an era when AI is increasingly embedded in applications ranging from robotics to healthcare.
A major theme of his work is developing ways to understand, verify and guide the behavior of learning-based systems. His research combines techniques from formal methods, program synthesis and reinforcement learning to give developers greater confidence that AI-enabled systems will behave as intended. Recent work includes specification-guided reinforcement learning, methods for making AI agents more reliable and cost-effective, and approaches that improve the safety and interpretability of systems used for complex decision-making.
Bastani also explores how machine learning can improve the process of developing software itself. His recent research investigates how large language models can optimize computer programs, assist with programming tasks and interact with software tools more effectively, while maintaining performance, security and reliability. Across these projects, he aims to connect advances in machine learning with rigorous computer science techniques so that AI systems can become more useful, dependable and practical in real-world settings.