Machine Learning Seminar: Debangshu Banerjee, "Formal Constraint-Guided Generative Modeling."

Abstract: Despite the impressive capabilities of generative models, they often ignore instructions, hallucinate, and, more concerningly, exploit task-specific checkers to inflate performance metrics. Providing formal guarantees of correctness and reliability for systems built on generative models remains an open challenge. In this talk, Debangshu will present his recent work on providing formal guarantees for generated outputs with respect to predefined specifications. More importantly, he will discuss how the formal structure of target outputs can be leveraged to guide the generation process. He will explore formal languages for specifying correctness and steering algorithms that guide generative models toward provable compliance, covering both single-turn and interactive, multi-turn generation.