UCLA

Generative AI has become a transformative paradigm that enables machines to produce high-quality content such as images, language, and audio. However, beyond creating charming and coherent outputs, these systems must reason -- steering their generations to satisfy specific properties. For instance, in science and engineering, this capability could ensure that synthesized molecular structures obey physical constraints or that design blueprints meet safety standards. While sound reasoning techniques from classical symbolic AI can rigorously guarantee these properties, they are often computationally prohibitive and difficult to scale. As a result, many recent approaches rely on scalable yet unsound methods, such as chain-of-thought prompting, which prioritize efficiency over rigorous correctness. In this dissertation, I will discuss how to design tractable generative AI models as drop-in replacements of existing models like autoregressive Transformers and diffusion models, with the distinguishing capability of sound reasoning. I will demonstrate how such tractable generative models enable high-fidelity yet controllable generations in various domains, and highlight the importance of building generative models with intrinsic reasoning capabilities.