Today, reasoning is commonly interpreted as large language models generating chains of thought. Yet historically, AI reasoning had a very different meaning: executing algorithms that manipulated symbols to perform logical or probabilistic deduction and derive definite answers to questions about knowledge. In this talk, I show that such old-fashioned ideas are very relevant to reasoning with large language models today. In particular, I will demonstrate that integrating symbolic reasoning algorithms directly into the architecture of language models enables state-of-the-art capabilities in controllable text generation, alignment, and mathematical reasoning. These capabilities are built on top of tractable probabilistic circuit models that approximate the distribution of the large language model’s future behavior, and allow for efficient reasoning on the GPU. I will further show that the same ideas naturally extend to neurosymbolic offline reinforcement learning and image diffusion.

Seminar Recording: https://drive.google.com/file/d/1PF5GxivBHKiloFdKTv59NMae5dtXDzSk/view?usp=sharing