Tractable Control for Autoregressive Language Generation

Explore a cutting-edge approach to constrained text generation in this 47-minute lecture by Honghua Zhang from UCLA. Delve into the challenges of generating text that satisfies complex constraints using autoregressive large language models. Learn about the innovative GeLaTo (Generating Language with Tractable Constraints) framework, which utilizes tractable probabilistic models (TPMs) to impose lexical constraints in autoregressive text generation. Discover how distilled hidden Markov models can efficiently compute conditional probabilities to guide GPT2 generation. Examine the state-of-the-art performance achieved by GeLaTo on challenging benchmarks like CommonGen, surpassing strong baselines. Gain insights into the potential of this approach for controlling large language models and the future development of more expressive TPMs. This talk, part of the Probabilistic Circuits and Logic series at the Simons Institute, offers a deep dive into advanced techniques for improving text generation with complex constraints.