Structured Inference with Large Language Gibbs

The article introduces Large Language Gibbs, a novel scheme for structured probabilistic inference utilizing large language models (LLMs) as transition operators for conditional distributions. This method iteratively resamples individual variables based on others, leveraging the next-token conditionals of LLMs to mitigate order-dependent biases and achieve a stationary distribution. The approach has shown promise in applications such as sampling from synthetic distributions and Bayesian structure learning, providing a viable alternative to traditional one-pass autoregressive generation for practitioners focused on structured reasoning in AI.