aiqbench analysis
This paper introduces Large Language Gibbs, a method that leverages LLM conditional distributions as transition operators within Markov Chain Monte Carlo (MCMC) frameworks to perform structured probabilistic inference. By iteratively resampling variables rather than relying on single-pass autoregressive generation, the approach mitigates order-dependent biases and achieves a stationary distribution that better reflects complex world priors. The study demonstrates practical viability in consistent reasoning tasks and Bayesian structure learning, offering a robust alternative to standard generation techniques for handling structured uncertainty.
Structured Inference with Large Language Gibbs
arXiv:2606.19264v2 Announce Type: replace-cross Abstract: The knowledge encoded in large language models (LLMs) can serve as a substrate for structured reasoning over variables describing a complex world, but accessing this knowledge in a probabilistically coherent manner poses a difficult inference problem. We propose Large Language Gibbs, a scheme for structured probabilistic inference that uses conditional distributions of an LLM as transition operators. Rather than sampling structured objects through single-pass autoregressive generation, we iteratively resample individual variables conditioned on others using an LLM's next-token conditionals. This approach avoids order-dependent biases and produces a stationary distribution that reflects a compromise between all local conditionals. We apply this approach to sampling from synthetic distributions, consistent reasoning tasks, and Bayesian structure learning. The results suggest that the use of LLM conditionals in MCMC is a practical alternative to one-pass generation for structured probabilistic inference under a world prior accessible through noisy LLM conditionals.
Abstract:The knowledge encoded in large language models (LLMs) can serve as a substrate for structured reasoning over variables describing a complex world, but accessing this knowledge in a probabilistically coherent manner poses a difficult inference problem. We propose Large Language Gibbs, a scheme for structured probabilistic inference that uses conditional distributions of an LLM as transition operators. Rather than sampling structured objects through single-pass autoregressive generation, we iteratively resample individual variables conditioned on others using an LLM's next-token conditionals. This approach avoids order-dependent biases and produces a stationary distribution that reflects a compromise between all local conditionals. We apply this approach to sampling from synthetic distributions, consistent reasoning tasks, and Bayesian structure learning. The results suggest that the use of LLM conditionals in MCMC is a practical alternative to one-pass generation for structured probabilistic inference under a world prior accessible through noisy LLM conditionals.
EMNLP 2026. Code: this https URL
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2606.19264 [cs.LG]
(or arXiv:2606.19264v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.19264 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 246 words
Related coverage
Test-Time Adaptation of Reasoning Strategies with Bayesian Nonparametric Memory
arXiv · Oct 6, 2026
From Plato to Aristotle: Toward a General Theory of Representation Hypotheses
arXiv · Oct 15, 2026
[AINews] not much happened today
Latent Space · Sep 10, 2026
Opaque recurrence, and other AI terms that you should probably know
TechCrunch · Sep 7, 2026