Get Started
Home
Topics
Search
Library
Research questionHow can noisy LLM conditionals support coherent probabilistic inference over structured variables without order-dependent generation?LLMs encode useful world knowledge, but their next-token conditionals do not automatically define a coherent joint distribution over multiple interdependent variables. One-pass generation can make inferred structures depend on the order in which variables are produced.
AI
Machine Learning
Natural Language Processing
Reasoning
Research Paper
Statistical Machine Learning
Technology
Latest papersRecent research connected to this question, newest first.Structured Inference with Large Language GibbsThe work studies Large Language Gibbs, which uses an LLM’s conditional distributions to iteratively resample individual structured variables. Evidence is reported on synthetic distributions, consistent reasoning tasks, and Bayesian structure learning, suggesting applicability to structured probabilistic inference when the relevant world prior is accessible through noisy LLM conditionals.research paper · Sep 2, 2026
Related questions
How can grammar-constrained decoding preserve syntactic validity without distorting an LLM’s output probabilities?How can language models perform complex logical reasoning without accumulating token-level errors?How can we detect internally incoherent language-model forecasts before relying on them for consequential decisions?Can language models infer others’ mental states as social interactions evolve under unreliable information?