Logical Discrete Graphical Models Must Supplement Large Language Models for Information Synthesis

TL;DR

The paper tackles the challenge of synthesizing information from multiple sources with reliability, arguing that large language models alone cannot achieve general intelligence for information synthesis due to hallucinations, complex reasoning, planning under uncertainty, and calculations. It proposes a logical discrete graphical modeling framework that encodes reasoning in first-order logic fragments (Horn clauses, query and planning fragments) and structures (knowledge, implication, and proposition graphs) to support deterministic and probabilistic inference. The framework includes deterministic conjunction/disjunction nodes and a probabilistic or gate modeled as a linear exponential family, enabling scalable, interpretable reasoning and grounding in causal structure. Training and inference are designed to work with unlabeled text via parse generation and expectation-maximization, combining structured logical parsing with probabilistic learning and string-to-logic grounding to reduce hallucinations and enhance exact reasoning and multi-viewpoint analysis. Overall, the approach aims to provide reliable, explainable information synthesis that can outperform retrieval-augmented or purely discriminative strategies by construction, with potential for improved exact computation and multi-source synthesis in real-world IR tasks.

Abstract

Given the emergent reasoning abilities of large language models, information retrieval is becoming more complex. Rather than just retrieve a document, modern information retrieval systems advertise that they can synthesize an answer based on potentially many different documents, conflicting data sources, and using reasoning. We review recent literature and argue that the large language model has crucial flaws that prevent it from on its own ever constituting general intelligence, or answering general information synthesis requests. This review shows that the following are problems for large language models: hallucinations, complex reasoning, planning under uncertainty, and complex calculations. We outline how logical discrete graphical models can solve all of these problems, and outline a method of training a logical discrete model from unlabeled text.