Mining the Mind: What 100M Beliefs Reveal About Frontier LLM Knowledge

2025-10-08

AI Summary

Contemporary large language models (LLMs), such as GPT-4.1, exhibit critical deficiencies in factual knowledge—namely, hallucination, internal inconsistency, and semantic ambiguity—yet existing evaluations rely on small-scale, biased benchmarks that inadequately reflect real-world knowledge reliability. Method: We systematically extract and analyze 100 million fact-belief statements generated by the model. To ensure representativeness and reduce sampling bias, we introduce Recursive Prompting to construct GPTKB v1.5, a large-scale, unbiased factual dataset. We then apply statistical modeling and cross-source validation against authoritative knowledge bases (e.g., Wikidata, DBpedia). Contribution/Results: Our analysis quantifies pervasive hallucinations and inconsistencies in LLM knowledge for the first time at scale, revealing significantly lower factual accuracy than reported on standard benchmarks, substantial distributional divergence from structured knowledge repositories, and severe methodological biases in current evaluation protocols. This work establishes a reproducible, large-scale empirical framework and a new benchmark paradigm for rigorous LLM knowledge assessment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Data Mining & Knowledge Management: Linked Open Data, Knowledge Graphs & KB Completion

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

LLMs are remarkable artifacts that have revolutionized a range of NLP and AI tasks. A significant contributor is their factual knowledge, which, to date, remains poorly understood, and is usually analyzed from biased samples. In this paper, we take a deep tour into the factual knowledge (or beliefs) of a frontier LLM, based on GPTKB v1.5 (Hu et al., 2025a), a recursively elicited set of 100 million beliefs of one of the strongest currently available frontier LLMs, GPT-4.1. We find that the models' factual knowledge differs quite significantly from established knowledge bases, and that its accuracy is significantly lower than indicated by previous benchmarks. We also find that inconsistency, ambiguity and hallucinations are major issues, shedding light on future research opportunities concerning factual LLM knowledge.

Problem

• Analyzing factual knowledge and beliefs in frontier LLMs using 100M data points
• Revealing significant accuracy gaps between LLM knowledge and established benchmarks
• Identifying inconsistency, ambiguity and hallucinations as major knowledge issues

Innovation

• Analyzes 100M beliefs from frontier LLM
• Uses recursively elicited GPTKB v1.5 dataset
• Reveals accuracy gaps via knowledge comparison