Supposedly Equivalent Facts That Aren't? Entity Frequency in Pre-training Induces Asymmetry in LLMs

2025-03-28

AI Summary

This work identifies a fundamental cause of hallucination in large language models (LLMs): systematic asymmetry in recognizing logically equivalent facts, rooted in skewed frequency distributions of entities as subjects versus objects in pretraining corpora. Method: We construct an entity-frequency index using OLMo/Dolma and design a controlled probe dataset grounded in Wikidata5M, employing zero-shot fact verification to isolate the effect of syntactic role frequency. Contribution/Results: We establish, for the first time, a causal link between structural imbalances in pretraining entity frequencies and LLM hallucination. Empirical analysis reveals that factual accuracy is significantly higher (p < 0.01) when entities exhibit high subject but low object frequency; performance reverses under low-subject/high-object conditions; and no significant asymmetry emerges when both frequencies are high. These findings demonstrate that hallucination arises not merely from semantic or logical deficits, but from statistical biases encoded during pretraining—specifically, subject–object frequency asymmetry.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• 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

Abstract

Understanding and mitigating hallucinations in Large Language Models (LLMs) is crucial for ensuring reliable content generation. While previous research has primarily focused on"when"LLMs hallucinate, our work explains"why"and directly links model behaviour to the pre-training data that forms their prior knowledge. Specifically, we demonstrate that an asymmetry exists in the recognition of logically equivalent facts, which can be attributed to frequency discrepancies of entities appearing as subjects versus objects. Given that most pre-training datasets are inaccessible, we leverage the fully open-source OLMo series by indexing its Dolma dataset to estimate entity frequencies. Using relational facts (represented as triples) from Wikidata5M, we construct probing datasets to isolate this effect. Our experiments reveal that facts with a high-frequency subject and a low-frequency object are better recognised than their inverse, despite their logical equivalence. The pattern reverses in low-to-high frequency settings, and no statistically significant asymmetry emerges when both entities are high-frequency. These findings highlight the influential role of pre-training data in shaping model predictions and provide insights for inferring the characteristics of pre-training data in closed or partially closed LLMs.

Problem

• Explains why LLMs hallucinate due to pre-training data asymmetry
• Links model behavior to entity frequency discrepancies in pre-training
• Demonstrates asymmetry in recognizing logically equivalent facts

Innovation

• Links model behavior to pre-training data
• Uses OLMo series to estimate entity frequencies
• Constructs probing datasets from Wikidata5M triples