Learning Rules, Memorizing Facts: An Information-Theoretic View of Hallucination
Abstract
Understanding the relationship between generalization and hallucination requires clarifying the role of memorization. We study this relationship from an information-theoretic perspective through a Selector-Density Conditional Mutual Information (CMI) framework. In the sparse random-fact limit, a uniformly bounded information budget per fact forces a nonvanishing false-positive rate at recall bounded away from zero, even when all facts are observed during training. In a rules-and-facts setting, we decompose model-data mutual information into rule learning, atomic-fact memory, and excess memorization, revealing competition between rule learning and factual memory even without excess memorization. A joint budget bound quantifies the resulting constraint on atomic hallucination under perfect atomic recall. Moreover, an additive Gaussian model shows that maintaining high recall for atomic and composite facts incurs false positives even under optimal data mixing. Finally, controlled relational experiments hint how training exposure, rule complexity, and model capacity shape these trade-offs.
Related
- Efficient Robust Learning at the Information-Theoretic Limit
- Dependency, Compression, and Synergy: A Unified Information-Theoretic View of Multimodal Learning
- Fundamental Limits of Information-Theoretic Secure Aggregation with User Dropouts in Federated Learning
- A Joint Information-Theoretic and Graph Learning Framework for Sensor Placement and Macroscopic Traffic State Reconstruction
- Measuring Learned Equivariance: An Information Theoretic Approach
BibTeX
@misc{xu2026learning,
title = {{Learning Rules, Memorizing Facts: An Information-Theoretic View of Hallucination}},
author = {Gengze Xu and Anxin Guo and Yifei Hu and Xiangyu Wu and Yong Liu and Ziqiao Wang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.16233},
eprint = {2610.16233},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.16233},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Learning Rules, Memorizing Facts: An Information-Theoretic View of Hallucination
AU - Gengze Xu
AU - Anxin Guo
AU - Yifei Hu
AU - Xiangyu Wu
AU - Yong Liu
AU - Ziqiao Wang
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2610.16233
UR - https://arxiv.org/abs/2610.16233
AB - Understanding the relationship between generalization and hallucination requires clarifying the role of memorization. We study this relationship from an information-theoretic perspective through a Selector-Density Conditional Mutual Information (CMI) framework. In the sparse random-fact limit, a uniformly bounded information budget per fact forces a nonvanishing false-positive rate at recall bounded away from zero, even when all facts are observed during training. In a rules-and-facts setting, we decompose model-data mutual information into rule learning, atomic-fact memory, and excess memorization, revealing competition between rule learning and factual memory even without excess memorization. A joint budget bound quantifies the resulting constraint on atomic hallucination under perfect atomic recall. Moreover, an additive Gaussian model shows that maintaining high recall for atomic and composite facts incurs false positives even under optimal data mixing. Finally, controlled relational experiments hint how training exposure, rule complexity, and model capacity shape these trade-offs.
ER -