arXiv 2026-05-04 preprint Hallucination in large language models (LLMs) is generally defined as the generation of information that does not correspond to facts1. However, output errors observed in actual use are not limited to factual errors. Structural inconsistency, instruction drift, persistence-related errors, and memory-related confabulation are difficult to explain fully through fact-centered definitions alone. This study examines hallucination and related output error patterns through a case-based analysis of actual LLM conversation logs. The cases examined in this study have different surface forms. However, they share a common structure: the output does not sufficiently satisfy a specific user criterion or evaluation criterion. In particular, these errors can be interpreted not as the result of a single cause, but as a combinational structure in which some criteria are weakened or violated while multiple criteria operate at the same time. Based on this view, this study reinterprets hallucination not only as an independent error type, but as a specific case within a broader category of output failure that includes violation of factuality criteria. It also applies the concept of Criterion Violation Output (CVO), proposed in previous work, to actual LLM conversation log cases. CVO refers to an output that fails to satisfy one or more evaluation criteria in a situation where multiple criteria operate at the same time. Through this approach, the study suggests that hallucination, structural errors, instruction drift, and goal misalignment can be interpreted together from the perspective of criterion violation structure. This study proposes that hallucination should be understood not simply as a knowledge error, but as a structural criterion violation that appears in the process of output generation and evaluation. Instead of evaluating LLM output errors only by asking whether the content is factually correct, this approach asks which criteria are satisfied and which criteria are violated. As a case-based exploratory analysis, this study provides a conceptual starting point for future research on quantitative analysis, modeling of interactions among criteria, and controllability of model outputs.
Lee Hochul (Mon,) studied this question.