Abstract: Based on Rhetorical Structure Theory(RST), this study evaluate the discourse-level performance of the large language model GPT-4o in Chinese news text simplification. Discourse structures are represented through rhetorical trees and semantic relations. A discourse-annotated dataset of news texts is constructed, and nine quantitative indicators are proposed to describe the properties of text simplification at the discourse level. The comparison between expert simplifications and model outputs shows that GPT-4o produced relatively flat discourse trees. Further analysis revealed that deletion and merging operations by the model often caused excessive simplification, leading to information loss and imbalance in referential coherence. The study proposes the incorporation of discourse structural information into text simplification evaluation and optimization, and highlights its potential for model constraint and self-correction mechanisms.