Linguistic Structure Enrichment Fails to Improve LLM Coherence Assessment

Experiments show that adding syntactic and rhetorical information to text actually decreases coherence prediction accuracy because the metadata is incompatible with current LLM architectures. The study also demonstrates using coherence as a zero-shot proxy for detecting disinformation.