MedFabric and EtHER: A Data-Centric Framework for Word-Level Fabrication Generation and Detection in Medical LLMs

MedFabric and EtHER introduce a data-centric pipeline generating word-level medical fabrications that preserve syntactic and stylistic fidelity while embedding subtle factual deviations, directly addressing distributional drift and coverage gaps in existing hallucination benchmarks. EtHER's modular detector combines Text2Table Decomposition, Word Masking and Filling, and Hybrid Sentence Pair Evaluation for granular fabrication localization. Clinical NLP researchers and medical AI safety teams should prioritize this work, as word-level detection granularity significantly outperforms sentence-level hallucination classifiers in actionable error attribution.