A hybrid PubMed-KG RAG system improves NLI faithfulness of multi-stakeholder clinical-trial summaries by ~0.013 (p<0.0001) across GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash, while Unsupported Claims remains the dominant human-annotated failure mode (mean 1.55/3).
Rousseau, Xin Li, Weijia Xu, Vetle I
1 Pith paper cite this work, alongside 188 external citations. Polarity classification is still indexing.
1
Pith paper citing it
188
external citations · OpenAlex
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Faithful by Design: Evaluating and Improving LLM-Generated Clinical Trial Summaries for Multi-Stakeholder Audiences
A hybrid PubMed-KG RAG system improves NLI faithfulness of multi-stakeholder clinical-trial summaries by ~0.013 (p<0.0001) across GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash, while Unsupported Claims remains the dominant human-annotated failure mode (mean 1.55/3).