G-CARL grounds patient-facing medical report explanations against retrieved claims and per-instance checklists

arXiv·low signalXBlueskyLinkedInCopy link

An August 20 arXiv paper defines Patient-oriented Medical Report Interpretation as a task and trains G-CARL with reinforcement learning that combines multi-source retrieval for atomic claim verification with context-aware, instance-specific weighted checklists. The pairing is meant to hold factuality and patient relevance at once without collapsing response diversity, and the authors report gains in claim-level precision and checklist recall plus clinician confirmation. They also released MMedReport, a real-world benchmark with a clinician-designed three-dimensional evaluation protocol.