G-CARL: Grounded Checklist-Aligned Reward Learning for Patient-Oriented Medical Report Interpretation

TL;DR - G-CARL is a reinforcement-learning framework for generating accurate, patient-friendly explanations of medical reports from user queries and dialogue history. It matters because it jointly targets verifiable medical factuality and personalized communication without forcing standardized responses.

• Introduces Patient-oriented Medical Report Interpretation, an open-ended multimodal generation task combining report evidence, user questions, and dialogue context.
• Uses multi-source retrieval to verify atomic medical claims and instance-specific weighted checklists to assess response coverage.
• Provides structured rewards for factuality, user-demand satisfaction, and expression quality while preserving response diversity.
• Introduces the real-world MMedReport benchmark; experiments and clinician preference evaluations show improvements over existing post-training baselines in overall quality, claim precision, and checklist recall.