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

Summary

Demand for personalized interpretation of medical reports for patients is growing, but existing medical vision-language research has not adequately addressed a task that requires both evidence-grounded medical accuracy and communication tailored to context. The authors introduce a new open-ended generation task, called patient-oriented medical report interpretation, that explains reports in accurate and accessible language based on a user's question and dialogue history. Accuracy and communication differ fundamentally in how they can be verified yet remain closely intertwined, making them hard to optimize together through ordinary supervised training or broad reinforcement learning alone. To address this, the authors propose G-CARL, a reinforcement learning framework that combines multi-source retrieval to verify claims at the sentence level with checklists weighted differently for each case. The team also built a benchmark drawn from real-world data and a three-dimensional evaluation protocol designed by clinicians. Experiments showed G-CARL consistently outperformed existing approaches in overall quality, sentence-level accuracy and checklist recall, and clinicians rated its output as more accurate in preference evaluations.