Summary
A new paper on arXiv (2608.20331) by Shiao Xie, Siyu Chen, Jianwei Lv, and Bo Yuan introduces the Patient-oriented Medical Report Interpretation task, which requires vision-language models to explain medical reports to patients in accurate yet accessible language, conditioned on user queries and dialogue history. The authors argue that factuality and user satisfaction are fundamentally different but tightly coupled objectives that existing supervised fine-tuning and monolithic reinforcement learning paradigms struggle to jointly optimize. They propose G-CARL, a framework that combines multi-source retrieval for atomic claim verification with context-aware, instance-specific weighted checklists measuring response coverage. This provides structured supervision for factuality, user-need fulfillment, and expression quality without restricting response diversity. The work also contributes MMedReport, a real-world benchmark, and a clinically designed three-dimensional evaluation protocol. Experiments show G-CARL consistently outperforms existing post-training baselines in overall quality, claim-level precision, and checklist recall.
Overview
Field: Computer Vision (CV)
Authors: Shiao Xie, Siyu Chen, Jianwei Lv, Bo Yuan
Published: 2026-08-22
arXiv: 2608.20331
Summary
Personalized interpretation of medical reports is increasingly important for patients, requiring both evidence-grounded medical factuality and context-dependent patient communication. Existing medical vision-language tasks fail to adequately capture these dual demands.
This paper proposes the Patient-oriented Medical Report Interpretation task, which requires a model to explain medical reports to patients in accurate and accessible language, based on the user's query and dialogue history.
Key idea
Factuality and user satisfaction differ fundamentally in verifiability yet are tightly coupled, making them hard to jointly optimize under standard SFT and monolithic RL paradigms. To address this, the authors propose G-CARL, a framework that combines:
- Multi-source retrieval for atomic claim verification (factuality)
- Context-aware, instance-specific weighted checklists measuring response coverage (user-need fulfillment)
This provides structured supervision for factuality, user-need satisfaction, and expression quality without constraining response diversity.
Benchmark and evaluation
The authors construct MMedReport, a real-world benchmark, together with a clinically designed three-dimensional evaluation protocol.
Results
Experiments show that G-CARL consistently outperforms existing post-training baselines in overall quality, claim-level precision, and checklist recall.
---
*Auto-collected on 2026-08-22.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178633810