Research map / Reinforcement learning / RL from feedback
OraPO: Oracle-educated Reinforcement Learning for Data-efficient and Factual Radiology Report Generation
Zhuoxiao Chen, Hongyang Yu, Ying Xu, Yadan Luo, Long Duong, Yuan-Fang Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FactS reward in place of accuracy reward (ablation) | GRPO · CheXpert Plus | macro F1: 0.089 -> 0.291 | Improved |
| add FactS reward in place of accuracy reward (ablation) | GRPO · CheXpert Plus | macro recall: 0.162 -> 0.605 | Improved |
| add oracle DPO step (OraPO) on top of GRPO+FactS (ablation) | GRPO + FactS · CheXpert Plus | macro F1: 0.291 -> 0.341 | Improved |
| add oracle DPO step (OraPO) on top of GRPO+FactS (ablation) | GRPO + FactS · CheXpert Plus | macro recall: 0.605 -> 0.832 | Improved |
| train OraPO with only 400 samples instead of FactS-only 1K (ablation) | OraPO + FactS · CheXpert Plus | macro F1: 0.291 -> 0.296 | Improved |
| add SFT to GRPO+FactS instead of DPO (ablation) | GRPO + FactS · CheXpert Plus | macro recall: 0.605 -> 0.176 | Hurt |
| add SFT to GRPO+FactS instead of DPO (ablation) | GRPO + FactS · CheXpert Plus | macro F1: 0.291 -> 0.106 | Hurt |
| use accuracy-reward GRPO on base model (ablation) | Qwen2.5-VL-3B · CheXpert Plus | macro F1: 0.034 -> 0.089 | Improved |
| use OraPO+FactS with 1K samples versus MambaXray-L with 1.27M | OraPO+FactS · CheXpert Plus | macro F1: 0.335 -> 0.341 | Improved |
| use OraPO+FactS versus MambaXray-L | OraPO+FactS · CheXpert Plus | macro precision: 0.377 -> 0.237 | Hurt |
| use OraPO+FactS versus best baseline | OraPO+FactS · MIMIC-CXR | macro F1: 0.340 -> 0.357 | Improved |
| use OraPO+FactS versus GPT-5 Thinking | OraPO+FactS · CheXpert gold-label validation set | macro F1: 0.301 -> 0.288 | Hurt |
Closest papers by meaning
- Radiology Report Generation via Multi-objective Preference Optimization (AAAI 2025 · similarity 0.87)
- Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.87)
- ReFINE: A Reward-Based Framework for Interpretable and Nuanced Evaluation of Radiology Report Generation (AAAI 2026 · similarity 0.86)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.85)
- DART: Disease-aware Image-Text Alignment and Self-correcting Re-alignment for Trustworthy Radiology Report Generation (CVPR 2025 · similarity 0.85)
- LLM-RG4: Flexible and Factual Radiology Report Generation Across Diverse Input Contexts (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
