Research map / Large language models / Alignment and preferences
Phrase-grounded APO for Improving Chest X-ray Report Generation
Raziuddin Mahmood, Tanveer Syeda-Mahmood
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 APO alignment at inference (parameter updates on self-generated preference pairs) | seven RRG models · ChestImaGenome Gold | average improvement over original reports: 39.98% average over metrics | Improved |
| add APO alignment on top of static report correction (ablation) | RRG models · ChestImaGenome Gold | average improvement over original reports: 26.52% -> 39.98% | Improved |
| add phrasal grounding loss term to DPO-style loss (ablation) | APO · datasets average (RQ score) | average RQ improvement: 11.4% -> 22.7% | Improved |
| replace APO with GRPO using RQ as reward (ablation) | RRG models · ChestImaGenome Gold | RQ score: overall improvement under 5% | No clear effect |
| apply GRPO alignment to a weaker generator (ablation) | XrayGPT · ChestImaGenome Gold | RQ score: 0.37 -> 0.36 | Hurt |
| apply APO to Maira-2 | MAIRA-2 · ChestImaGenome Gold | RQ score: 0.52 -> 0.61 | Improved |
| apply APO to XrayGPT | XrayGPT · ChestImaGenome Gold | RQ score: 0.37 -> 0.52 | Improved |
| apply APO across datasets (MS-CXR, ChestXray8, VinDr-CXR, ChexPert) | RRG models · multiple datasets | RQ score average improvement: around 22.7% | Improved |
| use DPO instead of APO on MS-CXR (ablation) | RRG models · MS-CXR | RQ average improvement: 18.9% -> 0.22% | Hurt |
| static correction of sentences by FC model and LLM | Llama3.2 correction · 1000 sampled sentences | clinical errors: 94.5% correct, 4.2% clinical errors | Mixed |
Closest papers by meaning
- Radiology Report Generation via Multi-objective Preference Optimization (AAAI 2025 · similarity 0.88)
- Reporting Practice Matters: The Impact of Reference Choice on Chest X-ray Report Evaluation (NeurIPS 2026 · similarity 0.87)
- DART: Disease-aware Image-Text Alignment and Self-correcting Re-alignment for Trustworthy Radiology Report Generation (CVPR 2025 · similarity 0.86)
- S2D-Align: Shallow-to-Deep Auxiliary Learning for Anatomically-Grounded Radiology Report Generation (AAAI 2026 · similarity 0.86)
- Bootstrapping Large Language Models for Radiology Report Generation (AAAI 2024 · similarity 0.86)
- ReFINE: A Reward-Based Framework for Interpretable and Nuanced Evaluation of Radiology Report Generation (AAAI 2026 · 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.
