@inproceedings{bose-etal-2026-grail,
title = "{GRAIL}: Grounded Visual Alignment of Medical Vision-Language Models for Radiology Report Generation",
author = "Bose, Sarosij and
Rajendran, Ravi K. and
Debnath, Biplob and
Karydis, Konstantinos and
Roy-Chowdhury, Amit and
Chakradhar, Srimat",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.536/",
pages = "11262--11283",
ISBN = "979-8-89176-462-0",
abstract = "Radiology Report Generation (RRG) aims to reduce the burden of clinical documentation, yet generating reports that remain visually grounded in the input image remains challenging. Current medical vision-language models (Med-VLMs) often rely on curated preference data, retrieval augmentation, or large-scale labeled corpora, but remain prone to hallucinated findings due to weak cross-modal alignment. To address this challenge, we propose $\textbf{GRAIL}$: Grounded Radiology Alignment of Medical Vision-Language Models for Radiology Report Generation, a two-stage multimodal alignment framework for grounded RRG. First, $\textit{Clinically Informed Textual Reasoning}$ improves report structure and medical terminology using verifiable language rewards with periodic clinical guidance. Second, $\textit{Expert-guided Visual Reasoning}$ leverages a domain expert to rank generated reports using image-report alignment scores, encouraging visually grounded report generation without curated preference pairs or retrieval data. Experiments on multiple benchmarks show that $\textbf{GRAIL}$ consistently improves both report generation quality and clinical accuracy, achieving relative gains over strong post-training baselines by up to $\textbf{15}${\%} in clinical accuracy and $\textbf{13}${\%} in generation quality. $\textbf{GRAIL}$ also outperforms proprietary general-purpose and medical VLMs, highlighting the effectiveness of self-supervised multimodal post-alignment for reducing hallucinations in visually grounded radiology report generation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bose-etal-2026-grail">
<titleInfo>
<title>GRAIL: Grounded Visual Alignment of Medical Vision-Language Models for Radiology Report Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sarosij</namePart>
<namePart type="family">Bose</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ravi</namePart>
<namePart type="given">K</namePart>
<namePart type="family">Rajendran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Biplob</namePart>
<namePart type="family">Debnath</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Konstantinos</namePart>
<namePart type="family">Karydis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amit</namePart>
<namePart type="family">Roy-Chowdhury</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Srimat</namePart>
<namePart type="family">Chakradhar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Radiology Report Generation (RRG) aims to reduce the burden of clinical documentation, yet generating reports that remain visually grounded in the input image remains challenging. Current medical vision-language models (Med-VLMs) often rely on curated preference data, retrieval augmentation, or large-scale labeled corpora, but remain prone to hallucinated findings due to weak cross-modal alignment. To address this challenge, we propose GRAIL: Grounded Radiology Alignment of Medical Vision-Language Models for Radiology Report Generation, a two-stage multimodal alignment framework for grounded RRG. First, Clinically Informed Textual Reasoning improves report structure and medical terminology using verifiable language rewards with periodic clinical guidance. Second, Expert-guided Visual Reasoning leverages a domain expert to rank generated reports using image-report alignment scores, encouraging visually grounded report generation without curated preference pairs or retrieval data. Experiments on multiple benchmarks show that GRAIL consistently improves both report generation quality and clinical accuracy, achieving relative gains over strong post-training baselines by up to 15% in clinical accuracy and 13% in generation quality. GRAIL also outperforms proprietary general-purpose and medical VLMs, highlighting the effectiveness of self-supervised multimodal post-alignment for reducing hallucinations in visually grounded radiology report generation.</abstract>
<identifier type="citekey">bose-etal-2026-grail</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.536/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>11262</start>
<end>11283</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GRAIL: Grounded Visual Alignment of Medical Vision-Language Models for Radiology Report Generation
%A Bose, Sarosij
%A Rajendran, Ravi K.
%A Debnath, Biplob
%A Karydis, Konstantinos
%A Roy-Chowdhury, Amit
%A Chakradhar, Srimat
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F bose-etal-2026-grail
%X Radiology Report Generation (RRG) aims to reduce the burden of clinical documentation, yet generating reports that remain visually grounded in the input image remains challenging. Current medical vision-language models (Med-VLMs) often rely on curated preference data, retrieval augmentation, or large-scale labeled corpora, but remain prone to hallucinated findings due to weak cross-modal alignment. To address this challenge, we propose GRAIL: Grounded Radiology Alignment of Medical Vision-Language Models for Radiology Report Generation, a two-stage multimodal alignment framework for grounded RRG. First, Clinically Informed Textual Reasoning improves report structure and medical terminology using verifiable language rewards with periodic clinical guidance. Second, Expert-guided Visual Reasoning leverages a domain expert to rank generated reports using image-report alignment scores, encouraging visually grounded report generation without curated preference pairs or retrieval data. Experiments on multiple benchmarks show that GRAIL consistently improves both report generation quality and clinical accuracy, achieving relative gains over strong post-training baselines by up to 15% in clinical accuracy and 13% in generation quality. GRAIL also outperforms proprietary general-purpose and medical VLMs, highlighting the effectiveness of self-supervised multimodal post-alignment for reducing hallucinations in visually grounded radiology report generation.
%U https://aclanthology.org/2026.findings-emnlp.536/
%P 11262-11283
Markdown (Informal)
[GRAIL: Grounded Visual Alignment of Medical Vision-Language Models for Radiology Report Generation](https://aclanthology.org/2026.findings-emnlp.536/) (Bose et al., Findings 2026)
ACL