What they found
The authors built ARIVE, a system that drafts radiology reports in steps. It looks at anatomical regions, reasons about what it sees, and only then writes the final report. To train it, they built 159,416 step-by-step reasoning trajectories that mix images and text, using MIMIC-CXR. They then tuned it further with rewards for anatomical accuracy and clinical content. The abstract says ARIVE beat strong baselines on clinical-efficacy measures on MIMIC-CXR and on MIMIC-ABN, which focuses on abnormal cases. In blinded expert grading its reports were competitive with leading models, and its reasoning trails got high ratings for diagnostic traceability. A paired review test suggests that seeing the reasoning can help reviewers spot errors in reports.
Why it matters
Current SI systems can draft plausible radiology reports, but it is hard to see why they wrote what they did. That makes reports hard to check or fix. Showing the evidence trail behind each finding could make SI drafts easier for doctors to review. The abstract gives no figures for the gains, so the size of the improvement is unclear.
How it changes the race
This is a focused step for medical SI, not a change in frontier capability. It points to a way to make specialist models easier to audit by training them to show their visual evidence before they answer.
Safety and control
Showing the reasoning behind each finding helps human reviewers check and correct SI output in a high-stakes setting. The abstract only says the reasoning "can aid" error spotting. It does not say whether the shown reasoning always matches what the model actually did.
The abstract, by the authors
Automated radiology report generation can draft clinically plausible reports, yet the pathway from visual evidence to reported findings remains opaque, leaving the diagnostic basis of these reports unexplained and the generated reports difficult to review or revise. We introduce ARIVE (Anatomy-guided Reasoning with Interleaved Visual Evidence), an explainable report generation framework that interleaves anatomical inspection with diagnostic reasoning prior to final report composition. To support this process, we constructed 159,416 multimodal chain-of-thought (MCoT) trajectories from MIMIC-CXR, wherein visual evidence from inspected anatomical regions informs subsequent clinical interpretation. Trained on these trajectories and further optimized using anatomical and clinical-content rewards, ARIVE achieved superior clinical-efficacy performance compared with strong baselines on MIMIC-CXR and the abnormality-centered MIMIC-ABN. ARIVE's reports were competitive with those of leading models in blinded expert grading, and its MCoT trajectories earned high ratings for diagnostic traceability. Furthermore, a paired review evaluation suggests that access to MCoT can aid in identifying errors within generated reports.
Quoted from arXiv:2611.14018. The research and its words belong to its authors.
