An arXiv preprint reports that an AI system called STRIVE scored substantially higher than existing methods at matching changes between a current chest X-ray study and the patient’s most recent prior study. On the paper’s Longitudinal Change Concordance measures, which assess agreement on reported changes over time, STRIVE recorded scores of 0.394 on the coarse measure and 0.283 on the fine-grained measure. The strongest baseline scores were 0.193 and 0.128, respectively, putting STRIVE at more than twice those levels.
Longitudinal reporting asks a system to capture what is present now and how it differs from an earlier study. STRIVE splits that task among specialized agents for current finding presence, attributes and change. It then applies a Consistency Gate before report generation and a Validation Agent after generation.
All three clinical decision agents first used supervised fine-tuning. The Temporal Change Agent was then refined with Progression-Aware GRPO, a training stage focused on temporal changes, and checkpoints were selected according to validation performance.
Evaluation used the official Longitudinal-MIMIC test partition. Each example paired a current study with the patient’s most recent prior study and report; the partition contained 2,058 studies and used released reference reports. Clinical-efficacy scoring used CheXbert labels from generated and reference reports, binarized into present-versus-other states and summarized with micro-averaged precision, recall and F1.
STRIVE led on several other measures
The paper’s other automated scores also favored STRIVE. It ranked first on every natural-language-generation measure except ROUGE-L, where it ranked second, and first on all three clinical-efficacy measures. Its clinical-efficacy recall was 0.665, compared with precision of 0.581.
On seven ReXrank measures, STRIVE ranked first on six and second on BERTScore.
The biggest difference was in change coverage
The analysis also reported how often reference-stated changes were omitted by generated reports. Baseline omission rates ranged from 0.71 to 0.87, whereas STRIVE’s rate was 0.46. The reported difference was an absolute reduction of 0.25 and a relative reduction of 35%. In the version without the Temporal Change Agent, the omission rate was 0.83.
STRIVE’s advantage was reported across all five fine-grained change labels. Of 3,275 reference changes, 2,219 were labeled stable. There were 87 resolved statements, and nine of ten baselines recovered none of those statements correctly.
Component tests were more mixed
In the reported component ablation, the version without the Temporal Change Agent was associated with an LCC-C score of 0.148 rather than 0.394 for the full version, and an LCC-F score of 0.095 rather than 0.283. Clinical-efficacy F1 remained essentially unchanged. The comparison describes an association in the reported ablation and does not by itself establish a causal effect.
The other ablations showed the following reported scores. Without progression-aware GRPO, LCC-F was 0.259. Without Validation, clinical-efficacy F1 was 0.615, LCC-C was 0.338 and LCC-F was 0.233. Without the Consistency Gate, the corresponding values were 0.620, 0.389 and 0.274.
A stage-by-stage trace recorded LCC-C and LCC-F scores of 0.342 and 0.241 from raw Temporal-Agent outputs. The scores were 0.354 and 0.250 after the Consistency Gate, 0.338 and 0.233 after report writing, and 0.394 and 0.283 after validation.
What the benchmark does not establish
To examine uncertainty in the comparison, the analysis used a paired nonparametric bootstrap that resampled the 2,058 test study pairs 2,000 times to calculate 95% confidence intervals. STRIVE’s intervals lay above all ten baselines. Against MedRAX, the reported margin was 0.201 for LCC-C, with a 95% confidence interval from 0.174 to 0.229, and 0.155 for LCC-F, with an interval from 0.126 to 0.183.
The evidence is limited to automated comparisons of generated reports with reference reports in the Longitudinal-MIMIC test set. The supplied analysis does not report external-dataset validation, radiologist adjudication, prospective workflow testing or patient outcomes. These results therefore describe performance under the reported benchmark conditions, rather than clinical usefulness or readiness for deployment.
The supplied document is an arXiv v1 preprint dated 25 Aug 2026. Its front matter lists author affiliations but provides no funding statement.
Paper data and sources
Original title: STRIVE: Multi-Agent Structured Temporal Reasoning with Integrated Verification for Longitudinal Radiology Report Generation
Authors: Junyeong Maeng, Eunsong Kang, Heung-Il Suk
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text