A specialized reasoning large language model for accelerating rare disease diagnosis: a randomized AI physician assistance trial

An open-source reasoning model named RaDaR outperformed larger proprietary systems in diagnosing rare diseases and boosted physician accuracy by more than 21 percentage points in a randomized trial, according to research posted to arXiv. The model, called Rare Disease navigatoR, was built with 32 billion parameters and trained on 49,170 publicly available free-text cases alongside 104,666 synthetic cases using reasoning-enhanced training [1][2]. In benchmark tests and validation across four external centers, RaDaR posted the strongest performance among evaluated open-source models, surpassing the 671-billion-parameter DeepSeek-R1 [1][2]. A retrospective cohort analysis found that RaDaR flagged the final diagnosis before it was documented in clinical notes in 61.06 percent of cases, a gap that corresponded to a potential lead time of 1.87 months and 50.18 percent of the within-center interval [1][2]. In a randomized physician-assistance trial, clinicians using RaDaR achieved a diagnostic accuracy improvement of 21.44 percentage points compared with colleagues who relied on internet search alone [1][2]. The study also reported that phenotype-anchored synthetic narratives provided a useful training signal for long-tail rare diseases, with performance scaling monotonically within the tested data range [1][2]. The paper was submitted to arXiv on June 23, 2026 [1]. arXiv is an open-access repository of electronic preprints that are moderated but not peer-reviewed; it hosts papers across physics, computer science, quantitative biology, and related fields [8]. As of late 2024, the repository was receiving roughly 24,000 new articles per month and had surpassed two million total submissions [8]. The platform also supports arXivLabs, a framework that allows community collaborators to build experimental tools — such as citation explorers and code finders — directly on article pages [6][7]. Rare diseases collectively affect millions of people worldwide, yet diagnosis is frequently delayed by a shortage of clinicians with specialized expertise [1][2]. Large language models have drawn interest as potential diagnostic aids, but earlier efforts have been limited by data scarcity and a lack of clinically grounded validation [1][2]. The RaDaR team framed their work as both a deployable reasoning model and a reproducible development framework for diagnostic AI operating under data constraints [1][2].