RareDxR1 rare-disease diagnosis model accepted to ICME 2026

RareDxR1 is an end-to-end reasoning-centric large language model for open-domain rare disease diagnosis, presented on arXiv as arXiv:2607.00147 and submitted on 30 Jun 2026. The paper, authored by Deyang Jiang, Haoran Wu, Ziyi Wang, Yiming Rong, Yunlong Zhao, Ye Jin and Bo Xu, reports that RareDxR1 operates directly on unstructured clinical notes and achieves state-of-the-art accuracy across different benchmarks. The manuscript runs seven pages, includes three figures, and was accepted to the IEEE International Conference on Multimedia and Expo (ICME) 2026.

What is RareDxR1 and how does it work?

RareDxR1 is an end-to-end large language model trained to diagnose rare diseases directly from unstructured clinical notes, avoiding prior phenotype extraction pipelines. The authors describe a progressive training framework that fuses knowledge internalization with autonomous evolutionary learning so that fragmented rare-disease knowledge is absorbed into the model parameters rather than retrieved at inference. The system also integrates a procedure the paper names Reflection-Enhanced Reasoning Sampling, abbreviated RERS, which the authors say synthesizes expert-level diagnostic trajectories by learning from failures without human annotation.

How does RareDxR1 differ from prior AI approaches?

RareDxR1 departs from pipeline and retrieval-augmented designs by internalizing knowledge and moving beyond closed-set phenotype decision rules. The paper contrasts its approach with typical pipeline-based phenotype extraction and retrieval-augmented generation, which the authors argue cause information loss due to predefined ontologies and retrieval bottlenecks. To bridge generation and expert reasoning, RareDxR1 trains with RERS and a dual-level curriculum reinforcement learning scheme that the authors propose for gradually mastering rare disease diagnosis.

What evidence do the authors present for performance?

The authors state that experimental results show RareDxR1 achieves state-of-the-art accuracy on multiple benchmarks, describing this as a breakthrough in open-domain rare disease diagnosis. The arXiv entry (arXiv:2607.00147) does not publish numeric benchmark scores in its abstract; the claim of state-of-the-art accuracy appears in the paper summary. The manuscript is seven pages long with three figures and the authors indicate they will make their code and dataset publicly available.

Why it matters

Rare disease diagnosis requires extracting subtle phenotypes from messy, unstructured clinical notes and exploring a large diagnostic search space. By training a model to internalize fragmented knowledge and generate diagnostic trajectories without human annotation, RareDxR1 aims to reduce dependence on curated ontologies and external retrieval during inference. If the paper's state-of-the-art claim holds under peer review and public release of code and data, clinicians and researchers could gain a model that reasons about rare disorders directly from notes rather than relying on brittle phenotype pipelines.

What to watch

Watch for the public release of the code and dataset the authors promise and for the ICME 2026 presentation, where more complete experimental details and benchmark numbers should appear. Check the arXiv entry arXiv:2607.00147 and the ICME proceedings for full results and replication material.

References and provenance

• Paper: "RareDxR1: Autonomous Medical Reasoning for Rare Disease Diagnosis Beyond Human Annotation," Deyang Jiang et al., arXiv:2607.00147 (submitted 30 Jun 2026). 7 pages, 3 figures. Accepted to IEEE ICME 2026.
• DOI listed on arXiv: https://doi.org/10.48550/arXiv.2607.00147