RareDxR1: AI Model Diagnoses Rare Diseases from Unstructured Notes

Rare disease diagnosis remains one of medicine’s toughest challenges—physicians must sift through vague, unstructured patient narratives and perform complex reasoning across a vast space of possibilities. Existing AI methods typically rely on pipeline-based phenotype extraction or retrieval-augmented generation (RAG), but these suffer from information loss due to rigid ontologies and retrieval bottlenecks. To overcome this, a team of researchers introduces RareDxR1, a novel large language model designed for end-to-end reasoning from raw clinical text. It internalizes fragmented rare-disease knowledge directly into its parameters, eliminating the need for structured phenotypes or external knowledge bases. The model is trained through a progressive framework that combines knowledge internalization with autonomous evolutionary learning, enabling it to reason like an expert.

Key innovations include Reflection-Enhanced Reasoning Sampling (RERS), which synthesizes expert-level diagnostic trajectories by learning from its own failures without any human annotation, and a dual-level curriculum reinforcement learning strategy that gradually masters the task from simpler cases to complex ones. Experimental results show RareDxR1 achieves state-of-the-art accuracy across multiple benchmarks, demonstrating a significant leap in open-domain rare disease diagnosis. The code and dataset will be made publicly available, and the paper has been accepted to IEEE ICME 2026.