NeRD: Neuro-Symbolic Rule Distillation for Efficient Ontology-Grounded Chain-of-Thought in Medical Image Diagnosis

A new framework called Neuro-Symbolic Rule Distillation, or NeRD, aims to make medical image diagnosis more efficient by generating reasoning chains that follow clinical ontologies, according to a paper submitted to arXiv on 14 June 2026 [1][2]. The framework addresses two persistent shortcomings in concept-driven interpretable machine learning. Concept Bottleneck Models require clinicians to score every predefined concept at inference time, a process the authors describe as imposing “a substantial burden on clinicians” [1][2]. Meanwhile, rationale-based generative approaches often select concepts based on class discriminability, which can cause the reasoning to drift away from established diagnostic ontologies [1][2]. NeRD produces reasoning chains that are sufficient for a diagnosis but avoid redundancy, without requiring manually crafted diagnostic rules [1][2]. The system was tested on two skin datasets, where it demonstrated strong diagnostic performance and interpretability [1][2]. A blinded expert evaluation confirmed the clinical plausibility of the rationales produced by the model [1][2]. The work also introduces what the authors call a first expert-in-the-loop study for Multimodal Chain-of-Thought-based diagnosis, achieving concept-level intervention that is both efficient and effective [1][2]. The paper appears in the Computer Vision and Pattern Recognition section of arXiv and is hosted under the arXivLabs framework, which allows community collaborators to develop and share new features on the platform [1]. Neural networks, the broader family of models to which NeRD belongs, consist of layers of artificial neurons that transform input signals through weighted connections [3]. Deep neural networks, defined as networks with at least two hidden layers, have become the backbone of modern computer vision systems, with architectural innovations such as convolutional neural networks driving performance gains in image-related tasks [3]. The push for interpretability in medical applications reflects a wider effort to make these models trustworthy in high-stakes settings [1][2]. The paper’s release coincides with growing infrastructure for sharing machine-learning research artifacts. Platforms such as Hugging Face now allow authors to link papers to models, datasets, and interactive demos, and a collaboration with arXiv embeds those demos directly on abstract pages [5][6]. The NeRD paper’s arXiv page includes links to code and data repositories, though no associated model or demo was confirmed at the time of submission [1].