Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels and has shown strong results in mathematics. We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines. Through a controlled experiment that keeps the questions, model, and optimizer fixed while changing only the answer space, we trace this failure to answer-space structure rather than domain difficulty. In small answer spaces, incorrect rollouts often collide on the same wrong pseudo-label and reinforce it; in large answer spaces, they disperse and receive little reward. This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement. PROSE scores each reasoning step with a medical process reward model, assigns the trajectory reward as the minimum score across steps, and enforces answer-format constraints. Without labels, PROSE substantially improves a general Llama model, surpassing purpose-built medical models and matching much larger systems. Because the process signal is internalized into the policy, the adapted model requires no reward model at inference and transfers its gains to unseen datasets. We further show that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy.
Kai-Long Fan, An-Qi Pu, Yi-Chen Wu et al.· 0 citations
Medical AI has produced many radiology models, particularly for chest X-rays (CXR), each excelling at isolated tasks like lesion detection or report generation. However, these models have disparate capabilities and limited generalizability due to training on restricted datasets, making clinical integration challenging. Large language models (LLMs) now enable interfacing heterogeneous models within agentic frameworks that automatically interpret and unify outputs in natural language. In this work, we present RadFabric, an agentic AI system that orchestrates fourteen specialized open-source CXR analytics models and two Vision-Language Models (VLM) through a modular protocol. RadFabric includes an Anatomical Interpretation Agent that grounds visual findings in anatomical context, and a trainable reasoning agent that synthesizes these anatomically-enriched outputs with VLM-generated radiology reports into transparent, step-by-step diagnoses, even when model outputs are heterogeneous or conflicting. This architecture enables explainable, robust diagnoses across common and rare pathologies while facilitating extensibility through additional agents. Evaluation results on the MIMIC-CXR dataset shows that RadFabric can achieve an AUC of 85.18% on task of detecting different legion types from the given CXR, outperforming all state-of-art CXR models. Notably, the reasoning agent particularly improves detection of uncommon findings, demonstrating enhanced interpretability, generalizability, and clinical applicability.
Wenting Chen, Yi Dong, Zhaojun Ding et al.· npj Digital Medicine· 2 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
