RPRO: Ranked Preference Reinforcement Optimization for Enhancing Medical QA and Diagnostic Reasoning
RPRO enhances medical question answering by combining reinforcement learning with preference-driven reasoning refinement to improve clinical chain-of-thought performance. The framework uses task-adaptive templates and probabilistic evaluation to automatically select and correct low-quality reasoning chains. Experiments show the 2B model outperforms much larger models on medical QA benchmarks, proving this approach is scalable.