Closing the Abductive Gap: Cognitive Feature Matching to Evaluate Hypotheses in Medical Diagnosis

Diagnosis in realistic, open-ended clinical settings remains a challenge for Large Language Models (LLMs) as they struggle with abductive reasoning -- the process of synthesizing patient observations to formulate the most plausible explanatory hypothesis. In this paper, we show that current approaches such as Chain-of-Thought prompting and general-domain reasoning metrics fail to improve diagnostic accuracy, and identify limitations in their ability to generate, evaluate, and rank hypotheses. To address this gap, we introduce the Abductive Reasoning Quality Score (ARQS), grounded in the psychologically principled Tversky’s feature contrast model, to explicitly capture the cognitive structure of clinical judgement. We design strategies to employ ARQS across single- and multi-turn environments. Experiments on benchmark datasets demonstrate that ARQS-based instructions outperform existing strategies for clinical abduction and process-reward models improving both accuracy and robustness of LLMs. ARQS also serves as a discriminative reward signal for Reinforcement Learning, improving accuracy by up to 26\%. Finally, in multi-turn simulated clinical environments, our ARQS-based protocol achieves superior diagnostic accuracy while reduced cost-of-care by reducing unnecessary medical testing.