Med-StepBench: A Hierarchical Reasoning Framework for Evaluating Hallucinations in Medical Vision-Language Models

2026-05-11

AI Summary

Existing medical vision-language models are prone to hallucinations in multi-step diagnostic reasoning that evade detection by current benchmarks. To address this, this work proposes the first hierarchical evaluation framework tailored for 3D PET/CT imaging, decomposing clinical diagnosis into four expert-designed stages. Leveraging over 12,000 3D scans and a million-scale image-sentence pairs, along with physician-validated annotations, the framework enables fine-grained assessment of both general-purpose and medical-specific models. The benchmark uncovers systematic errors masked by aggregate metrics and reveals model susceptibility to clinically plausible yet adversarial intermediate interpretations when reliable visual evidence is absent, thereby establishing a rigorous foundation for developing safe and trustworthy medical vision-language models.

Abstract

Large vision-language models (VLMs) demonstrate strong performance in medical image understanding, but frequently generate clinically plausible yet incorrect statements, raising significant safety concerns. Existing medical hallucination benchmarks primarily focus on 2D imaging with one-shot diagnostic questions, offering limited insight into whether predictions are grounded in correct localization and abnormality identification, allowing critical reasoning errors to remain hidden behind seemingly correct diagnoses. We introduce Med-StepBench, the first large-scale benchmark for step-wise hallucination detection in 3D oncological PET/CT, comprising over 12,000 images and more than 1,000,000 image-statement pairs across volumetric and multi-view 2D data, which decomposes clinical reasoning into four expert-designed diagnostic stages. Using clinician-verified annotations, we perform the first step-level evaluation of general-purpose and medical VLMs, revealing systematic failure modes obscured by aggregate accuracy metrics. Furthermore, we show that current VLMs are highly susceptible to adversarial yet clinically plausible intermediate explanations, which significantly amplify hallucinations despite contradictory visual evidence. Together, our findings highlight fundamental limitations in grounding multi-step clinical reasoning and establish Med-StepBench as a rigorous benchmark for developing safer and more reliable medical VLMs.

Problem

• medical hallucination
• vision-language models
• clinical reasoning
• PET/CT
• benchmark

Innovation

• step-wise hallucination detection
• medical vision-language models
• 3D PET/CT
• hierarchical clinical reasoning
• adversarial intermediate explanations