MedScope-4B Reduces Medical VLM Hallucinations via Counterfactual Training
November 11, 2026
MedScope-4B, a 4B parameter model based on Qwen3.5, uses counterfactual lesion erasure to prevent reporting findings not present in images. It achieved an evidence sensitivity gap of 0.758, whereas frontier models claimed erased lesions in up to 62% of cases.
HOW THIS AFFECTS YOU
●
builderYou can use label-derived trajectories to train smaller VLMs that are more faithful to visual evidence.
●
researcherThe evidence sensitivity gap (ESG) serves as a metric to detect when VLMs rely on prompt bias rather than images.
●
healthThis method helps prevent medical AI from seeing symptoms that are not actually in the scan.