EVADE: Evidence-Verified Agentic Diagnosis with Escape

TL;DR - EVADE is a training-free diagnostic framework that makes a frozen medical vision-language model compare answers from full and self-localized zoomed image views, abstaining when they disagree. It improves calibration and selective safety without sacrificing accuracy.

• On VQA-RAD, SLAKE, and PathVQA with Qwen2.5-VL-7B, EVADE reduced expected calibration error by up to 45% versus zero-shot.
• Cross-view consistency avoids relying on textual self-verification, which can itself hallucinate.
• EVADE was the only evaluated method to improve both calibration and selective risk while maintaining accuracy.
• Self-proposed crops localized diagnostic structures better than center or random crops, but reliability gains came from agreement gating and abstention rather than answer revision.