Is the Mirage Effect a bug, or is it Geometric Reconstruction in action? A framework for why VLMs perform better "hallucinating" than guessing, and what that may tell us about what's really inside these models

About this article

Last week, a team from Stanford and UCSF (Asadi, O'Sullivan, Fei-Fei Li, Euan Ashley et al.) dropped two companion papers. The first, MARCUS, is an agentic multimodal system for cardiac diagnosis - ECG, echocardiogram, and cardiac MRI, interpreted together by domain-specific expert models coordinated by an orchestrator. It outperforms GPT-5 and Gemini 2.5 Pro by 34-45 percentage points on cardiac imaging tasks. Pretty Impressive! But - the second paper is more intriguing. MIRAGE: The Illusion...