Researchers led by Mochamad Rizal Fauzan posted a paper on arXiv on 2026‑10‑19. They tested five open‑source vision‑language models (2‑5 billion parameters) on 1,000 COCO images while gradually reducing visibility with simulated fog and water. Accuracy fell to chance when visibility was between 12.1 m and 15.9 m in fog, but the models still reported internal confidence scores of 0.68‑0.84 and verbal confidence up to 1.0. Only one model showed confidence that tracked accuracy.
Why it matters
Users of VLMs in low‑visibility settings, like autonomous drones, may be misled by over‑confident outputs.
