Multimodal Large Language Models as Image Classifiers
The study investigated how Multimodal Large Language Models (MLLMs) perform as image classifiers under different evaluation protocols like Open-World, Multiple-Choice, and Closed-World. Findings show that performance is heavily dependent on ground truth quality; corrected labels significantly narrow the gap with supervised models. MLLMs are sensitive to annotation noise, suggesting current performance gaps often stem from flawed evaluation rather than inherent model deficiency.