Towards Annotation-Free Validation of MLLMs: A Vision-Language Logical Consistency Metric

arXiv:2605.06201v1 Announce Type: new Abstract: Dominant accuracy evaluation might reward unwarranted guessing of Large Language Models, and it might not be applicable to novel tasks for model validation without ground-truth (gt) annotation. Based on basic logic principle, we propose a novel...