Perceptually Perturbed Judgment Dataset

Multimodal LLMs show bias, researchers develop fix

Researchers have identified a significant bias in multimodal large language models when they are used as judges. These models often prioritize plausible text narratives over perceptually correct visual information, a ph…