multimodal-evaluation

Researchers introduce MM-JudgeBias, a benchmark that exposes systematic compositional biases in multimodal large language models when used as automatic judges, testing 26 SOTA MLLMs across 1,800 samples.