Jinbae Im

MLLM judges are surprisingly unreliable, exhibiting systematic biases and instability when integrating visual and textual cues, even under semantically irrelevant perturbations.