Can MLLMs Critique Like Humans? Evaluating Open-Ended Aesthetic Reasoning in Multimodal Large Language Models

Photo Critique Dataset, we score five open-weight MLLMs against multiple ranked human critiques per photo with reference-based similarity metrics, under six prompt conditions that disentangle persona framing, aspect hinting, length control, and single- versus multi-pass generation, and add an image-grounding control that feeds each model the wrong photograph. We find that reference-based similarity gives a misleading picture. Stricter lexical and learned metrics show only weak alignment with human critiques, while a coarse embedding cosine reports broad topical overlap that the grounding control traces to a stable house style rather than image-specific observation. Behaviorally, the models diverge from humans in consistent ways the scores do not surface: even under a length cap they write two to three times as much, cover nearly every aesthetic aspect where humans are selective, engage each aspect more uniformly and at greater depth, and repeat themselves across critiques of the same photo where humans vary. We argue that reference-based similarity rewards a fluent, comprehensive critique style rather than the selectivity and specificity of human critique, and discuss implications for evaluating and training open-ended multimodal generation.