@inproceedings{ge-etal-2025-mllm,
    title = "{MLLM}-Bench: Evaluating Multimodal {LLM}s with Per-sample Criteria",
    author = "Ge, Wentao  and
      Chen, Shunian  and
      Chen, Hardy  and
      Chen, Nuo  and
      Chen, Junying  and
      Chen, Zhihong  and
      Xie, Wenya  and
      Yan, Shuo  and
      Zhu, Chenghao  and
      Lin, Ziyue  and
      Song, Dingjie  and
      Wang, Xidong  and
      Gao, Anningzhe  and
      Zhiyi, Zhang  and
      Li, Jianquan  and
      Wan, Xiang  and
      Wang, Benyou",
    editor = "Chiruzzo, Luis  and
      Ritter, Alan  and
      Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/fix-sig-urls/2025.naacl-long.256/",
    pages = "4951--4974",
    ISBN = "979-8-89176-189-6",
    abstract = "Multimodal large language models (MLLMs) have broadened the scope of AI applications. Existing automatic evaluation methodologies for MLLMs are mainly limited in evaluating \textit{objective} queries without considering \textit{real-world} user experiences, inadequately addressing the nuances of creative and associative multimodal tasks. However, the \textit{open-ended} and \textit{subjective} nature of such tasks poses a significant challenge to the evaluation methodology, where it is difficult to define the ground-truth answers for them. To this end, in our paper, we propose a new evaluation paradigm for MLLMs, which is evaluating MLLMs with \textit{per-sample criteria} using potent MLLM as the judge. To validate the feasibility and effectiveness of this paradigm, we design a benchmark, dubbed \textbf{MLLM-Bench}, by curating the evaluation samples across six comprehensive cognitive levels. We benchmark 26 popular MLLMs in a pairwise-comparison fashion, showing diverse performance across models. Moreover, the validity of our benchmark manifests itself in reaching 88.02{\%} agreement with human evaluation. We contend that the proposed paradigm explores the potential of MLLMs as effective evaluation tools with the help of per-sample criteria."
}