Multimodal large language models (MLLMs) (e.g., GPT-4V, LLaVA, and Claude-3) have broadened the scope of AI applications. Yet, evaluating their performance presents a significant challenge owing to the inherently subjective nature of tasks that do not yield clear-cut solutions especially for those open-ended queries. Existing automatic evaluation methodologies are mainly limited in evaluating objective queries without considering real-world user experiences, inadequately addressing the nuances of creative and associative multimodal tasks. 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 \textit{MLLM-Bench}, with the evaluation samples across six critical levels following the revised Bloom's Taxonomy with the ethical consideration. We benchmark 21 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, and that MLLM-Bench will serve as a catalyst for encouraging the development of user-centric MLLMs tailored to real-world applications. Our benchmark data, online leaderboard and submission entry are at https://mllm-bench.llmzoo.com.
View on arXiv