arXiv 2506.215722025-11-13cs.CL

通过结构方程建模对齐多模态大语言模型基准与人类偏好

Aligning MLLM Benchmark With Human Preferences via Structural Equation Modeling

Interdisciplinary Artificial Intelligence Research Institute, Wuhan College（交叉学科人工智能研究 institute，武汉学院）；School of Computer and Artificial Intelligence, Wuhan University of Technology（计算机与人工智能学院，武汉理工大学）；Shanghai Artificial Intelligence Laboratory（上海人工智能实验室）；Sanya Science and Education Innovation Park, Wuhan University of Technology（三亚科学教育创新园，武汉理工大学）；Institute of Automation, Chinese Academy of Sciences（自动化研究所，中国科学院）；School of Mathematics and Statistics, Northwestern Polytechnical University（数学与统计学院，西北工业大学）；Institute of Artificial Intelligence (TeleAI) of China Telecom（中国电信人工智能研究所（TeleAI））

Shengwu. Xiong, Tianyu. Zou, Cong. Wang, Xuelong Li

备注：: 12 pages, 9 figures

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对现有多模态大语言模型基准的缺陷，本文提出结合结构方程建模与皮亚杰能力层级的框架，构建GOLD基准，实现了更优的可解释性与认知一致性。

摘要：

多模态大语言模型（MLLM）评估长期面临缺乏结构化、可解释且具备理论支撑基准的根本性挑战；现有基于启发式分组的任务存在认知目标模糊、能力重叠、指标冗余、诊断能力弱的问题。为此，本文提出结构方程建模对齐框架，可量化内部效度、维度可分性与组件贡献度，同时引入受皮亚杰理论启发的能力层级，将MLLM能力划分为感知、记忆、推理三层。基于该理论重组现有任务构建GOLD基准，实验表明其相比既往基准具备更优可解释性、更低指标冗余度与更清晰的认知一致性。

英文摘要：

Evaluating multimodal large language models (MLLMs) is fundamentally challenged by the absence of structured, interpretable, and theoretically grounded benchmarks; current heuristically-grouped tasks have vague cognitive targets, overlapping abilities, redundant indicators, and weak diagnostic power. We therefore propose a structural-equation-modeling-aligned framework that quantifies internal validity, dimensional separability, and component contributions, and introduce a Piaget-inspired capability hierarchy that stratifies MLLM abilities into Perception, Memory, and Reasoning. Reorganizing existing tasks under this theory, we build the GOLD benchmark, whose experiments show superior interpretability, lower indicator redundancy, and clearer cognitive consistency than prior benchmarks.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。