MMIR
收藏资源简介:
MMIR是一个用于评估多模态大型语言模型(MLLMs)在检测和推理布局丰富的多模态内容中的不一致性的第一个基准数据集,包含534个具有挑战性的样本,跨越五个推理密集型的不一致性类别:事实矛盾、身份误归因、上下文不匹配、数量差异和时间/空间不连贯。
MMIR is the first benchmark dataset for evaluating Multimodal Large Language Models (MLLMs) on detecting and reasoning about inconsistencies in layout-rich multimodal content. It comprises 534 challenging samples spanning five reasoning-intensive inconsistency categories: factual contradictions, misattributed identities, contextual mismatches, quantity discrepancies, and temporal/spatial incoherences.
MMIR 数据集概述
基本信息
- 数据集名称: MMIR (Multimodal Inconsistency Reasoning)
- 主页: https://jackie-2000.github.io/mmir.github.io/
- 数据集地址: https://huggingface.co/datasets/rippleripple/MMIR
- 论文地址: https://arxiv.org/pdf/2502.16033
- GitHub仓库: https://github.com/eric-ai-lab/MMIR/
数据集简介
MMIR是首个用于评估多模态大语言模型(MLLMs)在布局丰富的多模态内容中检测和推理不一致性的基准测试。数据集包含534个具有挑战性的样本,涵盖以下五个推理密集型不一致类别:
- 事实矛盾 (Factual Contradiction)
- 身份归属错误 (Identity Misattribution)
- 上下文不匹配 (Contextual Mismatch)
- 数量差异 (Quantitative Discrepancy)
- 时空不连贯 (Temporal/Spatial Incoherence)
数据集构建
- 通过四阶段筛选流程精心构建
- 确保高质量、多样化和具有挑战性的测试案例
评估结果 (部分模型表现)
| 模型 | Web (开放式) | Office (开放式) | Poster (开放式) | 总体 (开放式) | Web (选择题) | Office (选择题) | Poster (选择题) | 总体 (选择题) |
|---|---|---|---|---|---|---|---|---|
| o1 (1217) | 47.91 | 59.19 | 38.73 | 51.40 | 47.91 | 58.52 | 46.47 | 52.15 |
| GPT-4o (1120) | 25.00 | 42.60 | 30.98 | 33.14 | 37.29 | 58.96 | 47.88 | 47.75 |
| Qwen2.5-VL-7B | 8.54 | 29.14 | 11.97 | 17.60 | 14.37 | 33.18 | 16.90 | 22.56 |
联系方式
- Qianqi Yan: qyan79@ucsc.edu
- Xin Eric Wang: xwang366@ucsc.edu
引用格式
bibtex @misc{yan2025multimodalinconsistencyreasoningmmir, title={Multimodal Inconsistency Reasoning (MMIR): A New Benchmark for Multimodal Reasoning Models}, author={Qianqi Yan and Yue Fan and Hongquan Li and Shan Jiang and Yang Zhao and Xinze Guan and Ching-Chen Kuo and Xin Eric Wang}, year={2025}, eprint={2502.16033}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2502.16033}, }




