MARBLE
收藏资源简介:
MARBLE是一个具有挑战性的多模态推理基准,旨在仔细审查多模态语言模型(MLLMs)在复杂多模态问题和环境中逐步推理的能力。MARBLE由两个极具挑战性的任务组成,M-Portal和M-Cube,需要利用空间、视觉和物理约束来制定和理解多步计划。
MARBLE is a challenging multimodal reasoning benchmark designed to meticulously assess the ability of multimodal language models (MLLMs) to gradually reason through complex multimodal problems and environments. Comprising two highly challenging tasks, M-Portal and M-Cube, it necessitates the utilization of spatial, visual, and physical constraints to formulate and understand multi-step plans.
MARBLE: 多模态空间推理与规划硬基准数据集
数据集概述
- 名称: MARBLE (Multimodal Reasoning Benchmark)
- 目标: 评估多模态语言模型(MLLMs)在复杂多模态问题和环境中的逐步推理能力
- 特点: 包含两个高难度任务(M-Portal和M-Cube),要求利用空间、视觉和物理约束制定和理解多步计划
- 当前模型表现: 12个先进模型在M-Portal上表现接近随机,在M-Cube上准确率为0%
数据集构成
- M-Portal任务:
- 评估模型计划正确性(plan-correctness)和填空能力(fill-the-blanks)
- 需要下载地图图像: https://huggingface.co/datasets/mrble/MARBLE/resolve/main/images.zip
- M-Cube任务:
- 包含完整任务(cube)、简化任务(cube_easy)和感知子任务(cube_perception)
评估方法
- 评估脚本:
- M-Cube:
eval_cube_api.py或eval_cube_local.py - M-Portal:
eval_portal.py或eval_portal_local.py
- M-Cube:
- 输出结果: 保存在
./output目录
相关资源
- 主页: https://marble-benchmark.github.io
- 论文: https://arxiv.org/abs/2506.22992
- 数据集下载: https://huggingface.co/datasets/mrble/MARBLE
引用信息
bibtex @article{jiang2025marble, title={MARBLE: A Hard Benchmark for Multimodal Spatial Reasoning and Planning}, author={Jiang, Yulun and Chai, Yekun and Brbic, Maria and Moor, Michael}, journal={arXiv preprint arXiv:2506.22992}, year={2025}, url={https://arxiv.org/abs/2506.22992} }
联系方式
- Yulun Jiang: yulun.jiang@epfl.ch




