Cube Bench
收藏资源简介:
Cube Bench是由莫纳什大学开发的基于魔方解谜任务的多模态空间推理基准测试集,旨在评估多模态大语言模型(MLLMs)在序列化空间推理中的表现。该数据集通过虚拟魔方模拟器动态生成,包含图像渲染、文本状态描述和候选动作集合三种模态数据,严格遵循确定性生成原则以确保可复现性。其核心设计聚焦于解构感知-评估-行动-反思-恢复的完整决策闭环,通过七项测试任务(如状态重建、最优移动预测、闭环控制等)量化模型在组合复杂度增长时的性能衰减。该数据集主要应用于多模态人工智能领域,为解决静态感知测试与动态交互需求间的评估鸿沟提供了标准化测量工具。
Cube Bench is a multimodal spatial reasoning benchmark developed by Monash University based on Rubik's Cube puzzle-solving tasks. It aims to evaluate the performance of multimodal large language models (MLLMs) in sequential spatial reasoning. This dataset is dynamically generated via a virtual Rubik's Cube simulator, and contains three modal data types: image renderings, textual state descriptions, and candidate action sets. It strictly follows the deterministic generation principle to ensure reproducibility. Its core design focuses on dissecting the complete decision-making closed loop consisting of perception, evaluation, action, reflection, and recovery, and quantifies the performance decay of models as combinatorial complexity increases through seven test tasks, such as state reconstruction, optimal move prediction, closed-loop control, and others. This dataset is primarily applied in the field of multimodal artificial intelligence, providing a standardized measurement tool to bridge the evaluation gap between static perception tests and dynamic interaction requirements.
Cube Bench数据集概述
数据集简介
Cube Bench是一个可复现的魔方基准测试套件,旨在通过魔方这一复杂、结构化的环境,评估大型多模态模型在感知、基础对齐和闭环控制方面的推理与规划能力。
核心评估维度
- 感知:评估模型能否从图像中准确识别魔方的状态。
- 基础对齐:评估模型能否将视觉状态映射到内部表示。
- 闭环控制:评估模型能否生成有效的操作步骤以达到目标状态(例如,解魔方)。
数据集获取与安装
- 获取方式:通过Git克隆仓库获取数据集与框架。
- 环境要求:建议使用Python 3.12环境,可通过Conda或Python venv创建。
- 依赖安装:需预先安装
requirements.txt中指定的依赖包。 - 包安装模式:建议以可编辑的开发模式安装本包。
- 数据预处理:运行基准测试前,需预计算IDA* / 最优距离图。此过程计算密集,耗时可能长达约8小时,具体取决于CPU性能。
- 安装验证:可通过导入
cube_bench包并检查版本来验证安装是否成功。




