VisEscape
收藏资源简介:
VisEscape是一个专为评估AI模型在具有挑战性的探索驱动决策制定条件下的表现而设计的虚拟逃脱室基准,包含20个虚拟逃脱室。该数据集通过部分可观测的状态和动态状态更新来模拟现实世界中的推理挑战,要求智能体探索环境、识别并整合相关信息、发现对象的使用方法,并解决复杂谜题以成功逃脱。数据集的创建过程中使用了多种3D建模软件,旨在评估多模态智能体在动态环境中的探索性问题解决能力。
VisEscape is a virtual escape room benchmark specifically designed to assess the performance of AI models in challenging exploration-driven decision-making scenarios, encompassing 20 virtual escape rooms. This dataset simulates real-world reasoning challenges via partially observable states and dynamic state updates, requiring AI Agents to explore the environment, identify and integrate relevant information, discern the proper use of objects, and solve intricate puzzles to achieve successful escape. A range of 3D modeling software was utilized during the dataset's development, aiming to evaluate the exploratory problem-solving capabilities of multimodal AI Agents in dynamic environments.




