EXPRESS-Bench
收藏资源简介:
EXPRESS-Bench是一个专门设计用于评估探索和推理能力的大型数据集,由中山大学的研究团队创建。该数据集包含777个探索轨迹和2044个问题-轨迹对,旨在支持主动探索并解决现有数据集在模糊性和依赖先验知识方面的关键局限。数据集通过记录代理的视觉观察和动作来生成问题-答案对,并通过人工筛选确保答案的相关性和唯一性。EXPRESS-Bench适用于评估Embodied Question Answering (EQA)任务,助力提升代理的探索效率和回答可靠性。
EXPRESS-Bench is a large-scale dataset specifically designed to evaluate exploration and reasoning capabilities, developed by the research team from Sun Yat-sen University. It contains 777 exploration trajectories and 2044 question-trajectory pairs, aiming to support active exploration and address the key limitations of existing datasets in terms of ambiguity and reliance on prior knowledge. The dataset generates question-answer pairs by recording agents' visual observations and actions, and ensures the relevance and uniqueness of answers through manual screening. EXPRESS-Bench is applicable for evaluating Embodied Question Answering (EQA) tasks, helping to improve agents' exploration efficiency and answer reliability.




