CUActSpot
收藏资源简介:
CUActSpot是由微软等机构联合构建的计算机使用代理基准测试数据集,旨在评估模型在复杂图形用户界面交互中的能力。该数据集包含206个精心设计的样本,覆盖GUI、文本、表格、画布和自然图像五种模态,并支持点击、拖拽和绘制等多种交互类型,数据通过人工标注和迭代筛选确保质量。数据集主要用于推动计算机使用代理在真实工作场景中的动作定位研究,解决现有基准测试中交互类型单一、缺乏复杂操作数据的问题,以提升模型在文档编辑、图像处理等任务中的可靠性。
CUActSpot is a benchmark dataset for computer-use agents jointly developed by Microsoft and other institutions, designed to evaluate the capabilities of models in complex graphical user interface (GUI) interactions. This dataset contains 206 meticulously curated samples covering five modalities: GUI, text, table, canvas, and natural image, and supports multiple interaction types including clicking, dragging, and drawing. Its quality is guaranteed through manual annotation and iterative screening. The dataset is primarily used to advance research on action localization for computer-use agents in real-world work scenarios, addressing the issues of single interaction types and lack of complex operation data in existing benchmarks, so as to improve the reliability of models in tasks such as document editing and image processing.

- 1Covering Human Action Space for Computer Use: Data Synthesis and Benchmark东南大学; 穆罕默德·本·扎耶德人工智能大学; 武汉大学; 中山大学; 微软 · 2026年



