GAMEPLAYQA
收藏资源简介:
GAMEPLAYQA是由南加州大学团队开发的面向3D虚拟智能体决策密集场景的多视频理解评测框架,包含9款多人在线游戏的同步多视角标注视频。数据集以1.22标签/秒的密度标注了智能体状态、动作及环境事件,形成2.4K个诊断性QA对,涵盖基础感知、时序推理和跨视频理解三个认知层级。通过结构化干扰项分类体系,该数据集可精准分析模型在快速决策、多智能体建模和时空 grounding 方面的幻觉现象,为具身AI和世界建模研究提供关键评估工具。
GAMEPLAYQA is a multi-video understanding evaluation framework developed by a team from the University of Southern California, designed for decision-dense scenarios involving 3D virtual intelligent agents. It includes synchronously multi-view annotated videos from 9 multiplayer online games. The dataset annotates agent states, actions and environmental events at a density of 1.22 labels per second, generating 2.4K diagnostic QA pairs that cover three cognitive levels: basic perception, temporal reasoning, and cross-video understanding. Equipped with a structured distractor classification system, this dataset can precisely analyze model hallucinations in rapid decision-making, multi-agent modeling and spatial-temporal grounding, serving as a critical evaluation tool for embodied AI and world modeling research.




