EgoSocialArena
收藏资源简介:
EgoSocialArena是由浙江大学开发的一个用于评估大型语言模型(LLMs)在第一人称视角下理论思维(ToM)和社会化能力的新框架。该数据集包含两个评估环境:静态环境和交互环境,涵盖七个场景,总计2195条数据。数据集的创建过程包括将现有的第三人称ToM基准转换为第一人称视角,并设计了多个独特的社会情境来测试LLMs的社会化能力。EgoSocialArena旨在解决LLMs在真实社交世界中的表现问题,特别是在理解和推理他人心理状态以及自身心理状态随社交情境变化的能力。
EgoSocialArena is a novel framework developed by Zhejiang University for evaluating large language models (LLMs) on theory of mind (ToM) and socialization capabilities from a first-person perspective. This dataset comprises two evaluation environments: static and interactive ones, spanning seven scenarios with a total of 2195 data instances. The development of EgoSocialArena involves two key components: converting existing third-person ToM benchmarks into first-person perspectives, and creating several distinct social scenarios to assess the socialization capabilities of LLMs. This framework is designed to address the performance limitations of LLMs in real-world social contexts, specifically their abilities to understand and reason about both others' mental states and their own mental states that shift in response to social scenarios.
EgoSocialArena
总体概述
- 数据集名称:EgoSocialArena
- 包含图片:
/image/figure1.png(标题为“Introduction”)
数据转换
- 描述:将第三人称视角转换为第一人称视角
- 包含图片:
/image/figure2.png(标题为“Conversion”)
示例
- 包含图片:
/image/figure3.png(标题为“Example”)




