Phy-Q

受人类智商计算方式的启发，我们定义了反映智能体物理推理智能的物理推理商（Phy-Q 分数）。我们的评估表明，1) 所有智能体的表现都远低于人类，2) 学习智能体，即使具有良好的局部泛化能力，也难以学习潜在的物理推理规则，无法进行广泛的泛化。我们鼓励开发可以达到人类水平 Phy-Q 分数的智能代理。Inspired by the calculation method of human IQ, we define the Physical Reasoning Quotient (Phy-Q score), which reflects the physical reasoning intelligence of AI agents. Our evaluations demonstrate that: 1) the performance of all AI agents is significantly inferior to that of humans; 2) even learning-based agents with strong local generalization capabilities struggle to learn the underlying physical reasoning rules and thus fail to achieve extensive generalization. We encourage the development of AI agents that can achieve human-level Phy-Q scores.

搜集汇总

背景与挑战

背景概述

Phy-Q数据集受人类智商计算启发，定义了物理推理商（Phy-Q分数）来评估智能体的物理推理智能。评估表明，当前智能体表现远低于人类水平，且难以泛化物理推理规则。该数据集由澳大利亚国立大学于2021年发布，旨在推动达到人类水平Phy-Q分数的智能代理开发。