ConceptARC
收藏资源简介:
ConceptARC是由圣达菲研究所创建的一个公开可用的ARC领域基准数据集,旨在系统评估抽象和泛化能力。该数据集围绕16个基本空间和语义概念组织,每个概念包含10个任务,每个任务有三个不同的测试输入,总计480个测试输入。ConceptARC的设计旨在依赖于核心概念的直接实例,并系统评估任务解算器对特定概念的理解程度,以便能够泛化。数据集的应用领域是评估人类和AI系统在抽象推理和概念泛化方面的能力,旨在推动AI系统在概念抽象方面的发展和有效评估。
ConceptARC is a publicly available ARC-domain benchmark dataset created by the Santa Fe Institute, designed to systematically evaluate abstract reasoning and generalization abilities. The dataset is organized around 16 fundamental spatial and semantic concepts, with each concept encompassing 10 tasks, and each task featuring three distinct test inputs, totaling 480 test inputs across the entire dataset. ConceptARC is constructed to leverage direct instantiations of core concepts, and systematically assess the extent to which a task solver understands a given concept to enable effective generalization. Its application domains cover evaluating the abstract reasoning and concept generalization capabilities of both humans and AI systems, with the core objectives of advancing the development of AI systems for conceptual abstraction and enabling the effective assessment of such systems.

- 1The ConceptARC Benchmark: Evaluating Understanding and Generalization in the ARC Domain圣达菲研究所 · 2023年



