In 2019, the 'Abstract and Reasoning Corpus for Artificial General Intelligence' (ARC-AGI) benchmark was introduced by François Chollet to evaluate intelligence based on skill-acquisition efficiency.

ARC-AGI defines artificial general intelligence (AGI) as a system capable of efficiently acquiring new skills beyond its training data, emphasizing general-purpose abilities over task-specific skills.

The benchmark focuses on fluid intelligence, requiring reasoning and problem-solving capabilities that do not depend on cultural knowledge or previously accumulated skills.

To ensure a fair comparison between AI and human intelligence, ARC-AGI employs core knowledge priors that isolate generalization abilities from domain-specific knowledge.

The design of ARC-AGI-2 is informed by prior human performance data, ensuring that tasks can be solved by humans within two attempts, which creates a level playing field for AI evaluation.

A new focus in ARC-AGI-2 is on efficiency measurement, which assesses not only problem-solving capabilities but also the cost-effectiveness of skill acquisition, marking a significant advancement in evaluating AI intelligence.

ARC-AGI is built on the principle of being 'Easy for Humans, Hard for AI', targeting human-intuitive tasks that challenge AI systems and reveal gaps in their reasoning.

In ARC-AGI-2, the difficulty is increased for AI while remaining accessible to humans, featuring challenging tasks such as symbolic interpretation and contextual rule application.

By 2024, the ARC Prize had developed into a foundation, with competitions leading to growing prize pools, culminating in the launch of ARC-AGI-2 on March 24, 2025.

The timeline of ARC-AGI competitions reflects a rising interest and participation, with the inaugural competition in 2020 achieving a 21% success rate, highlighting the benchmark's challenging nature.