A Survey on the Abstraction and Reasoning Corpus

Abstract:

Chollet (2019) proposed a definition of intelligence that emphasizes efficiency in skill acquisition rather than performance on a predefined set of tasks, and introduced the Abstraction and Reasoning Corpus (ARC-v1, or ARC-AGI-1) as a challenge benchmark for machine learning research.
In the following years, ARC and the associated competitions have highlighted fundamental limitations of classical deep learning approaches and underscored the need for new ideas in abstract reasoning. This has incentivized extensive trial-and-error exploration, resulting in a wide variety of methods applied to the corpus.
As ARC-v2 was released in March 2025, this literature survey provides a systematic breadth-first overview of the methods applied to ARC-v1 in the six years since its introduction, prior to version 2, and covers early developments for ARC-v2 and ARC Prize 2025.
We apply a taxonomy distinguishing inductive (which explicitly construct transformation rules) and transductive approaches (which directly map inputs to outputs), examine the ecosystem of enabling techniques and auxiliary datasets, and synthesize patterns, trade-offs, and underexplored areas across the research landscape.
Our goal is to provide newcomers with a comprehensive foundation for understanding existing approaches and identifying promising research directions in abstract reasoning.

Submission Type: Long submission (more than 12 pages of main content)

Changes Since Last Submission: Last update 15 Apr 2026: Adapted to reviewers' feedback; Added clarity on some points. Added references in Sec. 5, Sec. 6 Advantages/Challenges paragraphs. Fixed typos & minutia

Assigned Action Editor: ~Emanuele_Sansone1

Submission Number: 7369