Tendem
Mini toolsARC AGI 2
New
Text
Gen 7
Overview
Second Abstraction and Reasoning Corpus for AGI, with 1,000 training and 120 evaluation grid tasks plus private test sets, designed to test fluid intelligence and program-synthesis style reasoning.
Description
ARC-AGI-2 is a benchmark dataset for general intelligence composed of colored-grid tasks that require inferring abstract rules from a few examples. It includes 1,000 public training tasks, 120 public evaluation tasks, and additional private test sets, and is intended to evaluate humans and AI systems on skill-acquisition efficiency and generalization rather than narrow task-specific skill.
About ARC Prize Foundation
ARC Prize launched in 2024 as an experiment to promote the relatively obscure, but important, ARC-AGI benchmark. Now, after our OpenAI o3 testing, nearly everyone in tech has heard of it. But there is much more work to do to reach AGI. There remains many things that are easy/fast for humans, but hard/slow for AI.
Website:
arcprize.org
Last updated: January 30, 2026
