ARC-AGI-2
An upgraded version of the ARC benchmark proposed by François Chollet for measuring the abstract reasoning and generalization capabilities of AI systems. It emphasizes compositional reasoning and reduces patterns that overlap with training data. Human solvers achieve an average accuracy above 85%, while mainstream large language models have consistently scored below 20%.
Timeline (last 90 days)
Gemini 3 Pro scored 31.1% on ARC-AGI-2, Claude Opus 4.6 scored 68.8%, and GPT-5.2 scored 52.9%.
ARC-AGI-2 was released in 2025, the human average score is about 85%, and mainstream models before 2024 were generally below 5%.
ARC-AGI-2, as an upgraded benchmark, raises the difficulty threshold and deliberately circumvents the shortcut of large language models memorizing answers through large-scale pretraining data.
ARC-AGI-2 places greater emphasis on compositional reasoning, requiring the simultaneous application of multiple rules to arrive at an answer, and deliberately reduces patterns that may overlap with training data.
On the ARC-AGI-2 test set, the system solved 230 out of 240 test tasks, with an overall accuracy exceeding 95%.
On the ARC-AGI-2 benchmark, Meta^n is the only method to score above zero.
ARC-AGI-2 is a benchmark specifically designed to resist skill memorization, ensuring that models cannot exploit previously seen patterns by rote learning.
The Chollet team continues to release new versions of ARC-AGI (such as ARC-AGI-2 and ARC-AGI-3) to increase difficulty and diversity.
All Facts (9)
ARC-AGI-2 was designed by AI safety researcher François Chollet; early GPT-4 scored less than 10% on this test, while the human average score is approximately 85%
60%UnverifiedARC-AGI-2 was released in 2025, the human average score is about 85%, and mainstream models before 2024 were generally below 5%.
50%UnverifiedGemini 3 Pro scored 31.1% on ARC-AGI-2, Claude Opus 4.6 scored 68.8%, and GPT-5.2 scored 52.9%.
50%UnverifiedARC-AGI-2, as an upgraded benchmark, raises the difficulty threshold and deliberately circumvents the shortcut of large language models memorizing answers through large-scale pretraining data.
50%UnverifiedARC-AGI-2 places greater emphasis on compositional reasoning, requiring the simultaneous application of multiple rules to arrive at an answer, and deliberately reduces patterns that may overlap with training data.
50%UnverifiedOn the ARC-AGI-2 test set, the system solved 230 out of 240 test tasks, with an overall accuracy exceeding 95%.
50%UnverifiedARC-AGI-2 is a benchmark specifically designed to resist skill memorization, ensuring that models cannot exploit previously seen patterns by rote learning.
50%UnverifiedOn the ARC-AGI-2 benchmark, Meta^n is the only method to score above zero.
50%UnverifiedThe Chollet team continues to release new versions of ARC-AGI (such as ARC-AGI-2 and ARC-AGI-3) to increase difficulty and diversity.
50%