ARC Prize Introduces ARC-AGI-2: A New Era for AI Benchmarking

ARC Prize has unveiled the challenging ARC-AGI-2 benchmark, coinciding with the announcement of its 2025 competition featuring $1 million in prizes. As artificial intelligence evolves from completing specific tasks to exhibiting more general, adaptable intelligence, ARC-AGI-2 aims to identify gaps in capabilities and spur innovation. The ARC Prize team emphasizes that effective AGI benchmarks serve as vital tools for tracking progress, discerning capabilities, and inspiring research. Since its establishment in 2019, the ARC Prize has provided a guiding framework for researchers aiming toward artificial general intelligence (AGI) with benchmarks that endure. Previous benchmarks like ARC-AGI-1 focused on measuring fluid intelligence, emphasizing the adaptability of learning rather than simple memorization. ARC Prize aspires to not only assess advancements but also to ignite new ideas and accelerate scientific breakthroughs. With the launch of OpenAI's o3 in late 2024, a significant shift occurred in AI capabilities, combining large language models with reasoning engines. Despite this progress, such systems still demand considerable human oversight during training, highlighting a necessity for a challenge like ARC-AGI-2, which evaluates true adaptability and efficiency. The ARC-AGI-2 benchmark is designed to be more challenging for AI while remaining accessible to humans. Although advanced AI reasoning systems score poorly on the benchmark, humans typically complete tasks quickly and accurately. The unique design philosophy targets tasks that humans find manageable but are challenging or even impossible for AI, showcasing areas where AI still lacks adaptability—an essential characteristic of human intelligence. Efficiency is also a focal point for the ARC Prize, as measuring performance by the cost of task completion is crucial. Current metrics reveal significant discrepancies between human and AI efficiency in achieving task success, underscoring the ongoing challenges in closing the adaptability gap. Launching this year on Kaggle, ARC Prize 2025 includes various prize categories, aiming to promote innovative approaches to solve the newly established ARC-AGI-2 challenges. Last year’s competition drew 1,500 teams and produced influential research, emphasizing the importance of creativity and exploration in the journey toward realizing AGI. Through this contest, the ARC Prize aspires to catalyze substantial progress in the field of artificial intelligence.