Google DeepMind Proposes AGI Measurement Framework Built on 10 Cognitive Abilities

Artificial general intelligence (AGI) still lacks a consistent way to measure performance across models and tasks, while existing benchmarks largely emphasize knowledge-based questions or individual reasoning skills. Google DeepMind has therefore proposed a cognitive taxonomy that divides intelligence into 10 abilities, including perception, reasoning, and social cognition. The framework is designed to help research institutions systematically compare models’ progress and shortcomings.

The latest paper establishes a foundation for evaluation based on 10 key cognitive abilities and notes that fields such as learning and metacognition still lack mature tests. Google DeepMind has also partnered with Kaggle to hold a hackathon inviting participants to turn the taxonomy into practical evaluations. Available information does not disclose the event date, prize money, or any other financial details.