DeepMind proposes ten-ability cognitive framework to measure AGI progress
A new DeepMind paper replaces binary AGI claims with empirical cognitive profiles, comparing AI systems to human benchmarks across ten distinct abilities including planning, learning, and social interaction.
DeepMind researchers have published a cognitive framework for measuring progress toward AGI that breaks general intelligence into ten distinct abilities drawn from human psychology—eight foundational and two composite. Co-authored by DeepMind co-founder Shane Legg and Matthew Botvinick (now at Anthropic), the paper proposes a three-stage evaluation protocol that produces multi-dimensional "cognitive profiles" for AI systems, including their scaffolding and tool use, rather than single benchmark scores. The framework addresses data contamination, narrow academic task focus, and subjective definitions that plague current AI testing.
The approach shifts evaluation from static academic tests to dynamic integrated systems. Instead of declaring a model "strong AI" based on one score, practitioners assess how well an AI system—including its tools and scaffolding—handles ten key aspects of human cognition such as planning, learning, and social interaction, measured against a representative human sample. This lets teams identify specific behavioral risks and plan deployment based on empirical safety profiles rather than speculative claims. The framework gives industry leaders and regulators a rigorous empirical methodology for tracking real progress and building informed policy. The paper, "Measuring Progress Toward AGI: A Cognitive Framework," appeared on arXiv on June 24, 2026 (2605.28405) with no accompanying code or model release.
