Google DeepMind introduces an AGI cognitive framework and a Kaggle hackathon

What Google DeepMind proposed

Google DeepMind published a March 17, 2026 framework for measuring progress toward AGI. The main argument is that benchmark leaderboards alone are too narrow for AGI discussions, so evaluation should be grounded more explicitly in cognitive science.

The framework identifies 10 canonical cognitive capabilities: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition. Instead of focusing only on absolute benchmark scores, the proposal asks how an AI system compares with the human distribution across those capability areas.

How the evaluation would work

The paper outlines a 3-stage protocol: build a broad task pool with held-out evaluation sets, collect matching human data on the same tasks, and then locate model performance relative to the human range. DeepMind paired the proposal with a Kaggle hackathon rather than leaving it as a purely conceptual paper. The challenge focuses on five capabilities where the authors see the largest evaluation gap: learning, metacognition, attention, executive functions, and social cognition.

DeepMind said submissions open from March 17 through April 16, 2026, with results scheduled for June 1 and a $200,000 prize pool. That makes the announcement noteworthy not just as a research position, but as an attempt to recruit outside researchers into building the measurement layer itself.

Why it matters

The broader implication is that AGI claims may increasingly depend on capability profiles instead of isolated leaderboard wins. If frontier systems are evaluated by which forms of cognition they actually demonstrate, buyers, researchers, and policymakers get a more nuanced picture of strengths and blind spots. The Kaggle component also signals that benchmark governance is becoming a strategic issue, not just a technical one.