Google DeepMind proposed a cognitive framework for measuring progress toward AGI

Google DeepMind proposed a cognitive framework to measure progress toward AGI not by individual benchmarks, but by a set of cognitive abilities. Along with this, the company launched a Kaggle hackathon for researchers to develop practical tests for such an evaluation system.

Why a framework is needed

Google DeepMind starts from a simple problem: virtually all major laboratories talk about AGI, but there is still no generally accepted way to measure how close systems have come to it. Individual tests in mathematics, code, or text generation show only fragments of the picture. If the goal is to understand the overall intelligence level of a model, then measurement should look broader than a single set of tasks or a single successful demo scenario.

In a new work titled Measuring Progress Toward AGI: A Cognitive Taxonomy, the team proposes drawing on cognitive science, psychology, and neuroscience. The logic is this: if AGI is understood as sufficiently general intelligence, then it should be evaluated through basic cognitive functions, not only through applied skills. This is not a ready-made "AGI or not AGI" scale, but a scientific framework that can be applied to specific tests.

"Cognitive science is an important part of the puzzle," write the authors.

What the approach consists of

At the center of the framework are ten abilities that, according to DeepMind's hypothesis, are important for general AI system intelligence. The list includes perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem-solving, and social cognition. This set is important because it covers not only the familiar strengths of modern models, such as text generation or logical chains, but also more complex things—for example, the ability to track one's own errors, flexibly switch between goals, and correctly work with social context.

To turn this taxonomy into a measurable tool, DeepMind proposes a three-step protocol. First, AI systems need to be run through a broad range of cognitive tasks for each ability, using held-out test sets to reduce the risk of data contamination. Then, for the same tasks, a human baseline is collected from a demographically representative sample of adults.

After that, the results of models are compared not against an abstract "passing score," but against the distribution of human results for each ability. The idea here is that comparison with humans should not be a general slogan, but careful empiricism. A model may be very strong in reasoning and memory, but significantly weaker in learning new instructions or in social interpretation.

In this case, the conversation about progress toward AGI becomes more substantive: it shows not only where the system impresses, but also where exactly it has structural gaps.

Kaggle hackathon

DeepMind is not limiting itself to one publication. Together with Kaggle, the company launched the hackathon Measuring progress toward AGI: Cognitive abilities to help the community build the missing evaluations in practice. Participants are invited to use the new Kaggle Community Benchmarks platform and test their ideas on a range of leading models. The focus is on areas where the gap in evaluations is currently most noticeable.

• Learning
• Metacognition
• Attention
• Executive functions
• Social cognition

The prize pool is $200,000. According to DeepMind's rules, the best two works in each of the five tracks will receive $10,000 each, and four more strong submissions regardless of track will receive $25,000 each. Submissions are open from March 17, 2026 to April 16, 2026, and the company promises to announce results on June 1, 2026. This is an important detail: DeepMind is not just publishing the framework as theory, but trying to quickly build an ecosystem of checks and external experiments around it.

What this means

Google DeepMind proposes viewing progress in AI not as a race for individual records, but as a systematic comparison of models' cognitive profiles with the human level. If the approach takes hold, the industry will have a clearer language for talking about "moving toward AGI"—with a breakdown by ability, understandable gaps, and reproducible tests, rather than just loud announcements of the next breakthrough.