Google DeepMind proposes a cognitive framework for measuring progress toward AGI

Google DeepMind has unveiled a scientific framework meant to help measure progress toward AGI through a set of cognitive abilities rather than individual benchmarks. The company also launched a Kaggle hackathon so researchers can build practical tests for such an evaluation system.

Why a framework is needed

Google DeepMind starts from a simple problem: almost every major lab talks about AGI, but there is still no generally accepted way to measure how close systems are to it. Separate tests in math, code or text generation show only fragments of the picture. If the goal is to understand a model's overall level of intelligence, measurement has to look wider than a single set of tasks or one successful demo scenario.

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

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

What the approach consists of

At the center of the framework are ten abilities that, according to DeepMind's hypothesis, matter for the general intelligence of AI systems. The list includes perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving and social cognition.

The set matters because it covers not only the familiar strengths of current models, such as text generation or chains of logic, but also harder things: for example, the ability to track one's own errors, switch flexibly between goals and handle social context correctly.

To turn this taxonomy into a measurable tool, DeepMind proposes a three-step protocol. First, AI systems are run through a broad range of cognitive tasks for each ability, using held-out test sets to reduce the risk of data contamination. Second, a human baseline is collected for the same tasks from a demographically representative sample of adults. Third, model results are compared not with an abstract "passing score" but with the distribution of human results for each ability.

The idea is that comparison with humans should be careful empirical work, not a general slogan. A model may be very strong at reasoning and memory yet fall noticeably behind at learning from new instructions or at social interpretation. In that case, talk of progress toward AGI becomes more concrete: you can see not only where the system impresses, but where exactly it still has structural gaps.

The Kaggle hackathon

DeepMind is not stopping at a single publication. Together with Kaggle, the company launched the hackathon Measuring progress toward AGI: Cognitive abilities, so the community can help build the missing evaluations in practice. Participants are invited to use Kaggle's new Community Benchmarks platform and test their ideas on a range of frontier models. The focus is on the areas where the evaluation gap is currently most visible.

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

The prize pool is $200,000. Under DeepMind's terms, the top two entries in each of the five tracks will receive $10,000 each, and the four strongest 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 detail matters: DeepMind is not just publishing a framework as theory, but trying to quickly build an ecosystem of checks and outside experiments around it.

What it means

Google DeepMind suggests viewing AI progress not as a race for individual records, but as a systematic comparison of models' cognitive profiles with the human level. If the approach catches on, the industry will gain a clearer language for talking about "approaching AGI", with a breakdown by ability, well-defined gaps and reproducible tests, rather than only loud claims about the next breakthrough.