AGIへの進捗を測る認知フレームワーク、DeepMindが提案 Measuring progress toward AGI: A cognitive framework

Google DeepMind has put forward a cognitive framework intended to bring more rigor and consistency to how progress toward artificial general intelligence (AGI) is measured. The proposal addresses a persistent problem in the field: the term AGI means very different things to different researchers, companies, and policymakers, making it hard to assess where frontier models actually stand.

Rather than relying on a single benchmark score or a Turing-style pass/fail judgment, DeepMind's framework decomposes human intelligence into ten cognitive domains drawn from decades of work in cognitive science and psychometrics. The dimensions cover areas such as reasoning, knowledge, memory, perception, and language, echoing well-established models of human cognition like the Cattell-Horn-Carroll (CHC) theory that underpins many standardized intelligence tests. For each domain, the framework asks how close current systems come to competent human performance and where significant gaps remain.

Applying this lens, DeepMind argues that today's frontier models perform near human level in some areas — notably broad knowledge recall and fluent language generation — while showing pronounced weaknesses in others, including long-term memory, continual learning, and robust planning. The result is less a single AGI score than a capability profile, making it easier to see which research problems are still open.

The paper builds on DeepMind's earlier 2023 work led by Shane Legg, which proposed Levels of AGI defined along axes of performance and generality. The new framework can be read as a more fine-grained extension, layering a structured cognitive taxonomy onto that earlier classification. It also reflects a broader industry trend: OpenAI is reported to use an internal five-tier definition of AGI progress, while community efforts such as François Chollet's ARC-AGI benchmark push for evaluations that resist memorization and reward genuine generalization.

The framing has implications beyond research planning. Clearer, domain-level descriptions of capability could give regulators, safety teams, and standards bodies a more concrete vocabulary for discussing risks. Capability evaluations are already central to frontier-model safety policies at DeepMind, OpenAI, and Anthropic, and a shared cognitive map could help align those efforts. It may also help defuse some of the marketing-driven ambiguity around AGI claims, by forcing specific assertions about which abilities a system does or does not possess.

There are caveats worth noting. Anchoring AGI evaluation to human cognitive structure is intuitive but not neutral: it may underweight the distinctly non-human strengths of large models, such as massive parallelism, very long context windows, or the ability to operate across many languages and modalities simultaneously. It could also struggle to capture emergent agentic behaviors that do not map cleanly onto any single cognitive domain. DeepMind appears aware of these tensions, and the framework is likely to evolve as empirical evaluations accumulate.

For practitioners, the most immediate value may be methodological. Treating AGI as a multidimensional profile rather than a finish line encourages more honest reporting of model strengths and weaknesses, and makes it harder to overclaim based on a few headline benchmarks. Whether the wider community converges on this particular taxonomy or develops competing variants, the direction of travel — toward structured, cognitively grounded evaluation — looks increasingly hard to avoid.