There is no accepted AGI benchmark, and the honest reason is that generality is not a single quantity. The field uses a battery of tests, each measuring something narrower, and reads them together.
The main families
- Knowledge and reasoning: MMLU and its successors, which measure exam-style competence across subjects.
- Fluid intelligence: ARC-AGI, which uses novel visual puzzles designed to resist memorisation.
- Agentic work: GAIA, OSWorld and SWE-bench, which ask a system to complete multi-step tasks with tools.
- Broad coverage: BIG-bench, a large collection of tasks contributed by many researchers.
Why scores mislead
Benchmarks leak into training data, reward pattern matching and measure the wrong thing when the goal is transfer. A high score proves the system has seen similar problems, not that it can handle a new one.
What a better test looks like
Unseen tasks, graded by outcome rather than style, with the ability to measure how quickly the system learns rather than what it already knows.