Chapter 07 of an AGI volume outlines how general intelligence should be evaluated beyond single benchmarks

The chapter describes AGI evaluation as a systematic assessment of whether an AI system goes beyond performing isolated tasks to exhibit broad intelligence. Unlike conventional model evaluation based on predefined datasets, it examines adaptation to unfamiliar situations, skill acquisition, cross-domain transfer, and reasoning under uncertainty. The authors argue that a single accuracy score cannot capture general intelligence, so evaluation should treat it as a multidimensional capacity spanning perception, memory, reasoning, learning, planning, and metacognition. The text stresses that generalization must be tested by changing tasks, rules, or environments, and it extends evaluation to long-horizon agent behavior and multimodal integration, with robustness as a further requirement.