Evaluating artificial general intelligence across architectures and task families
Year: 2026
Type: preprint
Abstract: Artificial general intelligence describes the ability to learn, reason, and adapt across a broad range of cognitive tasks. Evidence for that ability depends on what transfers, to which tasks, and under what conditions. This critical narrative review compares symbolic architectures, connectionist learning, reinforcement learning, neuroscience-inspired and embodied systems, universal formal agents, and neurosymbolic hybrids. It relates each research program’s assumptions to the capabilities its evaluations can establish. Formal results on optimization and sample complexity clarify the role of task distributions and hypothesis restrictions. A finite-class calculation illustrates how a sufficient sample bound changes with description length and error tolerance. A second calculation shows how arithmetic and geometric means can reverse the ranking of two hypothetical capability profiles and how changing the task-family weights can reverse it again. These results motivate an evaluation protocol that reports family-level outcomes, human comparators, resource conditions, and uncertainty alongside the aggregate. Published studies demonstrate transfer within defined settings and strong performance in some specialized tasks. Whether comparable performance can be sustained across unfamiliar tasks and changing environments remains open. The protocol and reference architecture identify evidence that would help answer that question.
Author: Gilbert, Chris
Institution: William V. S. Tubman University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234417067
