General Reasoning: Benchmarking AI's Fundamental Cognitive Abilities for K-12 Education

General reasoning — encompassing mathematical problem-solving, reading comprehension, multi-step logical inference, and multimodal interpretation — represents the most extensively benchmarked dimension of AI performance in K-12 education. Our analysis covers 109 papers that collectively reveal a field making rapid progress in measuring what AI systems can do, yet largely failing to measure what matters most: whether these systems actually help students learn.

The dominant finding across this body of research is sobering. Even the most capable models available during the study periods — including GPT-4 and its multimodal variants — typically achieved only 50–70% accuracy on K-12 tasks, significantly below human performance benchmarks of 80–95%. More concerning still, several critical studies demonstrate that models frequently rely on shallow heuristics and pattern matching rather than genuine conceptual understanding. The GSM-PLUS benchmark, for instance, revealed 20% accuracy drops when problems were only slightly rephrased — suggesting that what looks like reasoning may often be sophisticated memorisation. This has profound implications for any deployment in educational settings, particularly in low- and middle-income countries (LMICs) where teacher oversight may be limited and the consequences of unreliable AI support are most acute.

The research landscape is heavily weighted towards mathematical reasoning and dominated by English and Chinese language contexts. Multilingual evaluation remains sparse — with only a handful of benchmarks covering languages such as Bangla, Vietnamese, Indonesian, and Korean — and LMIC-focussed research is notably underrepresented. Perhaps most critically, the field has invested heavily in measuring answer correctness while largely neglecting the pedagogical dimensions that determine whether AI tools genuinely support learning: scaffolding quality, cognitive load, metacognitive development, and the risk of cognitive offloading.