AI’s advantage in mathematics may stem from vastly larger working memory, not just reasoning

A new research essay argues that the dominant explanation for AI’s growing prowess in mathematics—superior reasoning—may overlook a simpler factor: access to a vastly larger working memory.

The author, Davide Piffer, suggests that modern language models’ large context windows function like external notebooks, allowing them to preserve entire problem statements, intermediate equations, abandoned approaches, definitions, constraints, and earlier conclusions in explicit form. This capacity, he argues, removes one of the most important biological limits on human reasoning: our restricted working-memory capacity.

Piffer contrasts human working memory—limited to a small number of unfamiliar elements at once—with AI’s ability to maintain dozens or even hundreds of such elements in active context. While humans rely on chunking and external aids like scratch paper to compensate, the scale of AI’s context window is orders of magnitude larger, effectively externalizing much of the reasoning process.

The essay grounds this argument in cognitive science, citing studies that link working memory to mathematical performance beyond general intelligence. For example, Alloway and Passolunghi (2011) found that working-memory measures contributed uniquely to mathematical performance even after accounting for verbal ability, while Alloway and Alloway (2010) showed early working-memory performance predicted later numeracy and literacy better than IQ in a six-year longitudinal study.

A meta-analysis by Friso-van den Bos and colleagues (2013) similarly found a consistent relationship between working memory and mathematics across primary-school studies, though the strength varied by task type. The author emphasizes that these findings do not imply working-memory training will reliably boost intelligence or math skills, but they do suggest a biological bottleneck in human reasoning that AI systems bypass through sheer memory capacity.

Piffer also clarifies that context windows are not identical to human working memory. While humans can silently maintain and transform internal states, standard language models rely more on the explicit sequence of tokens in their context. This makes their reasoning more externalized—akin to using scratch paper at scale—where the text itself is part of the mechanism of reasoning rather than just a record of it.

The essay concludes that AI’s advantage in mathematics may be partly illusory, reflecting the removal of a constraint that suppresses human performance rather than a fundamental leap in reasoning ability.