Formal linguistic competence is the set of capacities needed to produce and comprehend a language: knowing which strings could be words, how morphemes combine, and how to build well-formed sentences (for example, that English speakers say the dogs in my bedroom are asleep rather than …is asleep, even though the verb is closer to a singular noun). Functional competence is the ability to use language in real-world situations, and it relies on capacities that are not specific to language: formal reasoning, world knowledge, situation modeling, and social cognition (Mahowald et al., 2024).
The distinction rests on neuroscience. In humans, the brain's language network is highly selective: it does not respond during arithmetic, logical reasoning, or mental-state attribution, and people with global aphasia can still play chess and solve arithmetic problems (Mahowald et al., 2024). Applied to language models, it yields an asymmetric diagnosis: their formal competence in English is near human level, their functional competence is patchy and improves less consistently with scale, which is why developers turned to RLHF and external modules. Hence the hypothesis that next-token prediction suffices for the former but not necessarily for the latter.
The framework defuses two mirror-image fallacies: "good at language, therefore good at thought," which already fooled users of ELIZA (see ELIZA effect), and "bad at thought, therefore bad at language," which leads critics to dismiss LLMs as models of language because of their reasoning errors (Mahowald et al., 2024).
Mind the name: the authors note that their "competence" is not the one in Chomsky's distinction between competence (idealized knowledge of a language) and performance (its actual use, with lapses and errors), since separating the two is often hard in models and humans alike (Mahowald et al., 2024, Glossary). Still, the old distinction returns in human–machine comparisons: Firestone (2020) warned that a failure may reflect performance constraints rather than a lack of competence. How far can functional competence learned from text alone go, when so much knowledge of the world travels in text?