What's the difference between formal and functional linguistic competence?
This explores the distinction the corpus draws between getting language *right* (grammar, fluency, well-formed sentences) and using language to actually *understand and communicate* (intent, implication, grounding) — and why LLMs seem to have far more of the first than the second.
This explores a split that runs through much of the collection: formal competence is knowing the rules of a language — producing grammatical, fluent, well-structured text — while functional competence is using language to do real things in the world: grasping what someone means, what they left unsaid, and what they actually know. The cleanest statement of the divide is neurological: brain evidence suggests these two capacities run on distinct machinery, and next-token prediction trains the first while never activating the networks that produce the second, because real understanding requires integrating reasoning, social, and world-knowledge circuits beyond the language system itself Are language models developing real functional competence or just formal competence?. So the gap isn't a bug to be patched — it's baked into what the training objective can and can't reach.
What makes this concrete is that the corpus keeps finding the *same shape* of failure across different angles. Models handle explicit, surface structure well — connectives, discourse markers, simple grammar — but break down precisely where meaning has to be inferred: implicit relations, embedded clauses, forward planning Where exactly do LLMs break down with language structure?. Grammatical competence even degrades predictably as sentences get structurally deeper and more recursive, which suggests the models learned surface heuristics rather than genuine structural rules Does LLM grammatical performance decline with structural complexity?. That's formal competence fraying at its own edges. Functional competence barely registers at all: LLMs pattern-match on what's said but can't reason about implicatures, presuppositions, or speaker intent, recognizing ambiguity at 32% where humans hit 90% Why do LLMs fail at understanding what remains unsaid?.
The functional gap shows up most vividly in *communicative work* the models skip. Humans constantly ground conversation — asking clarifying questions, acknowledging, checking understanding — and LLMs produce 77.5% fewer of these acts. Worse, preference optimization actively strips them out, because raters prefer a confident complete answer to a model that pauses to check it understood you Why do language models sound fluent without grounding?. The fluency you hear is partly the *absence* of the very work that signals real understanding. A related move: in academic writing the models nail structural coherence but avoid evaluative stance-taking, favoring description over judgment — formal polish without the functional act of taking a position Why do ChatGPT essays lack evaluative depth despite grammatical strength?.
Here's where the corpus gets interesting and refuses to settle into a tidy verdict. One line of work reframes functional competence as something *acquired over time through participation* rather than possessed innately — as LLMs become established partners in human language games, they pick up elementary social grounding comparable to a young child's, making "do they understand?" a time-indexed question rather than a flat no Can LLMs acquire social grounding through linguistic integration?. But another note draws a hard line inside that optimism: social grounding and genuine *linguistic agency* are distinct, and agency in the enactive sense requires embodiment and stakes — precariousness — that no amount of use can supply Do LLMs gain true linguistic agency through integration?. So the formal/functional split isn't even the end of the story; functional competence itself splits into pieces a model might earn and pieces it categorically can't.
Two more findings round out the picture from unexpected directions. The fact that pre-training models on *formal languages* (think hierarchical grammars) transfers to natural language — cutting token needs by a third and improving syntactic generalization — tells you formal competence really is a separable, transferable skill with its own mechanism Can formal language pretraining make language models more efficient?, What formal languages actually help transformers learn natural language?. And stepping back, the broadest framing reads all of this as an epistemic problem: models track statistical regularities with high fidelity yet show structurally specific failures — hallucination, reasoning collapse, premise-sensitivity — that mark the measurable, unavoidable distance between capturing patterns and actually knowing things What do language models actually know?. Formal vs. functional, in the end, is one name for that distance.
Sources 11 notes
Neuroscience evidence shows next-token prediction produces formal linguistic competence but not functional competence, because functional understanding requires integration of diverse brain networks beyond language circuits that the prediction objective never activates.
LLMs perform well on explicit, consistent structures (causal connectives, discourse markers, simple grammar) but fail where structure must be inferred (implicit relations, embedded clauses, forward planning). This asymmetry reveals they've learned surface statistics without deep structural understanding.
LLMs show systematic performance decline as syntactic depth and embedding increase. Simple sentences are handled well while complex structures with recursion and embedding fail consistently, suggesting LLMs learned surface heuristics rather than structural grammar rules.
Research shows LLMs pattern-match on explicit language but cannot reason about implicatures, presuppositions, or speaker intentions. They fail at scalar implicature adaptation, ambiguity recognition (32% vs 90% human accuracy), and implicit warrant validation in arguments—core features of pragmatic competence.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
Analysis of 145 ChatGPT and 145 student essays revealed LLMs favor manner nouns (method, approach) while avoiding status and evidential nouns (claim, evidence). This systematic preference for description over evaluative stance-taking explains perceived vagueness without invoking vocabulary or grammatical deficits.
Social grounding is acquired through participation in language games rather than possessed innately. As LLMs become established communicative partners in human linguistic practice, they develop elementary social grounding comparable to young children, making the question of LLM understanding time-indexed.
Social grounding and linguistic agency are distinct properties. LLMs acquire more social grounding through integration into language communities, but remain categorically incapable of linguistic agency in the enactive sense, which requires embodiment and precariousness no amount of use can provide.
Pre-pretraining 1B models on hierarchical formal languages achieves equivalent loss and better syntactic generalization using 33% fewer natural language tokens. The mechanism persists: attention heads trained on formal languages remain critical for syntactic performance on natural language.
Transfer from formal to natural language succeeds only when formal languages satisfy two conditions: they capture hierarchical dependencies (Chomsky hierarchy) AND are learnable by transformers with length generalization (circuit complexity). Formal languages meeting both constraints outperform matched natural language training.
LLMs achieve high fidelity in capturing language patterns yet show systematic, structurally specific failures—hallucination, reasoning collapse, and premise-sensitivity. The gap between statistical tracking and real knowledge is measurable and unavoidable.