Are language models developing real functional competence or just formal competence?
Neuroscience suggests formal linguistic competence (rules and patterns) and functional competence (real-world understanding) rely on different brain mechanisms. Can next-token prediction alone produce both, or does it leave functional competence behind?
Fedorenko and colleagues (Dissociating language and thought) ground the LLM competence debate in neuroscience. Formal linguistic competence — knowledge of linguistic rules and patterns, grammatical structure, syntactic regularities — relies on dedicated language circuits in the brain. Functional linguistic competence — understanding and using language in the world — requires integration of diverse brain networks beyond language circuits: memory, reasoning, social cognition, sensorimotor systems.
The critical finding: word-in-context prediction, the training objective of most LLMs, produces formal competence as an emergent outcome. It does not and cannot produce functional competence, because functional competence requires the integration of systems that are architecturally distinct in the brain and not activated by the prediction objective.
LLMs are "qualitatively different in their formal linguistic capacities from models before roughly 2018" — a genuine discontinuity in formal competence. But this formal competence arises from an objective that leaves functional competence behind. The two competences are not on a continuum; they are served by different mechanisms.
The predictive implication is architectural. Models that succeed at real-life language use will need to mimic the division of labor between formal and functional competence in the human brain — through modularity: separate circuits for form-level processing and for world-connected functional processing. LLMs that add retrieval, tool use, and memory may be approximating this modularity, but from the outside rather than by design.
This is distinct from Bender & Koller's claim that meaning cannot be acquired from form alone (which rests on the joint-attention/communicative-intent argument). The Fedorenko finding adds a mechanistic neuroscience foundation: even if we grant that some meaning can emerge from distributional learning, the kind of competence that requires world integration is neurologically segregated and cannot be produced by the same mechanism as syntactic pattern-learning.
Source: Philosophy Subjectivity
Related concepts in this collection
-
Can language models learn meaning from text patterns alone?
Explores whether training on form alone—predicting the next word from prior words—could ever give language models access to communicative intent and genuine semantic understanding.
shares the formal/functional gap; Fedorenko adds neuroscience mechanism; Bender/Koller add communicative-intent argument
-
What makes linguistic agency impossible for language models?
From an enactive perspective, does linguistic agency require embodied participation and real stakes that LLMs fundamentally lack? This matters because it challenges whether LLMs can truly engage in language or only generate text.
enactive view is a third account of why functional competence requires more than formal pattern-learning
-
Why does ChatGPT fail at implicit discourse relations?
ChatGPT excels when discourse connectives are present but drops to 24% accuracy without them. What does this gap reveal about how LLMs actually process meaning and logical relationships?
behavioral evidence for formal/functional gap: explicit connectives (formal cues) work; implicit relations (functional understanding) fail
Click a node to walk · click center to open · click Open full network for a force-directed map
Original note title
formal and functional linguistic competence are neurologically distinct — next-token prediction produces formal competence but not functional competence