Formal Linguistic Competence

• Formal linguistic competence is defined as the internalized, abstract system of rules governing language production and comprehension.
• It underpins generative grammar theory and is evaluated via empirical tasks such as grammaticality judgments and computational perplexity tests.
• Contemporary research explores its neural correlates and the integration of symbolic and neural models to enhance language processing.

Formal linguistic competence is the internalized, abstract capacity to generate, recognize, and judge the well-formedness of linguistic structures—covering the rules and patterns that govern phonology, morphology, syntax, and (in some frameworks) elements of compositional semantics. This notion is foundational in generative linguistics, forms the basis for rigorous formal modeling and benchmarking in both human and artificial systems, and underlies current efforts to link linguistic structure to neural, cognitive, and computational substrates.

1. Theoretical Foundations and Formal Models

Formal linguistic competence, as formalized in generative grammar, is the idealized speaker–listener’s knowledge of the abstract system of rules (grammar) that delimits which utterances are grammatical in a language. The grammar, denoted $G_{\text{nat}}$, provides both weak adequacy (generating all and only the grammatical sentences $L_{\text{nat}}$), and strong adequacy (assigning correct internal structure—e.g., parse trees—to those sentences) (Luque, 2017). In classical formal language theory, this is represented by a 4-tuple grammar

$G = (N,\,\Sigma,\,P,\,S)$

where $N$ is the set of non-terminals, $\Sigma$ the terminal alphabet, $P$ the rules, and $S$ the start symbol. The language generated is $$L(G) = \{ w \in \Sigma^* \mid S \Rightarrow^*_G w \}$$.

The choice of formalism is crucial. Regular grammars or finite automata cannot capture essential natural language phenomena (e.g., center embedding, cross-serial dependencies). Context-free grammars (CFGs) model nested dependencies but fail on cross-serial dependencies found in languages like Swiss German. Context-sensitive grammars (CSLs) capture all, but are too computationally complex for real-time human processing. Moderately context-sensitive formalisms (MCS)—e.g., Tree Adjoining Grammars (TAGs), Linear Indexed Grammars (LIG), Combinatory Categorial Grammars (CCG), and Linear Context-Free Rewriting Systems (LCFRS)—capture the full range of syntactic dependencies needed for natural language while supporting polynomial time parsing (Luque, 2017).

This formal perspective defines the gold standard for human linguistic competence: the ability to recognize and generate all and only the expressions in $L_{\text{nat}}$ and to assign their appropriate structural descriptions in polynomially bounded time.

2. Operationalizations in Human and Machine Contexts

In human linguistic research, formal competence is assessed through native speakers' grammaticality judgments, controlled production/comprehension tasks, and lesion/fMRI studies targeting the left-hemisphere language network (e.g., Broca’s and Wernicke’s areas). These regions respond robustly to manipulations of phonology, morphology, syntax, and compositional semantics (Mahowald et al., 2023, AlKhamissi et al., 3 Mar 2025). Lesion studies reveal dissociations: patients can suffer massive grammatical deficits while sparing non-linguistic reasoning, supporting the theory that formal linguistic competence is neurally and functionally separable from broader cognition (Mahowald et al., 2023).

In LMs, formal competence is typically assessed by evaluating the model’s ability to discriminate grammatical from ungrammatical constructions in minimal pair tests (BLiMP, SyntaxGym), probing the encoding of grammatical categories and structure using classifier-based probes (Holmes), and measuring the perplexity or cross-entropy loss on held-out data (Steuer et al., 2023, Waldis et al., 2024, Zhao et al., 2023). For instance, in BLiMP, a model is presented with (grammatical, ungrammatical) sentence pairs differing in controlled ways; success is registered when the model assigns higher probability to the grammatical variant. Holmes, in contrast, assesses the linear separability of embedded grammatical distinctions in the LM’s representation space across >200 datasets spanning morphology and syntax (Waldis et al., 2024).

3. Dimensions and Empirical Factor Structure

Empirical investigations often decompose formal competence into graded sub-dimensions. For example, in controlled studies of Italian student writing, formal linguistic competence is distinguished into (i) communicative competence—mastery of register and context-appropriate structure—and (ii) grammatical competence—control over syntax, morphosyntax, coherence, and well-formed lexicon. These factors are positively correlated but non-identical, reflecting partially dissociable skill sets (Dallari et al., 29 Jan 2026).

In the Blackbird Language Matrices (BLM) framework, formal linguistic competence encompasses detection and application of rule-governed relations—agreement, argument structure, recursion—across multiple levels of linguistic structure and across sentence aggregates. BLM’s error analysis further reveals systematic error types corresponding to different rule violations, demonstrating fine-grained granularity of formal competence (Merlo et al., 24 Feb 2026).

4. Neurocognitive and Model-Internal Localization

Multiple studies reveal that the acquisition and brain alignment of formal linguistic competence in LLMs and humans are tightly coupled. As LLMs are trained, left-right asymmetries in fMRI predictivity emerge in tandem with the acquisition of formal grammatical skills, but not with world knowledge, arithmetic, or semantic reasoning tasks (Bonnasse-Gahot et al., 13 Feb 2026). Linear readout from LLM activations best predicts neural data specifically in regions implicated in formal linguistic processing, providing a computational-neuroscientific bridge (AlKhamissi et al., 3 Mar 2025, Bonnasse-Gahot et al., 13 Feb 2026).

At the architectural level, a small, sharply localized “core linguistic region” within LLMs—constituting ≈1% of parameters—can be identified via fine-tuning perturbation analysis. Disruption of this core systematically destroys formal linguistic competence (as measured by perplexity), but leaves domain knowledge capacities largely intact, indicating modular separation of grammatical from factual knowledge (Zhao et al., 2023).

5. Benchmarks, Evaluation Paradigms, and Scaling

The modern evaluation landscape for formal linguistic competence relies on large-scale, linguistically controlled benchmarks:

• BLiMP: 67 minimal-pair syntactic/morphological/semantic phenomena, each testing alignment to grammatical structure (Steuer et al., 2023, Renduchintala et al., 20 Apr 2026).
• SyntaxGym: Suite of syntactic structure judgments testing agreement, filler-gap dependencies, and islands (AlKhamissi et al., 3 Mar 2025).
• Holmes: Linear probing of internal representations for grammatical encoding, separating formal (morphology, syntax) from functional (semantics, reasoning, discourse) phenomena (Waldis et al., 2024).
• KoGEM: Korean-specific formal competence across phonology, morphology, syntax, standard and prescriptive rules, augmented to dissociate rote memorization from application of abstract rules (Kim et al., 2 Jun 2025).
• BLM: Structured, multi-sentence tasks designed to probe systematization and rule-encoding rather than superficial co-occurrence (Merlo et al., 24 Feb 2026).

Key empirical results include positive scaling of formal competence with model size up to a sub-linear plateau and sharp improvements in difficult or rare grammatical phenomena when synthetic, targeted data is injected during pre-training (Renduchintala et al., 20 Apr 2026, Waldis et al., 2024). Data composition, task structure, and architectural choices (encoder vs decoder; instruction tuning) jointly affect formal competence, with hybrid or explicitly guided curricula (e.g., L2T) accelerating acquisition (Yamaguchi et al., 6 Jan 2026).

6. Controversies and Alternative Formalizations

The generative tradition defines formal competence through idealized grammars and strong adequacy conditions—requiring explicit, symbolic, deep-structure representations. Critics argue that neural models lacking such explicit structures “merely” mimic statistical frequencies (the so-called “stochastic parrot” hypothesis) and fail to exhibit genuine competence (Borchmann, 14 Oct 2025). However, statistical-empiricist alternatives (Mańczak’s theory) reject the necessity of deep structure, defining competence as mastery of conditional token distributions anchored in frequency of use. In this view, cross-entropy minimization on corpus statistics offers a concrete, operational construct for formal competence, validated by synthesis (i.e., the ability to generate what is attested or expected) (Borchmann, 14 Oct 2025).

This empiricist formalism offers a natural fit for the statistical learning mechanisms of contemporary LLMs, aligning rule acquisition with frequency effects and providing a spectrum (rule vs exception) rather than a dichotomy.

7. Open Problems and Prospects

Open research directions for formal linguistic competence include (a) refining the boundaries of moderately context-sensitive formalisms to match natural language coverage and parsing tractability (Luque, 2017), (b) integrating real cognitive constraints (memory, frequency, attention) into theoretical and practical models (Dallari et al., 29 Jan 2026), (c) expanding annotated corpora and benchmarks for rare or complex phenomena (cross-serial dependencies, long-distance binding) (Renduchintala et al., 20 Apr 2026), and (d) hybridizing symbolic and neural models to combine expressive formal power with efficient, data-driven learning (Yamaguchi et al., 6 Jan 2026, Merlo et al., 24 Feb 2026). The modularity and fragility of core linguistic regions in LLMs, as well as the limited transfer between grammatical and world knowledge skills, argue for further architectural and data-centric innovations (Zhao et al., 2023, Renduchintala et al., 20 Apr 2026).

In sum, formal linguistic competence is both a theoretical construct—anchored in the grammar-driven, rule-governed knowledge underlying natural language—and a practical target for computational and neurocognitive modeling. Its rigorous operationalization, measurement, and mechanistic dissection are essential to both scientific understanding of language and to the development of robust, human-like linguistic systems.