Language Dominance and Proficiency

• Language dominance and proficiency are measures of a speaker's relative control over languages, impacting syntactic transfer and priming effects.
• Controlled prime–target experiments reveal that moderate L1 dominance maximizes priming magnitudes, while extreme dominance can diminish L2 proficiency.
• Both neural models and human studies show that shared abstract grammatical representations link language proficiency with effective bilingual processing.

Cross-linguistic priming is the phenomenon whereby prior exposure to a specific syntactic or grammatical structure in one language increases the likelihood that a bilingual individual, or a multilingual LLM, will subsequently produce or process an analogous structure in another language. This effect, originally established in psycholinguistics as evidence for abstract, language-independent grammatical representations, has become a central tool for probing shared syntax and cross-linguistic influence (CLI) in both human bilinguals and contemporary LLMs. Recent computational research has adopted cross-linguistic priming as a causal method to demonstrate the existence and dynamics of shared grammatical subspaces in trained neural architectures (Arnett et al., 5 Mar 2025, Issam et al., 29 Jan 2026, Michaelov et al., 2023, Arnett et al., 2023, Zhang et al., 2024).

1. Foundations and Psycholinguistic Significance

Cross-linguistic structural priming provides a robust paradigm for investigating abstract grammatical knowledge beyond surface-level lexical overlap. In human experiments, exposure to a syntactic frame (the “prime”) in language $\ell_1$ increases the probability of producing a structurally analogous sentence (the “target”) in language $\ell_2$, independent of semantic or lexical repetition. This effect extends robustly from monolingual to cross-linguistic contexts and has been observed in Dutch–English, Spanish–English, German–English, Greek–English, and Mandarin–English bilinguals (Michaelov et al., 2023).

The theoretical import of cross-linguistic priming lies in its capacity to demonstrate shared syntactic representations: when a structure in L₁ primes the same structure in L₂, it supports the existence of language-agnostic (“abstract”) grammatical subspaces. This abstractness rules out explanations based on associative learning or simple translation, instead implicating deeper representational commonalities—a hallmark of the “shared syntax hypothesis” in bilingual cognition (Arnett et al., 5 Mar 2025, Michaelov et al., 2023).

2. Experimental Paradigms and Methodologies

In both human and computational research, the measurement of cross-linguistic priming proceeds via controlled prime–target experiments. For LLMs, the standard protocol includes:

• Model Setup: Training bilingual (or multilingual) Transformer models (e.g., GPT-2–style or XGLM) under carefully manipulated exposure regimes—either simultaneous or sequential L1/L2 introduction, with tightly controlled training data amounts and language order (Arnett et al., 5 Mar 2025, Arnett et al., 2023).
• Stimuli: Using minimal prime–target pairs exemplifying canonical grammatical alternations (e.g., double-object vs. prepositional-object datives, active vs. passive voice, s-genitive vs. of-genitive).
• Priming Metric: For a prime sentence $P$ in language $\ell_1$ and two contrastive target sentences $T_{S_1}$, $T_{S_2}$ in language $\ell_2$, models compute:

$$P(T_{S_1}\mid P) = \exp\left(-\sum_{t=1}^{|T_{S_1}|}\log P(w_t^{S_1} \mid w_{<t}^{S_1}, P)\right)$$

The normalized target probability is then

$$P_N(T_{S_1}\mid P) = \frac{P(T_{S_1}\mid P)}{P(T_{S_1}\mid P) + P(T_{S_2}\mid P)}$$

Priming magnitude for structure $S_1$ is

$$\Delta_{\text{priming}} = P_N(T_{S_1}\mid P_{\text{match}}) - P_N(T_{S_1}\mid P_{\text{mismatch}})$$

• Significance Testing: Linear mixed-effects models are used with fixed effects for prime type and random intercepts for items; FDR correction is applied for multiple comparisons (Arnett et al., 5 Mar 2025, Michaelov et al., 2023).

This experimental machinery allows for direct analogies to human structural priming paradigms and supports causal inferences about abstract syntax in neural models.

3. Empirical Results and Quantitative Patterns

Cross-linguistic priming is reliably observed in both human subjects and neural LLMs, but effect sizes and patterns depend systematically on linguistic and experimental variables:

• Magnitude and Directionality: When English is the target language, both simultaneous and sequentially trained bilingual models show significant cross-linguistic priming for all tested grammatical alternations, with normalized probability increases of 10–20 percentage points ($\Delta_{\text{priming}} \approx 0.15$) for typologically close L1s (Dutch, Spanish). For typologically distant languages (Greek, Polish), effect sizes drop to $<0.07$ and may vanish after catastrophic forgetting in sequential setups (Arnett et al., 5 Mar 2025).
• Asymmetries: Priming from L1 to L2 is robust and significant, while the L2 to L1 effect is weaker or absent—particularly when the target is morphologically rich or has free word order. This matches human findings that L1→L2 priming is typically stronger and that directionality cannot be reduced to order of language acquisition alone. Structural salience in the target language is critical (Arnett et al., 5 Mar 2025, Michaelov et al., 2023).
• Model Class Effects: Transformers consistently outperform RNNs (GRUs) in cross-linguistic priming accuracy (e.g., 33.33% vs. 25.84%), and yield higher normalized conditional probabilities for primed structures. Cue-based retrieval via self-attention (Transformer) more faithfully captures abstract syntax than strictly sequential residual-activation mechanisms (RNN) (Zhang et al., 2024).
• Comparison with Human Data: Model priming magnitudes are on the same order as human priming in many L₁→L₂ configurations (e.g., Dutch→English $\Delta P_{\text{mod}} = 0.12$, $\Delta P_{\text{human}} = 0.11$), although absolute values are lower and some language pairs fail to reach significance, mimicking human asymmetries (Michaelov et al., 2023).

4. Mechanisms and Representational Evidence

Several mechanistic analyses corroborate behavioral priming outcomes with direct evidence from the internal representations of LLMs:

• Hidden-State Diagnostics: The ratio of L1- to L2-typical tokens in logit predictions rises with later L2 introduction and increases further following an L1 prime, especially for models with highly entrenched L1s. This suggests graded co-activation of L1 representations in L2 processing (Issam et al., 29 Jan 2026).
• Neuron Overlap: The intersection of L2-selective neurons across models pre-trained on different L1s is inversely correlated (r ≈ –0.85, $p<0.01$) with syntactic distance. Thus, typological similarity leads to more shared neural circuitry for grammatical structures, directly linking population-level priming to representational overlap (Issam et al., 29 Jan 2026).
• Emergence During Training: Cross-linguistic priming effects emerge rapidly—within less than 1 million L2 tokens—after L2 exposure in sequentially trained bilingual models, well before the L2 cross-entropy plateaus (Arnett et al., 2023). This is attributable to very sparse data contamination, indicating that only minimal L2 input suffices for shared structure induction.

5. Modulators of Priming: Syntactic Distance, Dominance, and Orthography

The strength, direction, and facilitation/interference effects of cross-linguistic priming are systematically governed by:

• Syntactic Distance: Closely related L1s (e.g., German or Spanish for L2 English) yield positive transfer (facilitation, negative CLI), while distant L1s (e.g., Korean, Turkish) introduce interference (positive CLI) after L1 priming (Issam et al., 29 Jan 2026). The number of grammatical phenomena with significant positive priming declines with increasing L1–L2 syntactic distance.
• Language Dominance and Proficiency: Age of L2 exposure (early/late) modulates both priming strength and directionality. Moderate L1 dominance (intermediate age of exposure) yields maximal priming magnitude, but as L1 dominance increases further, L2 proficiency declines and priming effects attenuate. This recapitulates human bilingualism findings concerning critical periods and proficiency windows for transfer (Issam et al., 29 Jan 2026).
• Orthography and Word Order: The presence of shared scripts amplifies cross-linguistic priming; romanizing previously non-Latin scripts (Greek, Korean) restores trends otherwise suppressed by script dissimilarities. Word-order similarity is not strictly necessary but enhances facilitation (Issam et al., 29 Jan 2026).

6. Theoretical Implications and Applications

Cross-linguistic priming in artificial neural learners has directly informed theoretical debates in psycholinguistics and computational linguistics:

• Shared Syntax Hypothesis: Behavioral and representational findings support a “shared” rather than “separate-but-connected” view of bilingual grammar for overlapping, grammatical structures, with partial co-activation occurring even for typologically distant languages (Arnett et al., 5 Mar 2025, Michaelov et al., 2023, Issam et al., 29 Jan 2026).
• Limits of Transfer: English’s fixed word-order and morphosyntactic transparency facilitate more robust cross-linguistic priming, while languages marked by rich morphology and flexible word order (Greek, Polish, Mandarin) show weaker or no transfer. This underscores structural salience over acquisition order as the key factor in syntactic adaptation (Arnett et al., 5 Mar 2025, Michaelov et al., 2023, Zhang et al., 2024).
• Applications as Causal Probes: Structural priming tasks create language-agnostic, task-independent causal probes for abstract syntax in LMs, bypassing the limitations of black-box probing classifiers. This approach has been used to validate and compare pre-training regimens, data contamination, low-resource adaptation, and representational emergence (Arnett et al., 2023).

7. Open Challenges and Future Directions

Emerging directions at the intersection of artificial and human bilingualism include:

• Layer-wise and Neuronal Localisation: Future work aims to identify which Transformer layers or specific neuron subpopulations encode abstract syntactic frames, going beyond output probabilities to enrich theories of shared grammatical representations (Michaelov et al., 2023).
• Extension to Low-Resource Families: The transferability and limits of cross-linguistic priming in low-resource or typologically divergent language pairs remain a significant challenge, with implications for both cognitive modeling and practical NLP deployment (Arnett et al., 5 Mar 2025, Arnett et al., 2023).
• Integration with Human Data: Direct comparison of model-based surprisal effects to human reading times, as well as intervention on model internal states, can provide mechanistic explanations for observed asymmetries and transfer limitations (Zhang et al., 2024, Michaelov et al., 2023).
• Interlanguage and Grammaticality: Priming of ungrammatical structures is sensitive to L1 dominance, suggesting a domain where “separate-but-connected” representations remain relevant—mirroring persistent CLI in human language learners (Issam et al., 29 Jan 2026).

In summary, cross-linguistic priming serves as a mechanistic window into the emergence, structure, and limits of shared grammatical knowledge in both biological and artificial multilingual learners, tying together behavioral, computational, and neuro-representational evidence (Arnett et al., 5 Mar 2025, Michaelov et al., 2023, Arnett et al., 2023, Issam et al., 29 Jan 2026, Zhang et al., 2024).