Cross-Linguistic Influence (CLI)

• Cross-Linguistic Influence (CLI) is the systematic effect of one language on another, impacting phonology, syntax, and lexical choice in both human and artificial systems.
• Computational approaches model CLI via sequential language training and simulated bilingual transformers, quantifying transfer through metrics like accuracy shifts and activation ratios.
• Empirical studies demonstrate that typological similarity and prompt engineering significantly enhance transfer efficiency in multilingual NLP and second language acquisition.

Cross-Linguistic Influence (CLI) captures the systematic impact of one language on another in the mind of a bilingual or multilingual speaker—and, more recently, in multilingual neural LLMs. In the computational linguistics literature, CLI encompasses facilitative and interfering effects that arise due to typological overlap, language dominance, proficiency distribution, and structural similarity between languages. Empirical and mechanistic studies have shown that CLI underlies both transfer of linguistic structure in second language acquisition and the transfer dynamics of cross-lingual LLMs.

1. Definition and Typology of Cross-Linguistic Influence

CLI is defined as the influence that the knowledge, usage, or structure of one language (typically the native or dominant language, L1) exerts on another acquired language (L2), affecting representation, processing, or production in bilingual individuals and artificial learners. CLI is formally characterized along two axes:

• Directionality: L1→L2 (native-to-non-native) or L2→L1, including bidirectional patterns.
• Polarity: Positive transfer (facilitation) when similarities ease acquisition or performance, versus negative transfer (interference) when mismatches or entrenched L1 structures impede L2 processing or production.

In both human and artificial bilingual systems, CLI can manifest in morphological, syntactic, phonological, and lexical domains. Models of CLI are often deeply informed by typological similarity, revealing systematic alignments between error patterns or transfer efficiency and the structural distance between language pairs (Berzak et al., 2014, Yadavalli et al., 2023, Dolicki et al., 2021).

2. Mechanistic and Computational Modeling of CLI

Contemporary research operationalizes CLI within controlled language modeling environments, yielding precise quantification and mechanistic insight beyond human behavioral studies. Key approaches include:

• Sequential LLM Training: Models such as SLABERT employ a two-stage transfer protocol: L1 pretraining on child-directed speech, and L2 embedding fine-tuning, freezing all but the token embedding layer. This setup isolates inductive bias transfer and enables per-phenomenon assessment of positive (Δ_f(L1) > 0) and negative transfer (Δ_f(L1) < 0) on minimal-pair grammaticality tasks (Yadavalli et al., 2023).
• Simulated Bilingualism in Transformers: GPT-2–style models are exposed to varying L1 dominance, L2 proficiency, and L1-L2 typological distance. CLI is measured in terms of accuracy shifts and co-activation metrics during BLiMP minimal-pair evaluations under L1-primed and unprimed conditions. Mechanistic evidence links CLI to activation ratios and recruitment of language-specific neural circuitry, with overlap measures negatively correlated with typological distance (Issam et al., 29 Jan 2026).
• ROSE Neural Oscillatory Framework: In the neurocomputational ROSE model, CLI in bilingual sentence production is attributed to competition between oscillatory attractor subspaces corresponding to L1 and L2 templates. Failure modes for CLI-induced transfer errors arise as (i) subspace competition (β-commit failures) or (ii) θ–γ phase misalignment during morphosyntactic sequencing, both supporting formalization of inhibition, co-activation, and transfer at the dynamical systems level (Uluslu et al., 26 Jan 2026).

3. Linguistic Determinants of CLI in NLP and Human Contexts

Empirical work investigating multilingual transformers and cross-lingual in-context learning has identified critical predictors of CLI and transfer strength:

• Structural and Typological Similarity: Transfer efficiency follows fine-grained syntactic and morpho-syntactic features (e.g., word order, relative clause attachment, pronoun politeness distinctions). Task-specific predictors outperform aggregate genetic, geographical, or syntactic distances for predicting zero-shot/few-shot transfer quality (Dolicki et al., 2021). Table 1, adapted from (Dolicki et al., 2021), highlights the strongest features for POS/NER/NLI tasks.

Task	Top Transfer Feature (WALS)	Impact Direction
POS tagging	96A (Order: Object-Verb / RelCl-Noun)	Shared order → ↑ accuracy
NER	89A (Numeral–Noun order)	Shared pattern → ↑ accuracy
NLI	45A (Pronoun Politeness)	Shared absence → ↑ accuracy

• Multilingual Prompt Design: In few-shot in-context learning, instructing LLMs with demonstrations from multiple high-resource languages (HRLs) consistently outperforms English-only or monolingual-HRL prompts for low-resource languages (LRLs) (Δgain(LRL) ≈ 15.6% above English; χ²≈33.2, p<1e-8). Exposure to mixed-script, non-English exemplars—even with context-irrelevant content—can further activate latent multilingual knowledge (Tu et al., 17 Feb 2025).
• Data-Derived and Phonological Factors: In neural text-to-text transfer (mT5), CLI emerges as a function of linguistic (syntactic, morphological, phonological) and data-based similarities. Meta-regression reveals that phonological and syntactic overlap are strong facilitators of zero- and few-shot transfer, while high morphological divergence can hinder performance (Muller et al., 2022).

4. Quantitative Analysis and Predictive Modeling of CLI

CLI intensity and transfer dynamics are now routinely modeled with formal statistical tools:

• Meta-Regression for Zero- and Few-Shot Performance: For multilingual models, the target language's task performance is regressed on a combination of source language performance and a suite of similarity features (syntactic, phonological, morphological, language-model pretraining score). The general model:

$$S_{𝒯}(0) = \beta_0 + \beta_1 S_{𝒮} + \beta_2\, \mathrm{Synt}_{(𝒯,𝒮)} + \beta_3\, \mathrm{Phono}_{(𝒯,𝒮)} + \cdots$$

• Few-Shot “Law”: Performance with $n$ target-language samples follows $S_{𝒯}(n) = S_{𝒯}(0) + \alpha \log(n+1)$, enabling direct estimation of annotation requirements (Muller et al., 2022).
• Unsupervised Typology Induction via CLI: ESL corpora of learner English can reconstruct native language similarity structure with 72.2% accuracy in typology prediction tasks, matching supervised WALS-based methods without any explicit use of typological documentation (Berzak et al., 2014).

5. Neural and Behavioral Correlates of CLI

Fine-grained probing of neural and behavioral signatures in both artificial and human bilingualism enables direct mapping of CLI processes:

• Co-activation and Language-Neuron Overlap: CLI magnitude in LMs is tracked by metrics such as the ratio of L1 tokens predicted during L2 processing and the percentage overlap of L2-relevant neurons between bilingual models (correlating at $r\approx -0.82$ with L1-L2 typological distance) (Issam et al., 29 Jan 2026).
• Oscillatory and Induced Neural Dynamics: The ROSE framework predicts that transfer errors align with specific oscillatory failures—β-band commitment deficits (competition) or θ–γ phase misalignment (sequencing). Spatiotemporal biomarkers (β burst amplitude, θ–γ phase–amplitude coupling at defined cortical loci) are linked to observable CLI-induced production errors (Uluslu et al., 26 Jan 2026).
• Bidirectional and Asymmetric Priming: Experimental manipulations reveal bidirectional priming for grammatical (shared) structures between typologically similar L1-L2 pairs, while transfer of ungrammatical (non-shared) patterns is asymmetric and contingent on L1 dominance (Issam et al., 29 Jan 2026). Orthographic overlap further amplifies CLI magnitude, as shown when scripts share representations.

6. Practical Guidelines for Harnessing or Mitigating CLI

Research provides concrete recommendations for applied multilingual NLP and language pedagogy:

• Source Language Selection: Use fine-grained typological features—especially those related to word order or tense/noun phrase structure—when choosing a source language for transfer. For any given downstream task, optimal source languages differ depending on which features most strongly impact target performance (Dolicki et al., 2021).
• Prompt Engineering in LLMs: For low-resource targets, mixed-HRL exemplars outperform English-only chemistry. Including 1–2 extraneous non-English sentences in prompts (“CIS-Multi”) can yield statistically significant further gains by activating broader multilingual subword and token spaces (Tu et al., 17 Feb 2025).
• Annotation Planning: Given a regression-based estimate of zero-shot transfer, annotation requirements for target language data can be directly computed using the few-shot law; for QA, each 10× increase in target examples yields ∼5 percentage points of accuracy gain (Muller et al., 2022).

7. Open Challenges and Future Research

• Causal Mechanisms and Spatiotemporal Biomarkers: Further interpretability is needed at the subspace and attention-head level in LLMs, and empirical work should leverage advanced MEG/EEG paradigms to validate oscillatory predictions of neurocomputational models (Uluslu et al., 26 Jan 2026, Tu et al., 17 Feb 2025).
• Task and Language Scope: Extension of CLI analyses to generative tasks (summarization, dialogue) and a broader array of language typologies remains an open agenda (Tu et al., 17 Feb 2025).
• Unsupervised Universal Typology from Transfer Artifacts: CLI-based unsupervised typological induction is promising for under-resourced languages, but requires incorporation of phonological and lexical signals beyond morpho-syntactic patterns (Berzak et al., 2014).
• Human vs. Model Divergence in CLI: Discrepancies between human and artificial representations (e.g., handling of layerwise vs. lexically-driven transfer) motivate comparative cognitive modeling (Yadavalli et al., 2023).

In sum, CLI occupies a central role in understanding and leveraging cross-lingual knowledge transfer for both cognitive science and multilingual NLP. The field increasingly anchors its insights in quantitative, mechanistic, and predictive frameworks, unifying behavioral, neurocomputational, and artificial learning perspectives.