Multilingual Language Models Encode Script Over Linguistic Structure

1. Problem Statement

Multilingual Large Language Models (LLMs) organize representations for typologically and orthographically diverse languages into a shared parameter space. However, the internal organization of these representations remains unclear. A fundamental question persists: Do these models learn an abstract, language-agnostic "interlingua" (a unified semantic space), or do they rely heavily on surface-form cues (such as script and token distribution)?

While previous work has identified "language-associated neurons" that control language selection, it remains unknown whether these units encode abstract linguistic identity (e.g., syntax, genealogy) or merely react to orthographic surface forms (e.g., Latin vs. Devanagari scripts). This ambiguity is critical for understanding cross-lingual transfer, robustness, and safety in multilingual systems.

2. Methodology

The authors investigate this question using Llama-3.2-1B and Gemma-2-2B, chosen for their compact, distilled nature where representational trade-offs are more explicit than in larger models. They employ a multi-faceted approach combining Language Activation Probability Entropy (LAPE) and Sparse Autoencoders (SAEs).

A. Identification of Language-Associated Units

• LAPE: Used to identify raw neurons in MLP sublayers that activate preferentially for specific languages (low entropy across language distributions).
• SAE-LAPE: Extends LAPE to sparse latent features extracted via Sparse Autoencoders, decomposing dense activations into interpretable, monosemantic features.

B. Experimental Perturbations

The study subjects these identified units to four controlled experiments to test their sensitivity to different linguistic properties:

1. Orthographic Perturbation (Romanization): Native scripts (e.g., Hindi, Chinese) are converted to Latin script (Romanized). The study compares the set of activated units in native vs. Romanized inputs to see if they overlap.
2. Structural Perturbation (Word Shuffling): Word order within sentences is randomly shuffled while preserving token identity and script. This tests reliance on syntactic structure vs. lexical/statistical cues.
3. Typological Probing: Linear probes are trained on the activations to predict typological features (genealogy, syntax, phonology) from lang2vec. This tests if deeper linguistic structure is linearly accessible.
4. Causal Intervention: Targeted ablation (zeroing out or replacing activations) is performed on specific neuron subsets (e.g., those invariant to shuffling vs. those sensitive to it) to measure their functional necessity for generation (measured via Perplexity).

3. Key Contributions & Results

A. Orthography is the Primary Organizing Factor

• Script-Bound Representations: Language-associated units are strongly conditioned on orthography. When non-Latin languages are Romanized, the model activates near-disjoint sets of neurons compared to the native script.
• No Anglicization: Romanized inputs do not converge toward English representations; instead, they occupy a distinct, isolated latent subspace.
• Fragmentation: Even minor orthographic changes (e.g., removing diacritics) induce significant reorganization of language-associated units, indicating "capacity fragmentation" where the model allocates separate features for superficially different realizations of the same language.

B. Limited Sensitivity to Word Order

• Robustness: Disrupting word order has a minimal effect on the identity of language-associated units, particularly in raw neurons.
• Sparse Feature Instability: While raw neurons remain stable, some sparse features (SAEs) show sensitivity to word order in Latin-script languages, suggesting they rely on local distributional statistics rather than deep syntactic structure.

C. Typological Structure Emerges with Depth

• Layer-wise Accessibility: Typological properties (genealogy, syntax, phonology) become increasingly linearly decodable in deeper layers.
• Invariance Correlates with Abstraction: Neurons that are invariant to script changes (active in both native and Romanized inputs) exhibit the strongest typological alignment. Conversely, script-specific neurons encode weaker typological signals.
• Hierarchy: Genealogical features are accessible early, while abstract phonological features emerge only in the deepest layers.

D. Causal Importance vs. Representational Accessibility

• The Critical Distinction: While typological structure is accessible in deeper layers, it is not causally necessary for generation.
• Invariance Drives Generation: Causal interventions reveal that units invariant to surface perturbations (both script and word order) are functionally critical for stable generation. Ablating these "overlap" neurons causes severe degradation (e.g., script mixing within words, language switching).
• Surface Sensitivity is Redundant: Units sensitive to surface form (e.g., script-specific or order-sensitive neurons) are less critical; ablating them often results in language switching rather than structural collapse, or has negligible impact.

4. Significance and Implications

• Rejection of a Unified Interlingua: The study concludes that multilingual LMs do not converge to a single, abstract interlingua. Instead, they maintain parallel, script-conditioned subspaces.
• Surface-Form Routing: Language control and generation are mediated by robustness to surface variations rather than a centralized abstraction module. The model "routes" inputs through different internal pathways based on the script.
• Cross-Lingual Transfer Limitations: The strong dependence on orthography explains why transliteration or script normalization alone yields limited gains in cross-lingual transfer without explicit adaptation.
• Safety and Robustness: The findings suggest that "jailbreak" or safety failures triggered by changing the script of a prompt may stem from the model routing inputs through different, less-safe internal subspaces rather than semantic differences.
• Interpretability Insight: There is a fundamental gap between representational accessibility (what can be probed) and functional necessity (what drives generation). Probing shows linguistic structure exists, but causal analysis shows the model relies on surface-invariant units to function.

Conclusion

The paper establishes that multilingual LMs prioritize surface-form cues (specifically orthography) over abstract linguistic structure in organizing their internal representations. While deeper layers do encode typological information, the functional core of language generation relies on units that are invariant to surface perturbations. This challenges the assumption of a unified semantic space and highlights the fragmented nature of multilingual representation in current compact models.