Nationality encoding in language model hidden states: Probing culturally differentiated representations in persona-conditioned academic text

Abstract: LLMs are increasingly used as writing tools and pedagogical resources in English for Academic Purposes, but it remains unclear whether they encode culturally differentiated representations when generating academic text. This study tests whether Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating research article introductions conditioned by British and Chinese academic personas. A corpus of 270 texts was generated from 45 prompt templates crossed with six persona conditions in a 2 x 3 design. Logistic regression probes were trained on hidden-state activations across all 35 layers, with shuffled-label baselines, a surface-text skyline classifier, cross-family tests, and sentence-level baselines used as controls. Probe-selected token positions were annotated for structural, lexical, and stance features using the Stanza NLP pipeline. The nationality probe reached 0.968 cross-validated accuracy at Layer 18, with perfect held-out classification. Nationality encoding followed a non-monotonic trajectory across layers, with structural effects strongest in the middle to upper network and lexical-domain effects peaking earlier. At high-signal token positions, British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative or process-oriented vocabulary, while Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural or internationalisation vocabulary. However, sentence-level analysis found no significant nationality differences in the full generated surface text. The findings extend probing methodology to a sociolinguistic attribute and have practical implications for EAP and language pedagogy.

• The paper demonstrates that nationality information is reliably encoded in intermediate hidden layers, achieving near-perfect probe accuracy.
• By employing centroid- and token-level probing methods, it reveals distinct linguistic structures—such as pre/post modifiers—that reflect cultural rhetorical differences.
• The findings highlight a risk of culturally hollow writing, where internal national signals do not appear in the generated text yet influence academic representation.

Nationality Encoding in LLM Hidden States: Probing Culturally Differentiated Representations

Overview and Motivation

The paper "Nationality encoding in LLM hidden states: Probing culturally differentiated representations in persona-conditioned academic text" (2604.10151) investigates whether LLMs, specifically Gemma-3-4b-it, internalize and encode information about nationality when conditioned on distinct academic personas and tasked with generating research article introductions. The study addresses a critical gap in current interpretability research: while surface-level output analyses have documented cultural or nationality bias, little is known about the internal representational structures underlying these phenomena, especially as they pertain to sociolinguistic attributes in educational domains.

By focusing on the hidden activations of the model across its layers, the authors seek to answer four central questions: (RQ1) whether nationality-linked information is encoded in hidden states; (RQ2) at which network depths such encoding is strongest; (RQ3) the linguistic features characterizing such encoded positions; and (RQ4) whether signals detectable in hidden states surface in the generated text.

Experimental Design and Methodology

The study employs Gemma-3-4b-it, an instruction-tuned 4B-parameter transformer, with a controlled 2 × 3 experimental design that manipulates both nationality (British, Chinese) and persona cohort (EMI/CMI postdoc, student). From a total of 270 generated academic texts—derived from 45 prompt templates across knowledge and genre spectrums—hidden states for all model-generated tokens are extracted across 35 layers.

Probing methodology consists of two key approaches:

• Centroid-level Probing: Logistic regression classifiers trained on layer-wise averaged (centroid) hidden states to predict nationality, medium, and academic role.
• Token-level Probing: Classifiers applied at each token position, yielding granular signals and permitting the extraction of high-signal tokens/windows that most strongly encode nationality-relevant distinctions.

Rigorous controls are implemented, including shuffled-label baselines, a surface-text TF-IDF skyline classifier, and cross-prompt-family generalization tests to ensure that probe performance genuinely reflects recoverable representational structure, not artifacts of probe expressivity or prompt-specific lexical cues. Furthermore, linguistic annotation (structural, lexical, stance) of high-signal tokens using the Stanza pipeline provides a detailed lens on how these internal encodings manifest in the micro-structure of text generation.

Numerical Results and Key Findings

Layerwise Encoding and Probe Performance

The nationality probe achieves cross-validated accuracy of 0.968 at Layer 18, with perfect (1.000) held-out accuracy, substantially outperforming both the shuffled baseline (approx. 0.5) and the surface-text skyline classifier (approx. 0.46). Selectivity scores remain high and cross-family generalization stays robust (≥0.84 in 5/6 configurations), except for prompts related to theoretical content, where the transfer signal attenuates.

Layer trajectory analysis reveals a non-monotonic pattern: nationality encoding is strongest in middle-to-upper layers (notably Layers 10 and 18), declining towards the output layer, consistent with studies showing a transition from local to higher-order abstractions in transformer models.

Structural, Lexical, and Stance Differentiation

Annotation of probe-selected windows identifies statistically robust structural and stance differences:

• British-associated representations: Higher postmodification, passive voice, hedging, boosting, and usage of evaluative/processual vocabulary.
• Chinese-associated representations: Greater premodification, nominal predicates, and a preference for sociocultural/internationalization-oriented vocabulary.

Contrasts between window- and token-level analyses suggest that modifier structures (pre/post) and phrase-internal slotting are the main carriers of this signal, while clause slot is not significantly involved. These structural patterns partially align with known cross-cultural rhetorical differences, but with some internal recombination and model-specific stylization.

Absence of Surface-Level Nationality Signals

Sentence-level statistical tests return a null result: none of the structural or stance differences observed at the most discriminative token positions are reliably observable in the global surface text. The implications are twofold: (1) the model’s internal nationality encoding is precise and layer-/position-specific; (2) ordinary reading of generated texts is unlikely to reveal these underlying cultural patterns, a phenomenon characterized as "culturally hollow writing."

Independence from Medium/Role Confounds

Control tests show that the nationality signal is not explained by instructional medium or academic role. Significant effects are confined to minority subgroups (notably the CMI-only cohort), warranting cautious interpretation but not overturning the broader conclusion of representational encoding.

Theoretical and Practical Implications

Culturally Hollow Writing and Representational Risk

The demonstration that LLMs encode nationality-linked rhetorical structures internally, but suppress them in surface output, provides empirical grounding for concerns around "culturally hollow writing." This is text that is fluent and genre-appropriate but lacks culturally situated depth—posing a representational risk, especially in English for Academic Purposes (EAP) and multilingual pedagogical domains.

The findings also operationalize aspects of algorithmic injustice: the internalization and transmission of dominant cultural-rhetorical patterns that are opaque to both student and educator end-users, reinforcing the need for critical AI literacy and scrutiny of LLM-mediated academic conventions.

Implications for Linguistic and Interpretability Research

Methodologically, the study extends the boundaries of probing from syntactic/semantic recovery to the sociolinguistic/cultural domain, demonstrating that fine-grained cultural differentiation is recoverable, structured, and linguistically meaningful in hidden states. The non-monotonic encoding trajectory highlights the importance of mid-network abstractions and layer-wise analysis in cultural or bias-focused interpretability studies.

For applied linguistics, mechanistic evidence of differentiated encoding substantiates concerns regarding de facto normativity in AI-generated academic output, even when explicit output-level bias is elusive.

Critical Evaluation and Limitations

The study’s primary limitation is the demonstration of recoverability, not causal use; it cannot show that these internal encodings are determinative for generation, only that they are present. The use of a linear probe restricts interpretability on encoding compactness, and exclusive dependence on model-generated text means that triangulation with human-authored corpora remains an open desideratum. Furthermore, the transfer attenuation for theory-related prompts highlights topic-dependent representational pathways.

Future Directions

Several avenues for further work are evident:

• Application of causal mediation/intervention methods (e.g., activation patching, amnesic probing) to adjudicate the generative influence of nationality representations.
• Triangulation with matched human-authored corpora to establish ecological validity.
• Comparative studies across LLMs with varied pretraining data to identify model- or dataset-specific effects.
• Focused studies on whether similar signals persist in strictly technical (non-pedagogical) academic domains.

Conclusion

This paper establishes that nationality, as a sociolinguistic attribute, is robustly encoded in the hidden states of a large transformer-based LLM when generating academic text under controlled persona prompts. The internal representations are linguistically characterizable, layer- and position-specific, and align in part with cross-cultural rhetorical theory. However, these encodings are not reliably visible at the level of generated output, foregrounding the tension between surface fluency and cultural depth, with significant implications for EAP, AI literacy, and the mitigation of representational bias in LLM deployment.