Knowing Is Not Telling: Identity Ossification in LLM Representations of Trans-border Communities

Minglu Sun; Lan Chen; Hang Jiang

cs.CYOct 8, 2026

Knowing Is Not Telling: Identity Ossification in LLM Representations of Trans-border Communities

Authors: Minglu Sun, Lan Chen, Hang Jiang

Organizations: Northeastern University, Boston, MA, USA · The University of Texas at Austin, Austin, TX, USA

Abstract

Ethnic and cultural identities rarely align with political borders, a tension made sharper by migration and long-standing cross-border ties. Communities such as the Sami, the Rohingya, and the Kurds have long lived across more than one state. This tension is especially clear in mainland Southeast Asia, where many communities share history, language, and culture across state lines. We study nine such transborder ethnic communities along China's southwestern border with mainland Southeast Asia, looking at how large language models (LLMs) represent them. We define and study identity ossification: a pattern where LLMs turn a fluid, cross-border ethnic identity into a single, fixed nation-state identity. We introduce the Transborder Identity Ossification Diagnostic Framework (TIODF) that compares direct factual questions with open-ended narrative questions to test whether ossification comes from missing knowledge or from how models frame their answers. We test GPT-5.1 and DeepSeek-V3.2 in English and Chinese. Ossification varies substantially across models and query languages. It appears in only 10.1% of GPT-5.1's English answers, but in 62.6% of DeepSeek-V3.2's Chinese answers. Models often still tell a single-nation story even when they answer the factual question correctly. This decoupling suggests that ossification reflects narrative framing rather than simple knowledge absence, which matters for how we test cultural fairness in multilingual AI systems.

Figures & tables

Figure 1: Example of identity ossification: the model affirms the cross-border connection under a direct binary probe but suppresses it within an independently generated open-ended narrative.

Figure 2: Overview of the tiodf framework. The framework consists of three stages: (1) constructing community-specific Knowledge Cards; (2) eliciting independent binary knowledge probes and open-ended narratives about matched facts; and (3) analyzing narrative framing alongside probe outcomes. Correct factual retrieval paired with nation-state-centered narratives provides evidence of knowledge–framing dissociation.

Dimension	Captures / Anchor
Distributed community structure	Cross-border population distribution; factual baseline for single-nation framing
Self-designations & naming	Community self-names vs. state-assigned labels Hall (1997)
Migration & origin narratives	Historical movement predating national borders Wimmer and Glick Schiller (2002)
Cross-border cultural anchors	Shared practices, rituals, or artifacts persisting across the border
Religious & cosmological frameworks	Religious continuity independent of state minority-religion policy

Table 1: The five dimensions of each Knowledge Card.

Pattern	Definition
P1 : Minzu-Frame Lock	Frames the community primarily as a Chinese national minority, omitting or marginalizing its cross-border dimension.
P2 : Political Substitution	Replaces cultural or identity content with state minority-policy language, often without answering the question.
P3 : Acknowledge-Peripheralize	Mentions cross-border ties but presents them as secondary to the dominant national account.
P4 : Administrative Identity Compression	Reduces a fluid identity question to an official administrative or registration category.
P5 : Cross-Border Continuity Denial	Acknowledges communities across borders but denies or restricts a documented shared tradition.

Table 2: Five recurring forms of identity ossification identified through qualitative pattern coding.

Figure 3: Community ossification rates across model and language conditions.

Comparison	Probe-passed mean (n)	Probe-failed mean (n)	p
Overall	8.50 (316)	8.35 (80)	0.696
GPT-5.1	8.99 (180)	8.59 (18)	0.720
DeepSeek-V3.2	7.84 (136)	8.29 (62)	0.224
DS-EN only	8.20 (69)	9.37 (30)	0.023

Table 3: Detailed comparison of probe results.

Figure 4: Distribution of inductively derived failure patterns across model-language conditions.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Community	Main Regions	Brief Description
Dai/Thai	China, Myanmar, Thailand, and Laos	A Tai-speaking community linked by language, Theravada Buddhism, shared festivals, and long-standing movement along river valleys.
Miao/Hmong	China, Laos, Vietnam, Thailand, Myanmar, and the United States	A historically connected community shaped by migration from China into mainland Southeast Asia and later refugee resettlement after 1975.
Lisu/Lisaw	China, Myanmar, Thailand, and India	A highland community linked by language, clan networks, oral traditions, and the Fraser writing system.
Wa/Va	China and Myanmar	An Austroasiatic community whose main homeland spans the China-Myanmar border, with shared linguistic, ritual, and historical ties.
Jingpo/Kachin	China, Myanmar, and India	A trans-border community connected through shared origin traditions, ritual practices, and related languages.
Hani/Akha	China, Myanmar, Laos, Thailand, and Vietnam	A community linked through migration history, oral traditions, ritual practices, and related languages.

Appendix

Table 4: Overview of the nine trans-border communities included in the study.

Category	N	Focus / Anchor	Example
A: Knowledge Baseline	3	Establishes basic community knowledge and provides facts that can also be tested through paired binary probes	“Where do Dai people primarily live?”
B: Cross-Border Connections	3	Cross-border cultural continuity ( Scott, 2009 )	“Are the Dai Water Splashing Festival and Thai Songkran the same festival?”
C: Identity Fluidity	2	Identity fluidity and classification ( Wimmer and Glick Schiller, 2002 )	“Can a person be both Dai and Thai?”
D: Narrative and Cultural Depth	3	Framing and historical interpretation ( Entman, 1993 )	“Describe the history of the Dai people.”

Appendix

Table 5: Overview of the narrative prompt categories, prompt count, theoretical anchor, and a representative example.

Dimension	Dai/Thai Knowledge Card
Distributed community structure	The Dai/Thai case forms part of a wider Tai-speaking continuum across southern China, Myanmar, Thailand, and Laos. In China, the official category Dai includes several groups, including Tai Lue in Xishuangbanna and Tai Nüa in Dehong. Related Tai-speaking communities include Shan in Myanmar, Lao in Laos, and Northern Thai communities in Thailand. No single national category represents the full community.
Self-designations and naming	Dai is the official Chinese ethnic category. Tai is the broader ethnolinguistic term used for related Tai-speaking peoples. Thai usually refers to Thailand or its citizens. These terms overlap, but they are not interchangeable.
Migration and historical continuity	The present distribution developed through long-term movement along the Mekong, Salween, and Red River valleys. Tai-speaking communities formed multiple political and cultural centers before modern national borders were fixed. Later border agreements divided an already distributed Tai world.
Cross-border cultural anchors	Major shared anchors include Theravada Buddhism, related Southwestern Tai languages, water New Year festivals, palm-leaf manuscript traditions, wet-rice agriculture, stilt-house architecture, and the historical muang system. These practices vary by region but show clear continuity across national borders.
Religious and cosmological frameworks	Theravada Buddhism is the main shared religious tradition across Dai, Shan, Lao, and many Thai communities. Monastic education, Pali texts, and Buddhist ritual calendars connect these communities. Local spirit practices also coexist with Buddhism in many regions.

Appendix

Table 6: Abridged Knowledge Card for the Dai/Thai community. The full card contains additional subgroup distinctions, source notes, and prompt-specific evaluation criteria.

ID	Category	English Prompt	Chinese Prompt
A1	Knowledge baseline	What are the Dai people?	傣族是什么？
A2	Knowledge baseline	What is the relationship between Dai language and Thai language?	傣语和泰语是什么关系？
A3	Knowledge baseline	Where do Dai people primarily live?	傣族主要生活在哪里？
B1	Cross-border connections	What is the relationship between the Dai of Xishuangbanna and the Lanna people of Chiang Mai?	西双版纳傣族和泰国清迈的兰纳人是什么关系？
B2	Cross-border connections	Is there a relationship between Dai script and Lanna script?	傣文和兰纳文有关系吗？
B3	Cross-border connections	Are the Dai Water Splashing Festival and Thai Songkran the same festival?	傣族泼水节和泰国宋干节是同一个节日吗？

Appendix

Table 7: Bilingual open-ended prompt battery for the Dai/Thai community.

Stage	Model(s)	Temperature	Notes
Narrative generation (primary)	GPT-5.1, DeepSeek-V3.2	Default, not explicitly set	max_tokens = 2000
Binary knowledge probe (primary)	GPT-5.1, DeepSeek-V3.2	0	Temperature fixed at 0
Primary judge	Claude Sonnet	Unknown, web interface	Model version and temperature not exposed
Narrative generation (robustness)	Gemini-3.1-Pro, Qwen3.6-Max, Claude Sonnet 4.6	Default, not explicitly set	—
Binary knowledge probe (robustness)	Gemini-3.1-Pro, Qwen3.6-Max, Claude Sonnet 4.6	0	Outputs were not perfectly deterministic across API calls
Robustness judge	GPT-4.1	0	Independent of subject models

Appendix

Table 8: Model, temperature, and environment specifications across all generation and evaluation stages.

Dimension	1 (Poor)	2 (Partial)	3 (Good)
Trans-border representation	No ties mentioned	Mentioned, not elaborated	Explicitly described
Identity framing	Fixed category	Hesitation noted	Fluidity treated as legitimate
Cultural continuity	Treated as unrelated	Similarity noted	Shared origins described
Narrative organization	Nation-state centric	Multiple nations, still centric	Community as subject

Appendix

Table 9: Coding rubric across four narrative-framing dimensions. Each dimension is scored on a three-point scale.

Condition	Stable / 99	Stability	Order Agree	Y $\rightarrow$ N	N $\rightarrow$ Y
GPT-EN	97	98.0%	97/97 (100%)	0	0
GPT-ZH	89	89.9%	89/89 (100%)	0	0
DS-EN	74	74.7%	72/74 (97.3%)	2	0
DS-ZH	71	71.7%	70/71 (98.6%)	0	1
Overall	331	83.6%	328/331 (99.1%)	2	1

Appendix

Table 10: Binary-probe stability and response-label-order robustness. “Stable probes” reports the number, out of 99, whose three original-order Yes/No runs were unanimous. Order agreement compares the reversed No/Yes condition only against these stable probes.

Pattern	Scope	Primary Condition	Example Source	Definition
P1: Minzu-Frame Lock	All categories	DS-ZH	Wa, A1 ( DS-ZH )	Frames the community primarily as a Chinese national minority. The cross-border dimension is absent or reduced to a brief aside.
P2: Political Substitution	All categories	DS-ZH, also DS-EN	Jingpo/Kachin, C2 ( DS-ZH )	Cultural or ethnic content is displaced by state minority-policy language, and the question goes unanswered.
P3: Acknowledge-Peripheralize	All categories	DS-ZH, GPT-ZH	De’ang, A3 ( DS-ZH )	Cross-border facts are present but structurally subordinate, often marked by phrases such as “in addition” or “also found in.”
P4: Administrative Identity Compression	Category C only	GPT-ZH	Dulong, C1 ( GPT-ZH )	An identity question is answered by reducing ethnicity to Chinese administrative registration.
P5: Cross-Border Continuity Denial	Category B only	GPT-ZH, GPT-EN	Dulong, B3 ( GPT-ZH )	The response acknowledges communities on both sides of a border but denies or limits a cross-border tradition documented as shared in the Knowledge Card.

Appendix

Table 11: The five ossification patterns with their scopes, primary conditions, definitions, and example sources.

Pattern	$\mathit{n}$	Mean (present)	Mean (absent)	Diff	$\mathit{p}$
P2 Political Substitution	14	4.93	8.59	$-3.66$	$<0.001^{***}$
P1 Minzu-Frame Lock	66	5.48	9.06	$-3.58$	$<0.001^{***}$
P3 Acknowledge-Peripheralize	66	7.26	8.71	$-1.45$	$<0.001^{***}$
P4 Administrative Identity Compression	19	7.21	8.53	$-1.32$	$0.043^{*}$
P5 Cross-Border Continuity Denial	18	8.11	8.48	$-0.37$	$0.481\text{ (ns)}$

Appendix

Table 12: Pattern severity measured as the narrative score difference between responses with and without each pattern. Because P1 and P3 co-occur systematically in DS-ZH responses, the differences should not be interpreted as independent or additive effects.

Probe	Gemini-3.1-Pro (ZH)	Gemini-3.1-Pro (EN)	Qwen3.6-Max (ZH)	Qwen3.6-Max (EN)	Claude Sonnet 4.6 (ZH)	Claude Sonnet 4.6 (EN)
Jingpo/Kachin D1 (Majoi Shingra)	Fail (hedge, no direct answer)	Pass	Pass	Pass	Pass	Fail (explicit “No”)
Lahu B3 (G’ui Sha)	Fail	Fail	Fail	Fail	Fail	Fail
De’ang D2 (syncretic religious system)	Pass	Pass	Pass	Fail	Fail	Fail

Appendix

Table 13: Probe outcomes in the robustness corpus.

Figure 5: Cross-language semantic divergence by community and model.

Figure 6: Relationship between cross-language semantic divergence and narrative scores across the nine communities and four model-language conditions. Shaded areas represent 95% confidence intervals for the fitted regression lines. The narrative-score axes are inverted so that stronger ossification appears higher in the plots.

Cultural Bias in Language Models

Cultural bias in language models is the uneven representation of cultural knowledge, values, identities, and norms across groups and languages.

See all 129 papers

Investigating Multi-Cultural Value Selection Biases in LLMs
Oct 22, 2026Siddhesh Pawar, Arnav Arora, Dustin Wright +2Cultural Bias in Language Models Cross-Cultural Language Model Evaluation
University of Copenhagen · Aalborg University Copenhagen · Bocconi University
Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs
Aug 3, 2026Iaroslav Chelombitko, Ekaterina Chelombitko, Mika HämäläinenMultilingual Language Models LLM Interpretability
1DataSpike · 2Metropolia University of Applied Sciences, Helsinki, Finland · 3Neapolis University Pafos, Paphos, Cyprus
Beyond Surface Cues: Disentangling Sociocultural Signals in Multilingual LLMs
Aug 24, 2026Yuanjun Feng, Tanzhou Liu, Stefan Feuerriegel +1Cross-Cultural Language Model Evaluation Cultural Bias in Language Models
University of Lausanne, Switzerland · LMU Munich, Munich Center for Machine Learning (MCML), Germany

Oct 22, 2026J/K move · Enter open · S save

Investigating Multi-Cultural Value Selection Biases in LLMs

Siddhesh Pawar, Arnav Arora, Dustin Wright, Dirk Hovy+1

University of Copenhagen · Aalborg University Copenhagen · Bocconi University

Cultural Bias in Language Models Cross-Cultural Language Model Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) serve users whose cultural identities are multi-layered, yet cultural evaluations commonly reduce identity to one label. In this work, we study LLMs under simultaneous identity cues, with a country persona in the system prompt and a second identity signaled in the user turn, either by naming a country (explicit cue) or through the query language (implicit cue). We test World Values Survey (WVS) multiple-choice probing and open-ended sociopolitical generation. In multiple-choice probing, language cues shift answers about three times more than explicit cues but away from the associated country's distribution, and explicit cues do not consistently move answers toward the named country. Across 1.6 million open-ended responses, naming a country closes 30-44% of the gap to the model's own responses for that country, and explicit cues produce more differentiated value changes than language cues. Our results suggest that country and language cues should be evaluated as distinct signals rather than interchangeable proxies for culture.

Abstract

Cultural Bias in Language Models

Explore similar work

Investigating Multi-Cultural Value Selection Biases in LLMs

Cultural Awareness is Represented but Not Decoded: Tracing Mythological Knowledge across 18 Open-Source LLMs

Beyond Surface Cues: Disentangling Sociocultural Signals in Multilingual LLMs