Cultural Adaptation within Repeated Social Interactions

Chen Cecilia Liu; Roi Reichart; Anna Korhonen

cs.CLNov 5, 2026

Cultural Adaptation within Repeated Social Interactions

Authors: Chen Cecilia Liu, Roi Reichart, Anna Korhonen

Organizations: Language Technology Lab - University of Cambridge · Technion - Israel Institute of Technology

Abstract

Much of human culture is tacit and is typically learned through situated social interaction. As LLMs are deployed as interactive agents across diverse communities, assessing their cultural competence demands moving beyond existing question-answering approaches, shifting from what they know to how they adapt. In this work, we broaden the scope of cultural adaptation in NLP by investigating whether LLMs can adapt through repeated interactions via a sandbox environment. By reframing cultural adaptation as a continuous process of norm assimilation through social interaction, our work demonstrates that memory mechanisms foster longitudinal learning across repeated interactions, facilitate transfer to different social settings, and enable multilingual norm assimilation across five languages. Together, our findings lay the groundwork for continuous, interaction-driven cultural adaptation in LLMs.

Figures & tables

Figure 1: Overview of the setup. Left: An NPC with an assigned cultural identity interacts with an unassigned Learner agent driven by a social goal. To incentivize learning implicit norms, the goal is granted only when the learner demonstrates associated cultural norms and avoids cultural faux pas. A judge model evaluates the dialogue session and assigns a social reputation score in the end. Right: We evaluate how reasoning and memory mechanisms shape longitudinal learning behaviours and norm assimilation across repeated interaction episodes.

Model	CN	FR	JP	NZ	SA	Avg.
Gemma4-E4B	0.01 $\pm$ 0.04	0.22 $\pm$ 0.13	-0.05 $\pm$ 0.01	0.33 $\pm$ 0.08	0.15 $\pm$ 0.02	0.13
Gemma4-E4B SR	0.00 $\pm$ 0.06	0.06 $\pm$ 0.03	-0.06 $\pm$ 0.04	0.42 $\pm$ 0.03	0.25 $\pm$ 0.03	0.13
Gemma4-E4B ${}_{\text{Memory}_{\text{BDI}}}$	0.04 $\pm$ 0.05	0.36 $\pm$ 0.02	0.10 $\pm$ 0.05	0.53 $\pm$ 0.05	0.23 $\pm$ 0.03	0.25*
Gemma4-E4B ${}_{\text{Memory}_{\text{R}}}$	0.23 $\pm$ 0.04	0.38 $\pm$ 0.02	0.23 $\pm$ 0.06	0.51 $\pm$ 0.12	0.35 $\pm$ 0.04	0.34*
Qwen3-8B	-0.47 $\pm$ 0.03	-0.24 $\pm$ 0.02	-0.27 $\pm$ 0.06	0.21 $\pm$ 0.05	-0.11 $\pm$ 0.10	-0.17
Qwen3-8B SR	-0.23 $\pm$ 0.02	-0.11 $\pm$ 0.07	-0.08 $\pm$ 0.00	0.27 $\pm$ 0.06	-0.06 $\pm$ 0.03	-0.04*

Table 1: Model performance across cultures: China (CN), France (FR), Japan (JP), New Zealand (NZ), and Saudi Arabia (SA). Social Reputation

\in[-1.0,1.0]

: episode-averaged score measuring adherence to cultural norms (

-1.0

indicates any violations or cultural faux pas). ∗ : a statistically significant improvement over the default (

p<0.05

Panel of Figure 2 — Figure 2: Distribution of social reputation judgments under $\text{Memory}_{\text{R}}$ across models and cultures for English-language interactions. For each cultural group, the three bars (left to right) represent sequential repeated interaction episodes. Stacked segments denote the proportion of scenarios judged as Positive (+1; norm-aligning), Neutral (0; partially adhering), or Negative (-1; norm-violating). China (CN), France (FR), Japan (JP), New Zealand (NZ), and Saudi Arabia (SA).

Gemma4-E4B
Model	Ep.1	Post	$p_{\text{within}}$	$\beta$	$p_{\beta}$
$\text{Memory}_{\text{BDI}}$	$+0.22$	$+0.25$	$0.1380$	$+0.05$	$0.1260$
$\text{Memory}_{\text{R}}$	$+0.19$	$+0.34$	$<0.0001^{*}$	$+0.10$	$0.0016^{*}$
Qwen3-8B
$\text{Memory}_{\text{BDI}}$	$-0.01$	$-0.00$	$0.3838$	$+0.01$	$0.7241$
$\text{Memory}_{\text{R}}$	$-0.17$	$+0.01$	$<0.0001^{*}$	$+0.08$	$0.0161^{*}$

Table 2: Longitudinal social reputation performance across interaction episodes. Evaluated across

N=250

paired scenario encounters across all cultures. Ep.1 : Episode 1. Post : average performance of Ep.2 and Ep.3.

\boldsymbol{\beta}

: the fixed-effect longitudinal trajectory slope from a linear mixed model (LMM), with significance

p_{\beta}

. ∗ indicates statistically significant (

p<0.05

Figure 3: Embedding similarity between inferred norms and ground truth across models and cultures under

\text{Memory}_{\text{R}}

. The dashed line indicates the background embedding similarity for randomly paired norms for that model family.

Figure 4: Change in social reputation scores on unseen scenarios using norms inferred from repeated interactions (

\text{Memory}_{\text{R}}

) relative to the default.

Model	CN	FR	JP	SA	Avg.
Gemma4-E4B	-0.06	0.21	0.07	0.25	0.12
Gemma4-E4B ${}_{\text{Memory}_{\text{R}}}$	0.23	0.34	0.20	0.54	0.33
Qwen3-8B	-0.21	-0.05	-0.14	-0.18	-0.15
Qwen3-8B ${}_{\text{Memory}_{\text{R}}}$	-0.06	-0.04	-0.07	-0.07	-0.06

Table 3: Averaged multilingual performance (social reputation score,

\text{Memory}_{\text{R}}

) across four cultures: China (CN), France (FR), Japan (JP), and Saudi Arabia (SA).

Figure 5: Distribution of social reputation judgments under

\text{Memory}_{\text{R}}

for 5 episodes. For each culture, the bars (left to right) represent sequential repeated interaction episodes. Stacked segments denote the proportion of scenarios judged as Positive (+1; norm-aligning), Neutral (0; partially adhering), or Negative (-1; norm-violating).

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Culture	Norms from Model	Ground Truth
New Zealand	Initial business interactions prioritize shared, relatable observations about the setting to establish a comfortable and grounded rapport.	New Zealanders may use humour throughout business discussions to lighten the setting. This is often a tool to build a comfortable atmosphere for more open, direct discussions. Reciprocate this where you can.
China	When approaching a respected master in traditional Chinese contexts, deference to the holistic philosophical framework of health (e.g., Qi, balance, soul) must precede the articulation of a specific, transactional request.	In the context of traditional Chinese medicine, Chinese people and practitioners emphasize the importance of prevention and the use of food as medicine, guided by the belief in Qi and life energy.
Japan	Respect for the host’s established pacing and procedural guidance supersedes the eagerness to initiate discussion or take initiative.	It is common for the host to give a quick speech greeting everyone before discussing the topic of business.

Appendix

Table 4: Examples of norms inferred versus the ground-truth norms from Gemma4-E4B.

Age Group	Percentage
17–24	0.95%
25–34	27.06%
35–44	33.47%
45–54	29.59%
55–64	8.07%
65+	0.87%

Appendix

Table 5: Age distribution of scenarios. Most NPC characters are within 25–54 years old (typical working age) since many ground truth norms from CulturalForest are related to business settings.

We have 20 synthetic {culture} profiles and settings that’s generated by AI models.

We’d like your help on assessing the following dimensions of these profiles and settings.

1. Acceptability

How possible is this profile and setting within your culture? Could someone like this plausibly exist in your culture?

Consider whether the combination of characteristics, occupations, resources, and events could plausibly exist. Do not penalize unusual or rare characteristics merely for being unusual.

Options:

Appendix

Table 6: Human evaluation instruction.

Example Scenario
General Setting & Context
Culture	New Zealand
Setting	A sunlit timber-clad meeting room in a converted packing shed outside Hastings… Large sliding glass doors stand partially open to the terrace, letting in a light breeze.
NPC
Demographics	Stacey Wilson, 41 years old, Female
Background	…regional logistics director in Hawke’s Bay who manages transport and cold-storage distribution networks…

Appendix

Table 7: Example synthetic scenario, shortened for presentation. The ground-truth norm set is sourced from CultureForest Ye et al. (2026) and consists of three norms. The scenario generated based on CultureBank Shi et al. (2024) consists of one norm.

Model	CN	FR	JP	NZ	SA	Avg.
Qwen3-32B	-0.05	0.21	0.14	0.36	0.23	0.18
Qwen3-32B ${}_{\text{Memory}_{\text{R}}}$	0.14	0.39	0.17	0.50	0.37	0.31

Appendix

Table 8: Qwen3-32B performance across cultures: China (CN), France (FR), Japan (JP), New Zealand (NZ), and Saudi Arabia (SA). Social Reputation

\in[-1.0,1.0]

: episode-averaged score measuring adherence to cultural norms (

-1.0

indicates any violations or cultural faux pas). Qwen3-32B Memory R demonstrates both a statistically significant improvement over the default baseline (

p=0.0001

) and sustained longitudinal learning (

\beta=+0.08\text{ pts/ep}

p=0.0160

Panel of Figure 6 — Figure 6: Distribution of social reputation judgments under $\text{Memory}_{\text{R}}$ across models and cultures for multilingual interactions. For each cultural group, the three bars (left to right) represent sequential repeated interaction episodes. Stacked segments denote the proportion of scenarios judged as Positive (+1; norm-aligning), Neutral (0; partially adhering), or Negative (-1; norm-violating).

Online Language Model Adaptation

See all 13 papers

Don' t Box Me In: Dynamic Cultural Adaptation and Cognitive Tracking for Social Understanding
Aug 23, 2026Chongyuan Dai, Yaling Shen, Shengeng Tang +2Cultural Knowledge in Language Models
Hefei University of Technology · Monash University
Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation
Apr 20, 2026Mehwish Nasim, Sanjeevan Selvaganapathy, Neel Ganapathi Sabhahit +6Multilingual Language Models Multilingual Language Model Evaluation
1Network Analysis and Social Influence Modelling (NASIM) Lab · 3Computational Social Science & Systems Analysis University of Münster · School of Computing Macquarie University
LLMs Infer Cultural Context but Fail to Apply It When Responding
Jun 16, 2026Yisong Miao, Jian Zhu, Vered ShwartzPragmatic Reasoning in Language Models Cultural Bias in Language Models
National University of Singapore · University of British Columbia · Canada CIFAR AI Chair, Vector Institute

Aug 23, 2026J/K move · Enter open · S save

Don' t Box Me In: Dynamic Cultural Adaptation and Cognitive Tracking for Social Understanding

Chongyuan Dai, Yaling Shen, Shengeng Tang, Hui Ma+1

Hefei University of Technology · Monash University

Cultural Knowledge in Language Models

Open paper AskPDF arXiv

Social interaction increasingly takes place in multicultural settings, where individuals may draw on multiple cultural influences and adapt their communicative behavior across contexts. Despite recent advances in equipping Large Language Models (LLMs) with social understanding capabilities, existing approaches often model culture as a static demographic attribute, limiting their ability to accommodate hybrid and dynamically expressed communicative preferences. Therefore, in this paper, we propose DyCAC, a training-free framework that achieves fluid social alignment by incorporating Dynamic Cultural Adaptation with continuous Cognitive tracking. Rather than inferring a fixed cultural identity, DyCAC models culturally relevant communicative preferences as a time-varying mixture of population-level cultural reference profiles. This reference-based representation is further calibrated using dialogue-style signals observed in the ongoing interaction, enabling the model to capture both composite cultural influences and turn-level shifts in communicative behavior. In parallel, a memory module driven by Theory of Mind (ToM) continuously tracks the cognitive states of the interlocutor. Extensive experiments on interactive social and cultural benchmarks demonstrate the superiority of our approach. The proposed framework outperforms existing baselines, exhibiting enhanced social intelligence and broad adaptability across varied multicultural contexts.

Abstract

Online Language Model Adaptation

Explore similar work

Don' t Box Me In: Dynamic Cultural Adaptation and Cognitive Tracking for Social Understanding

Do LLMs Use Cultural Knowledge Without Being Told? A Multilingual Evaluation of Implicit Pragmatic Adaptation

LLMs Infer Cultural Context but Fail to Apply It When Responding