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.
Figures & tables
Figure 1: An overview of the DyCAC framework. xt and yt are the dialogue input and generated response, respectively. pt is the perceptual state, and mt is the memory updated by reasoner ΦToM . For cultural inference, variables Ht,πt,i,ℓt,i,qt,i represent the culture space, cultural prior, compatibility score, and normalized weight, respectively. The final cultural profile ct interpolates between the inferred profile cˉt and direct style c^t via weight λ .
Model
Method
Bel [0, 10]
Rel [-5, 5]
Kno [0, 10]
Sec [-10, 0]
Soc [-10, 0]
Fin [-5, 5]
Goal [0, 10]
Overall
Qwen2.5-7B
Vanilla
8.73
3.24
5.65
-0.34
-0.14
0.41
7.39
3.56
CoT
8.68
3.16
5.90
-0.37
-0.16
0.47
7.48
3.59
ReAct
8.76
3.19
5.76
-0.31
-0.09
0.41
7.40
3.59
MAD
8.57
3.23
5.32
-0.23
-0.10
0.56
7.61
3.57
MetaMind
8.79
3.60
5.93
-0.12
-0.09
0.48
7.45
3.72
SDPO
8.66
3.50
5.38
-0.15
-0.07
0.63
7.43
3.63
Table 1: Evaluation results on the SOTOPIA benchmark. Each experiment is run three times, and the reported results are averaged. The best performance is bolded. ∗p<0.05 and ∗∗p<0.01 indicate statistically significant differences compared with the vanilla method.
Model
Method
Multimodal
Text-Only
AR
EN
ES
HI
JA
SW
ZH
Avg.
AR
EN
ES
HI
JA
SW
ZH
Avg.
Gemma3-12B
Vanilla
34.51
34.51
40.05
27.13
28.57
30.15
15.04
29.99
37.42
34.51
44.80
23.73
33.19
34.91
11.23
31.40
SocialGaze
34.38
38.66
41.03
36.73
33.33
30.83
29.89
34.98
42.52
49.73
46.19
33.82
30.50
33.25
31.19
38.17
SocialAgent
34.76
34.26
34.59
33.50
34.50
33.50
27.14
33.18
41.17
49.49
39.54
42.02
37.39
36.54
33.19
39.91
CulturalDebate
38.99
39.14
43.80
38.30
32.83
32.50
28.39
36.28
43.85
52.46
44.63
42.36
35.36
38.08
32.73
41.35
MetaMind
32.57
36.03
42.17
36.65
33.07
29.26
31.70
34.49
38.33
47.36
38.16
40.23
35.37
36.27
33.56
38.47
Table 2: Results on the CEDAR benchmark. The best performance in each column is highlighted. Languages: AR = Arabic, EN = English, ES = Spanish, HI = Hindi, JA = Japanese, SW = Swahili, ZH = Chinese.
Method
Bel
Rel
Kno
Sec
Soc
Fin
Goal
Overall
[0, 10]
[-5, 5]
[0, 10]
[-10, 0]
[-10, 0]
[-5, 5]
[0, 10]
Vanilla
8.73
3.24
5.65
-0.34
-0.14
0.41
7.39
3.56
DyCAC (Full)
8.89
3.33
6.11
-0.16
-0.12
0.60
7.93
3.80
Ablation Study on Framework Modules
w/o Perception
8.75
3.16
5.60 ∗∗
-0.16
-0.13
0.53
7.74
3.64
w/o Cultural Profile Inference
8.72
2.98 ∗
5.69 ∗
-0.15
-0.15
0.41
7.77
3.61
Table 3: Ablation of the framework modules and ToM mechanism on the SOTOPIA benchmark using Qwen2.5-7B. ∗p<0.05 and ∗∗p<0.01 denote statistically significant performance drops compared to the full framework.
Figure 2: Performance comparison between Soft Cultural Adaptation (SCA) and Single-Culture Adaptation (SiCA) on the SOTOPIA and CEDAR benchmarks.
Model
Method
Avg.
Gemma3-12B
Vanilla
29.99
31.40
First-Person Role-Play
29.72
31.66
Third-Person Role-Play
33.35
34.75
SiCA
34.99
41.15
SCA (Ours)
37.55
41.90
Qwen3.5-35B
Vanilla
31.50
37.12
Table 4: Method comparison on the CEDAR benchmark. The two Avg. scores denote the results on multimodal (left) and text-only (right) subsets, respectively.
Figure 3: A qualitative example of dynamic cultural adaptation in our framework. As the dialogue progresses, our framework captures subtle semantic shifts (Left) to continuously update the latent cultural profile, translating these beliefs into effective strategic actions (Right).
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Dimension
Level
Culture-Conditioned Strategy
PDI
high
Show respect for hierarchy and preserve face; phrase pushes as suggestions rather than commands.
medium
Be polite and respectful without sounding stiff.
low
Speak as an equal; be plain, direct, and collaborative.
IDV
high
Frame tradeoffs in terms of personal choice and individual benefit.
medium
Balance personal benefit with relationship impact.
low
Frame choices in terms of harmony, reciprocity, and shared outcomes.
Appendix
Table 5: Detailed cultural strategy directives.
Figure 4: Efficiency comparison of our framework and baseline methods on SOTOPIA. We report average latency, average number of output tokens, and average total tokens consumed per dialogue turn.
Subset
Qwen2.5-7B
Qwen2.5-7B + DyCAC
Δ
High-shift
0.7322
0.8182
+0.0860
Low/no-shift
0.7043
0.7529
+0.0486
Appendix
Table 6: Performance stratified by the frequency of dynamic cultural shifts on SOTOPIA.
Large language models (LLMs) are increasingly deployed in culturally sensitive real-world tasks. However, existing cultural alignment approaches fail to align LLMs' broad cultural values with the specific goals of downstream tasks and suffer from cross-culture interference. We propose CultureManager, a novel pipeline for task-specific cultural alignment. CultureManager synthesizes task-aware cultural data in line with target task formats, grounded in culturally relevant web search results. To prevent conflicts between cultural norms, it manages multi-culture knowledge learned in separate adapters with a culture router that selects the appropriate one to apply. Experiments across ten national cultures and culture-sensitive tasks show consistent improvements over prompt-based and fine-tuning baselines. Our results demonstrate the necessity of task adaptation and modular culture management for effective cultural alignment.
Large language models are increasingly used for social decision-making situations that require balancing cultural norms with personal preferences. For example, a user preferring honesty might ask whether to correct a coworker publicly when local norms favor indirect feedback. Yet existing research studies cultural alignment and personalization largely separately. We introduce PACT, the Personal-Preference and Cultural-Norm Trade-off framework, which evaluates whether models choose to follow a cultural norm or allow personal preferences. We find that LLMs vary in how rigidly they enforce cultural norms, with behavior shifted more by country context (7.8%) than age (1%) and gender (0.7%) and shifting non-uniformly after instruction tuning. Furthermore, our five-country human study on PACT shows that culture-following in humans is mainly driven by scenario country, with the lowest agreement when participants judge their own cultural contexts, showing within-culture pluralism. Finally, human-LLM alignment experiments show that models can match majority choices, but fail to capture response distributions and uncertainty (with best correlations reaching only 0.24). Together, these findings motivate alignment evaluations that go beyond majority to capture cultural pluralism and disagreement in social judgment.
Angana Borah, Isabelle Augenstein, Rada Mihalcea
University of Michigan - Ann Arbor · University of Copenhagen
Recent work has shown that LLMs overrepresent dominant cultures, particularly Western ones, while marginalizing others. We investigate whether this affects models' ability to generate culturally adapted responses by evaluating their use of local measurement units based on the user's perceived cultural background. We introduce Cultural and Pragmatic Response Inference (CAPRI), a dataset of conversations with varying levels of cultural cues. Experiments with state-of-the-art LLMs show that models can infer cultural background and recall relevant conventions, but often fail to utilize the information to adapt their answers to the relevant cultural conventions, unless explicitly prompted to perform the tasks sequentially. We further evaluate adaptation to the interpretation of time and quantity expressions, two subjective language grounding dimensions that are affected by culture. We find that models increasingly adapt their answers as cultural cues accumulate, but their priors are not culture-neutral, sometimes aligning with the model's country of origin. Overall, CAPRI provides a resource for future research aimed at narrowing the gap between cultural knowledge and culturally adaptive language generation.
Yisong Miao, Jian Zhu, Vered Shwartz
National University of Singapore · University of British Columbia · Canada CIFAR AI Chair, Vector Institute