Mind the Gap in Cultural Alignment: Task-Aware Culture Management for Large Language Models

Binchi Zhang; Xujiang Zhao; Jundong Li; Haifeng Chen; Zhengzhang Chen

cs.CLFeb 25, 2026

Mind the Gap in Cultural Alignment: Task-Aware Culture Management for Large Language Models

Authors: Binchi Zhang, Xujiang Zhao, Jundong Li, Haifeng Chen, Zhengzhang Chen

Organizations: University of Virginia · NEC Laboratories America

Abstract

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.

Figures & tables

Figure 1: Overview of the CultureManager pipeline. CultureManager consists of three parts: Search Query Generation (left), Task-aware Cultural Data Synthesis (center), and Culture Management (right).

Method	CultureLLM Benchmark			CulturalBench
	ar-hate	ar-offensive	ar-average	China
Original	23.21	38.72	30.97	36.07
Prompt	36.53	76.15	56.34	25.00
TaskSFT	26.60 $\pm$ 2.45	63.50 $\pm$ 0.86	45.05	45.41 $\pm$ 1.16
CultureSFT	33.11 $\pm$ 5.98	52.79 $\pm$ 1.60	42.95	42.54 $\pm$ 1.15
CultureSFT-all	32.77 $\pm$ 4.31	42.41 $\pm$ 4.09	37.59	44.04 $\pm$ 1.22

Table 1: Experimental results of F1 scores on different culture-sensitive downstream tasks. “ar”, “bn”, “zh”, “es”, “tr” are abbreviations of “Arabic”, “Bengali”, “Chinese”, “Spanish”, and “Turkish”. The “average” column is computed by averaging the task metrics under the corresponding culture. For training-based methods, we record the mean and standard deviation of metrics across five runs with different random seeds. We bold the highest score and underline the second-highest score.

CultureLLM Benchmark			CulturalBench
Culture	Task	Labels (1 / 0)	Culture	Labels (1 / 0)
Arabic	hate speech detection	180 / 495	Chinese	62 / 174
Arabic	offensive language detection	402 / 1,598	Chinese	62 / 174
Bengali	racism detection	48 / 951	German	35 / 93
Bengali	threat detection	70 / 929	German	35 / 93
Chinese	gender bias detection	72 / 928	Korean	44 / 120

Table 2: Statistics of downstream task datasets.

Figure 2: Experimental results of the ablation study. “specific-only”: use only task-specific queries in data synthesis; “agnostic-only”: use only task-agnostic queries in data synthesis; “all-culture”: abandon knowledge management and train the model on all datasets.

Panel of Figure 3 — Figure 3: Accuracy of the culture router on different culture-sensitive tasks.

Figure 4: Results of task adaptation on different cultures. SFT represents CultureSFT; SFT+Adaptation means firstly fine-tuning the LLM using CultureSFT and then continually fine-tuning it using the synthetic data; TaskSFT denotes directly fine-tuning the original LLM using the synthetic data. The metric of each culture is computed as the mean value across the tasks under that culture.

Method	ar-hate	ar-offensive	ar-average
Original	53.34	76.88	65.11
Prompt	52.58	81.75	67.17
CultureSFT	52.66	73.91	63.29
CultureLLM	53.38	72.13	62.76
CultureManager	55.89	79.33	67.61
	bn-racism	bn-threat	bn-average

Table 3: Experimental results of culture-sensitive downstream tasks with Llama-3.3-70B-Instruct.

Cultural Alignment

Cultural alignment studies how AI systems represent and respond to the values, norms, languages, and lived contexts of different communities.

See all 45 papers

AlignCultura: Towards Culturally Aligned Large Language Models?
Apr 21, 2026Gautam Siddharth Kashyap, Mark Dras, Usman NaseemCross-Cultural Language Model Evaluation Cultural Alignment
School of Computing, Macquarie University, Australia
Prompt Programming for Cultural Bias and Alignment of Large Language Models
Mar 17, 2026Maksim Eren, Eric Michalak, Brian Cook +1LLM Alignment Cultural Alignment
Advanced Research in Cyber Systems, Los Alamos National Laboratory, Los Alamos, New Mexico, USA · Analytics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA
The Culture Funnel: You Can't Align What isn't in the Data
Jun 11, 2026Ananya Sahu, Mehrnaz Mofakhami, Daniel D'Souza +3Cultural Alignment Cultural Bias in Language Models
CohereCorrespondingLabs

Apr 21, 2026J/K move · Enter open · S save

AlignCultura: Towards Culturally Aligned Large Language Models?

Gautam Siddharth Kashyap, Mark Dras, Usman Naseem

School of Computing, Macquarie University, Australia

Cross-Cultural Language Model Evaluation Cultural Alignment Cultural Bias in Language Models Language Model Bias Evaluation Cultural Knowledge in Language Models

Open paper AskPDF arXiv

Cultural alignment in Large Language Models (LLMs) is essential for producing contextually aware, respectful, and trustworthy outputs. Without it, models risk generating stereotyped, insensitive, or misleading responses that fail to reflect cultural diversity w.r.t Helpful, Harmless, and Honest (HHH) paradigm. Existing benchmarks represent early steps toward cultural alignment; yet, no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO's principles of cultural diversity w.r.t HHH paradigm. Therefore, to address this gap, we built Align-Cultura, two-stage pipeline for cultural alignment. Stage I constructs CULTURAX, the HHH-English dataset grounded in the UNESCO cultural taxonomy, through Query Construction, which reclassifies prompts, expands underrepresented domains (or labels), and prevents data leakage with SimHash. Then, Response Generation pairs prompts with culturally grounded responses via two-stage rejection sampling. The final dataset contains 1,500 samples spanning 30 subdomains of tangible and intangible cultural forms. Stage II benchmarks CULTURAX on general-purpose models, culturally fine-tuned models, and open-weight LLMs (Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B). Empirically, culturally fine-tuned models improve joint HHH by 4%-6%, reduce cultural failures by 18%, achieve 10%-12% efficiency gains, and limit leakage to 0.3%.

Abstract

Cultural Alignment

Explore similar work

AlignCultura: Towards Culturally Aligned Large Language Models?

Prompt Programming for Cultural Bias and Alignment of Large Language Models

The Culture Funnel: You Can't Align What isn't in the Data