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.
Introduction
While Large Language Models (LLMs) demonstrate remarkable capabilities, they often exhibit a Western-centric bias. Current cultural alignment strategies typically rely on broad prompting (e.g., "Act like you are from Turkey") or fine-tuning on general value surveys like the World Values Survey (WVS). However, these methods face two significant hurdles: the Generalization Gap and Multicultural Interference.
The Generalization Gap refers to the disconnect between broad cultural values and specific task requirements. For instance, knowing that a culture values family does not inherently help a model detect offensive slang or subtle gender bias in that language. Multicultural Interference occurs when training a single model on diverse, often conflicting cultural norms, leading to degraded performance—a "jack of all trades, master of none" scenario.
Methodology: The CultureManager Pipeline
To address these challenges, we introduce CultureManager, a three-stage pipeline designed for task-specific cultural alignment.
The pipeline consists of three core components:
- Task-Aware Data Synthesis: Instead of static surveys, the system uses an LLM to generate search queries based on a target culture and task. A web search agent retrieves real-world context, which is then converted into labeled training samples matching the downstream task format.
- Modular Culture Management: To prevent interference, CultureManager keeps the base LLM frozen and trains lightweight LoRA (Low-Rank Adaptation) adapters for each specific culture. This isolates cultural knowledge and prevents cross-contamination.
- Culture Routing: During inference, a "culture router" analyzes the input text to identify the relevant culture and dynamically activates the corresponding adapter.
Experimental Results
We evaluated CultureManager using the CultureLLM Benchmark (covering five cultures across ten tasks) and CulturalBench.
As shown in the results, CultureManager consistently outperforms both zero-shot prompting and broad fine-tuning baselines. For example, in Arabic hate speech detection, it achieved an F1 score of 44.98, a significant jump from the base model's 23.21. Furthermore, the culture router demonstrated high precision, achieving over 95% accuracy in identifying the correct cultural context for most tasks.
Discussion and Conclusion
Our findings emphasize that cultural sensitivity is highly contextual. Broad value alignment often misses nuances—like the varying interpretations of gestures or slang—that web-grounded synthetic data can capture. Moreover, the success of the modular adapter approach confirms that "siloing" cultural knowledge is more effective than blending it into a single set of model weights.
While CultureManager currently focuses on a predefined set of national cultures, future work will explore more granular sub-cultures and mixed-culture inputs where multiple contexts overlap. By bridging the gap between general values and task-specific needs, CultureManager provides a scalable path toward truly global and culturally aware AI.
Summary updated