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

Summary

Large language models often carry a Western-leaning cultural bias. Existing cultural alignment methods, such as persona prompting or fine-tuning on value surveys like WVS, have two weaknesses. Broad value data transfers poorly to concrete culture-sensitive tasks such as offensive language detection. Training one model on many cultures causes their conflicting norms to interfere with each other.

The authors propose CultureManager to address both problems. It starts from task labels, a few unlabeled task examples and a list of target cultures. GPT-4o writes both task-specific and general cultural search queries. A lightweight SearchGPT agent then collects culturally grounded web pages, and GPT-4o summarizes them and turns them into labeled training samples in the target task's format. Each culture gets its own LoRA adapter trained on its pooled multi-task data. At inference, a zero-shot prompted router picks which adapter to use.

In experiments spanning ten tasks where culture matters and the cultures of five countries, CultureManager consistently beats prompt-based and fine-tuning baselines. The results suggest that effective cultural alignment needs task-aware data and modular per-culture handling.

Key points

• Data synthesis pipeline: For each query, the search agent fetches the top-10 Google results. GPT-4o condenses them into a culture-grounded passage, which anchors the generated samples in real web content.
• Adapter settings:
  • CultureLLM benchmark: LoRA rank 8, α=32, dropout 0.05, applied to q_proj and v_proj.
  • CulturalBench benchmark: rank 16, applied to all attention projections.
• Culture router: It reaches over 95% accuracy on most tasks using a roughly 7B-parameter model with zero-shot prompting. Spam detection is the weakest task.
• Llama-3.3-70B-Instruct on Bengali: CultureManager's average score is 50.73, compared with 41.46 for CultureLLM and 33.01 for the original model. Scale alone does not provide cultural competence.
• Further findings:
  • Two-stage training (CultureSFT followed by synthetic task data) beats either stage alone on four of the five cultures.
  • Zero-shot GPT-5.1 matches or trails a fine-tuned 7B CultureManager on several tasks.