Culture Funnel paper argues alignment needs better cultural training data

A new arXiv paper argues that current cultural alignment methods focus too heavily on inference-time interventions. The authors describe a “cultural data funnel” in which explicit cultural signals decline during post-training.

The paper uses a tagging framework across pretraining, fine-tuning, alignment, and reasoning datasets to show how geographically concentrated and task-specialized data can dominate later stages. Multilinguality helps, but does not guarantee balanced cultural representation.

The work is a useful reminder that alignment problems often start in the data pipeline, not only in prompts or refusal policies.