Frontier vision-language models learn from web data that overrepresents North America and Western Europe. As a result, a model may describe a Brooklyn bagel shop accurately while misidentifying a Thai night market, a Filipino jeepney, or a durian stall. Those errors affect image captioning, visual question answering, retrieval, and text-to-image generation.
A collaboration involving dozens of Southeast Asian researchers has proposed a framework for measuring these geographic gaps and a lightweight adaptation method for reducing them. The research paper calls the evaluation framework Anthropogenic Regional Adaptation and the adaptation method GG-EZ, short for Geographical-generalization-made-easy. Across three model families, the authors report cultural-relevance gains of 5% to 15% while retaining more than 98% of aggregate global benchmark performance.
Western-heavy data creates regional blind spots
Image-understanding models inherit the geographic distribution of their training corpora. They may replace a specific local object with a broad label, misread clothing or architecture, or retrieve images associated with a better-represented culture. Generative models exhibit a related failure mode by inserting Western aesthetics into prompts about weddings, markets, homes, and landscapes elsewhere.
A single global benchmark can hide these failures because strong results on well-represented content raise the overall average. Anthropogenic Regional Adaptation evaluates a model’s performance on people, objects, languages, and settings associated with a region, then tracks whether localization damages broader capabilities. This gives developers separate regional and global measurements instead of one aggregate score.
GG-EZ adapts a model in three passes
GG-EZ combines data curation, supervised fine-tuning, and linear model merging. Each stage uses established techniques, which keeps the workflow accessible to teams that cannot pretrain a model from scratch.
- Filter regional data. The pipeline selects images and text with a geographic association, then applies UnifiedReward, a model-based quality scorer, using a defined threshold. For the Southeast Asia experiments, the researchers also translate English instruction data into major regional languages.