Adobe Research Proposes Locale-Aware Boosting to Fix Ranking Bias in Global Markets

A recent preprint (2605.11272) from Adobe Research explores the issue of cross-locale behavioral bias in learning-to-rank (LTR) models utilized by Adobe Express. The predominance of the US in content generation and interaction results in LTR models, which rely on behavioral feedback, favoring templates popular in the US at the expense of local content visibility in other regions. The researchers found that training based solely on clicks does not adequately capture localization features that are semantically relevant. By incorporating vision-language model (VLM) graded relevance labels for additional supervision, semantic alignment improves, though local content visibility remains compromised. They introduce a multi-objective framework that integrates behavioral supervision, VLM relevance signals, and locale-aware boosting, enhancing relevance across five locales while maintaining stable localization and emphasizing the need to separate exposure bias from relevance optimization.