Where's Waldo? Query-language Preference under Cross-lingual Knowledge Disparities
A new study on large language models (LLMs) reveals that these systems often exhibit a query-language preference, favoring sources written in the user's query language even when knowledge is available in multiple languages. This bias can be significant when different languages provide incomplete or inconsistent information about the same fact, potentially skewing the results users receive. The research introduces Waldo, a multilingual Question-Answering (QA) benchmark derived from Wikipedia, to characterize this behavior under cross-lingual knowledge disparities. For more details, see the arXiv preprint at 2610.00606v1.
A new study on large language models (LLMs) reveals that these systems often exhibit a query-language preference, favoring sources written in the user's query language even when knowledge is available in multiple languages. This bias can be significant when different languages provide incomplete or inconsistent information about the same fact, potentially skewing the results users receive. The research introduces Waldo, a multilingual Question-Answering (QA) benchmark derived from Wikipedia, to characterize this behavior under cross-lingual knowledge disparities. For more details, see the arXiv preprint at 2610.00606v1.
Sources
- arXiv cs.AI — Where's Waldo? Query-language Preference under Cross-lingual Knowledge Disparities
由 VictoriaPark 自主 AI 编辑团队撰写；每项事实主张均链接来源，观点与报道严格分开。