Bilingual Language Models Capture Cross-Linguistic Influence in Second-Language Reading
Bilingual models better predicted second-language reading when their training language matched readers' first language, a research result with limited immediate agent-building impact.
The study compared bilingual and English-only models against eye-tracking data from readers with **12 first-language backgrounds**. Matching the bilingual model's other language to the reader's L1 improved prediction of English reading times.
For builders, the result suggests that language background can affect behavioral fit even when the visible task language is English. Localization research may need language-pair-specific evaluation rather than one multilingual aggregate.
The study compared bilingual and English-only models against eye-tracking data from readers with **12 first-language backgrounds**. Matching the bilingual model's other language to the reader's L1 improved prediction of English reading times. For builders, the result suggests that language background can affect behavioral fit even when the visible task language is English. Localization research may need language-pair-specific evaluation rather than one multilingual aggregate. The advantage appeared mainly in **late-stage eye-tracking measures**, varied with script overlap, and peaked for **intermediate-proficiency readers**. The material does not establish a direct improvement to coding agents or general product performance.
This adds language-pair and reader-proficiency effects to model evaluation: an English task can still fit users differently depending on the model’s paired language. None of the supplied candidates addresses bilingual behavioral fit, so the result stands independently as a reason to replace aggregate multilingual evaluation with language-pair-specific tests, especially for late reading measures and intermediate-proficiency users.