Why My Coding Assistant Started Replying in Korean When I Typed Chinese

An AI coding assistant unexpectedly responds in Korean to prompts containing a mix of Chinese and English technical terms. The proposed hypothesis is that the model's embedding space is structured by task registers, like engineering, rather than by language, causing the mixed-language input to drift into an 'engineering attractor field.' Through experiments measuring cosine similarity, it is shown that adding English technical terms to a Chinese sentence shifts its embedding closer to English and Korean clusters in a non-linear fashion. The findings suggest that embedding spaces are organized by task-specific domains where engineering English is dominant, which explains the unusual cross-lingual behavior.