When answering questions posed in Chinese, French, German or Russian, large language models seem to process the queries in English, which could create cultural issues
I just quickly skimmed the source paper, but I’m wondering if they can distinguish between a latent “concept” space that’s actually based on English, versus one that’s derived from all source languages and just weighted more towards English?
I just quickly skimmed the source paper, but I’m wondering if they can distinguish between a latent “concept” space that’s actually based on English, versus one that’s derived from all source languages and just weighted more towards English?