Researchers discover reusable cross-lingual signals for figurative language in LLMs

A new paper by Linfeng Liu and colleagues explores whether the internal signals driving figurative language generation in multilingual LLMs are language-specific or reusable across languages. Using activation steering — a technique that shifts model behavior by adjusting internal representations along specific directions — the team estimated a direction for figurative categories (e.g., metaphor, simile) from activation differences between figurative and literal examples in one language, then applied it during generation in another language. Across five figurative categories, six languages (including German, English, Chinese), and four multilingual LLMs, they found that steering directions work reliably within their own language (most robust for metaphor and simile). Crucially, these directions transfer across languages: a direction learned in one language increases target figurative behavior when applied to another, with German emerging as the most receptive target.

Going further, the researchers showed that directions assembled from multiple other languages can match or even surpass a target language's own native steering direction. Conversely, removing this shared cross-lingual component weakens native steering. These results provide direct evidence of a reusable but target-dependent cross-lingual signal for figurative generation, suggesting that multilingual LLMs share underlying representations for figurative language that can be manipulated across languages. This could enable more efficient multilingual text generation and stylistic control without needing separate training data for each language.