Cross-Lingual Steering for Figurative Language Generation

2026-05-28

AI Summary

This study investigates whether the internal representational signals driving figurative language generation in multilingual large language models exhibit cross-lingual universality. Leveraging activation steering, the authors extract directional vectors from the activation differences between figurative and literal expressions in one language and apply them to steer generation in other languages. The work provides the first direct evidence of reusable yet target-language-dependent cross-lingual signals for figurative language production. Notably, combining steering directions derived from multiple source languages consistently outperforms those from any single native language. The experiments span six languages, five figurative language types, and four state-of-the-art multilingual models, with German showing the highest sensitivity to activation steering.

Abstract

Multilingual large language models can generate figurative language, but whether the internal signals driving this behavior are language-specific or reusable across languages is unclear. Using activation steering as a probe, we estimate a direction for a figurative category from figurative--literal activation differences in one language and apply it during generation. Across five figurative categories, six languages, and four multilingual LLMs, these directions steer reliably within their own language, most robustly for metaphor and simile. More importantly, they transfer across languages: a direction learned in one increases the target behavior when applied to another, with German among the most receptive targets. Going further, directions assembled from other languages can match or even surpass a target language's own native direction, while removing this shared component weakens native steering. Together, these results provide direct evidence of a reusable but target-dependent cross-lingual signal for figurative generation.

Problem

• cross-lingual
• figurative language
• activation steering
• multilingual LLMs
• language transfer

Innovation

• cross-lingual steering
• figurative language generation
• activation steering
• multilingual LLMs
• transferable representations