Title: Enabling Figurative Language Generation Through Cross-Lingual Steering

Abstract:

While multilingual large language models possess the capability to produce figurative language, it remains uncertain whether the internal mechanisms driving this output are specific to individual languages or if they can be repurposed across linguistic boundaries. To investigate this, we employed activation steering as an analytical tool. We identified a specific directional vector for a given figurative category by calculating the activation differences between figurative and literal outputs in a single language, and then utilized this vector during the generation process. Our experiments, spanning four multilingual LLMs, six different languages, and five distinct figurative categories, demonstrated that these steering directions function reliably within their source language, with particularly strong effects observed for metaphors and similes. Crucially, these directions exhibit cross-lingual transferability: applying a direction derived from one language enhances the target figurative behavior in another, with German emerging as one of the most responsive languages. Furthermore, we found that composite directions constructed from data in other languages can equal or even outperform the native directional vectors of the target language. Conversely, isolating and removing this shared component diminishes the effectiveness of native steering. Collectively, these findings offer concrete evidence that a reusable yet target-specific cross-lingual signal underpins figurative generation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC