Cross-Lingual Steering for Figurative Language Generation

Researchers have made new findings in the field of multilingual large language models, discovering they can generate figurative language and creating a dataset to analyze suicide memes on social media.

A recent study found that multilingual large language models are capable of generating figurative language^[1]. Meanwhile, a new dataset called FigSIM has been developed to analyze suicide memes on social media, which are becoming increasingly common. FigSIM consists of 1049 memes annotated for fine-grained suicide severity levels, figurative phenomena, and suicide-related content^[1]. In related research, Kleinberg and Mullainathan proposed a formal framework for language generation in the limit in 2026^[3]. Li, Raman, and Tewari introduced refined notions of non-uniform and uniform generation, while Raman and Raman introduced a noisy model that allows the adversary to insert extraneous strings^[3]. The study found that a single noisy string strictly reduces the set of collections that can be generated, and generation with a single noisy string is equivalent to generation with any finite amount of noise^[3].