Yusuke Sakai

The paper proposes a method to identify 'hub texts' that exploit vulnerabilities in cross-modal encoders, demonstrating that a single text can achieve unrealistically high similarity scores across diverse images in tasks like image captioning and retrieval.

This paper analyzes the multilinguality of LLMs by examining their structural properties, finding that low-resource languages are structurally more distinct from English than high-resource languages, and that language-specific fine-tuning alters structure while maintaining inter-language relationships.