Cross-lingual Functional Vectors for Emotion Detection in Large Language Models
Function vectors transferred emotion-recognition behavior across languages without inference-time demonstrations.
The paper tests multilingual, multi-label emotion detection as a harder benchmark for FV steering. FVs extracted from one source language improved performance in another language in both clean and perturbed zero-shot settings. The authors argue the gains suggest the vectors carry task-relevant signals beyond language-specific lexical patterns. They also report stable attention-head ranges for effective FV construction within each LLM, consistent across languages. ArXiv · AI/CL/LG's note
The paper tests multilingual, multi-label emotion detection as a harder benchmark for FV steering. FVs extracted from one source language improved performance in another language in both clean and perturbed zero-shot settings. The authors argue the gains suggest the vectors carry task-relevant signals beyond language-specific lexical patterns. They also report stable attention-head ranges for effective FV construction within each LLM, consistent across languages. ArXiv · AI/CL/LG's note
score 3