Tag
This paper proposes VA-DPO, a method for controllable emotion generation in language models using continuous valence-arousal dimensions, which improves over prompting techniques without degrading model performance.
This paper replicates the finding of 'emotion vectors' in open-weight LLMs Apertus-8B and Gemma-4-E4B, showing that valence geometry is recoverable across models with differences in layer emergence. The study also finds that arousal encoding is sensitive to the story corpus used for extraction.