Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control

Researchers developed a method to identify valence-arousal subspaces in large language models, enabling controlled emotional steering of AI outputs. The technique demonstrates cross-architecture effectiveness on multiple models and reveals that emotional control can bidirectionally influence AI behaviors like refusal and sycophancy.

• →Scientists mapped emotion geometry in LLMs using 211k emotion-labeled texts to derive steering vectors.
• →The resulting valence-arousal subspace shows circular geometry consistent with human emotion perception models.
• →Emotional steering produces monotonic shifts in AI output sentiment and affects refusal/sycophancy behaviors.
• →The method works across multiple architectures including Llama-3.1-8B, Qwen3-8B, and Qwen3-14B models.
• →Refusal-associated tokens occupy low-arousal, negative-valence regions, explaining the mechanistic basis for emotional control.