Affective directions in activation space: An interpretable inference-time control mechanism

Emotion is a fundamental component of human communication, shaping understanding, trust, and engagement across domains such as education, healthcare, and mental health. Although large language models (LLMs) exhibit strong reasoning and knowledge-generation capabilities, they still struggle to express emotions in a consistent, controllable, and contextually appropriate manner. This limitation restricts their potential for authentic human–AI interaction. We propose a controllable emotion generation framework based on Emotion Vectors (EVs) —latent representations derived from internal activation shifts between neutral and emotion-conditioned responses. By injecting these vectors into the hidden states of pretrained LLMs during inference, our method enables fine-grained, continuous modulation of emotional tone without any additional training or architectural modification. We further provide a first-order theoretical analysis of emotional steering and its interaction with a semantic readout under explicit local assumptions. Experiments across multiple LLM families demonstrate adjustable emotional expression and characterize the model-dependent trade-off between steering strength, topical relevance, and fluency. The framework provides a general, training-free control interface with explicit intensity scaling and vector composition. Comparisons with prompting and fine-tuning characterize the emotional and quality effects of the evaluated configurations. Emotion Vector (EV) steering thus offers an efficient and interpretable means of bridging rational reasoning and affective understanding in large language models, and points to a promising direction for building emotionally resonant AI systems capable of more natural human–machine interaction.