From Rational Answers to Emotional Resonance: The Role of Controllable Emotion Generation in Language Models (2502.04075v2)

Abstract: Purpose: Emotion is a fundamental component of human communication, shaping understanding, trust, and engagement across domains such as education, healthcare, and mental health. While 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. Methods: 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 theoretical analysis proving that EV steering enhances emotional expressivity while maintaining semantic fidelity and linguistic fluency. Results: Extensive experiments across multiple LLM families show that the proposed approach achieves consistent emotional alignment, stable topic adherence, and controllable affect intensity. Compared with existing prompt-based and fine-tuning-based baselines, our method demonstrates superior flexibility and generalizability. Conclusion: Emotion Vector (EV) steering provides an efficient and interpretable means of bridging rational reasoning and affective understanding in LLMs, offering a promising direction for building emotionally resonant AI systems capable of more natural human-machine interaction.

• The paper presents a novel Emotion Vector (EV) pipeline that enables LLMs to produce controlled emotional responses without retraining.
• Experimental results show that EV injection enhances emotional expressivity with metrics like Target Emotion Confidence while maintaining semantic accuracy.
• Visual analyses and cosine distance distributions confirm robust emotional clustering, indicating practical applications in education, healthcare, and mental health support.

From Rational Answers to Emotional Resonance: The Role of Controllable Emotion Generation in LLMs

Introduction

The paper "From Rational Answers to Emotional Resonance" explores the capability of LLMs to generate controllable emotional responses, moving beyond mere rationality and towards affective resonance in human-machine interactions. The challenge addressed is the inherent limitation of LLMs in producing consistent, contextually appropriate emotional expressions, which hinders their utility in domains such as education, healthcare, and mental health where affective communication is crucial.

Methodology

The methodology centers on the creation and utilization of Emotion Vectors (EVs), which are derived as latent shifts in internal model activations between neutral and emotion-conditioned prompts. This approach allows for emotional modulation of LLM outputs by injecting EVs into the hidden states during inference without necessitating further model training or changes to architecture. The EV pipeline is a two-stage process involving EV extraction and steering during text generation (Figure 1).

Figure 1: Overview of the Emotion Vector (EV) pipeline.

Emotion Vectors are computed at each transformer block by taking the difference in token-averaged hidden states between emotion-conditioned and neutral prompts. These vectors are then averaged across multiple queries to stabilize the representation of specific emotions.

Theoretical Insights

The paper provides a theoretical basis asserting that steering LLMs with EVs enhances emotional expressivity while maintaining semantic integrity. Through detailed mathematical analysis, it is demonstrated that EV injection can increase emotional reading without distorting the semantic content, supported by a rigorous first-order linear expansion of the latent space mappings.

Experimental Results

Extensive experimental evaluations show that EV-steered models achieve stable emotional alignment with intended affective states across various LLM architectures and sizes. Performance is measured using metrics such as Target Emotion Confidence (TEC) and Emotion Probability Scores, which highlight the capability of EVs to consistently modulate emotional intensity across different prompts and models without sacrificing linguistic fluency.

Figure 2: Target Emotion Confidence (TEC) scores across different EV intensities for each target emotion.

Visualization and Analysis

Visualization techniques such as t-SNE are employed to map EVs, evidencing clear clustering of emotions across different layers of the model. This clustering indicates the robust performance of EVs in distinguishing between different emotional states despite variations in the input queries (Figures 3 and 5).

Figure 3: t-SNE plots of Emotion Vectors from different layers.

Figure 4: A t-SNE plot of Emotion Vectors, each forming distinct clusters.

Furthermore, the analysis of cosine distance distributions reinforces the ability of EVs to maintain separations between different emotional states by showing distinct within-class and between-class distance distributions (Figure 5).

Figure 5: Histograms of cosine distance distributions for each emotion.

Implications and Future Directions

The findings indicate that EVs provide a flexible and efficient solution for integrating emotional expressivity in LLMs, offering potential enhancements in areas requiring delicate emotional interactions such as personalized tutoring systems, empathetic healthcare bots, and mental health support applications. These results suggest that future research could further explore the integration of EVs with multimodal systems and investigate their applicability in real-time interactive settings.

Conclusion

The paper provides a comprehensive framework for enhancing LLMs with controllable emotional capabilities via the innovative use of Emotion Vectors. These contributions set the stage for building more nuanced and context-aware AI systems capable of authentic human-AI interactions, a crucial advancement for deploying AI in socially significant domains.