Do 'emotional neurons' exist in large language models? An ACL 2025 Findings study reveals the neural mechanisms of emotion processing in LLMs

A South Korean research team published a study in ACL 2025 Findings, which for the first time systematically confirmed the existence of neuron groups dedicated to processing specific emotions in models like Llama-3.1. The distribution of these 'emotional neurons' varies with model size and architectural depth, and different emotions show significant differences in sensitivity to neuron removal. This study provides key insights into the interpretability and emotion processing mechanisms of LLMs.

LLMs perform well in sentiment analysis tasks, but core questions remain: Do models truly 'understand' emotions or just imitate statistically? Are there neural mechanisms dedicated to processing specific emotions? This study builds an experimental framework based on Paul Ekman's theory of six basic emotions (joy, sadness, anger, fear, disgust, surprise) to explore the internal emotion processing mechanisms of LLMs.

The core foundation of the study is the EmoPrism large-scale synthetic dialogue dataset, which contains 293,725 single-emotion annotated dialogues. The construction process is: 315 sub-topics expanded to 5040 topics → 302,400 dialogues synthesized → labeled by three models + majority voting to determine tags. Synthetic data can precisely control emotion distribution and avoid the problem of contamination in real data, and has been open-sourced (CC-BY-4.0).

Identification strategy: Count the activation times of FFN neurons on each emotion token and calculate the entropy value. Neurons with low entropy values respond selectively to specific emotions; the top 1% neurons with the lowest global entropy are defined as emotional neurons and assigned to the emotion they activate most strongly. Experiments were conducted on the Llama-3.1-Instruct 8B/70B models, and it was found that there are emotion-specific clusters whose distribution varies with model size and number of layers.

Function verification via 'zero ablation' experiments: Set the output of emotional neurons to zero and measure changes in emotion classification accuracy. Results show: Anger and fear are sensitive to neuron removal (accuracy drops sharply); joy and surprise show little change or improvement in performance (possibly due to neuron overlap compensation). This indicates that processing strategies for different emotions are not uniform.

Emotional information runs through multiple layers of the model: Shallow layers are responsible for basic emotion feature extraction, while deep layers participate in complex emotion understanding and context integration. There are significant differences in model size: The 8B and 70B models differ in the distribution and function of emotional neurons, suggesting that model expansion may bring qualitative changes in emotion processing capabilities.

This study opens a new direction for LLM interpretability; confirming the existence of emotional neurons provides intervention targets for model editing and safety alignment. In applications, precise regulation of emotional output can be achieved by adjusting specific neurons, helping to develop safer and more controllable AI assistants. The dataset and code have been open-sourced to support reproducibility and expansion of subsequent research.

The ACL 2025 Findings study, through rigorous experiments, for the first time confirms the existence of emotional neurons in mainstream LLMs, deepens the understanding of LLM emotion processing mechanisms, and provides important implications for developing more interpretable and controllable AI systems. Participation from the open-source community is expected to drive more breakthroughs in this field.