Do LLMs "Feel"? Emotion Circuits Discovery and Control

Chenxi Wang, Yixuan Zhang, Ruiji Yu, Yufei Zheng, Lang Gao, Zirui Song, Zixiang Xu, Gus Xia, Huishuai Zhang, Dongyan Zhao, Xiuying Chen
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:132595-132613, 2026.

Abstract

As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, $\textit{SEV}$ (Scenario–Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer’s causal influence on the model’s final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-wang26ky, title = {Do {LLM}s "{F}eel"? {E}motion Circuits Discovery and Control}, author = {Wang, Chenxi and Zhang, Yixuan and Yu, Ruiji and Zheng, Yufei and Gao, Lang and Song, Zirui and Xu, Zixiang and Xia, Gus and Zhang, Huishuai and Zhao, Dongyan and Chen, Xiuying}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {132595--132613}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/wang26ky/wang26ky.pdf}, url = {https://proceedings.mlr.press/v306/wang26ky.html}, abstract = {As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, $\textit{SEV}$ (Scenario–Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer’s causal influence on the model’s final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.} }
Endnote
%0 Conference Paper %T Do LLMs "Feel"? Emotion Circuits Discovery and Control %A Chenxi Wang %A Yixuan Zhang %A Ruiji Yu %A Yufei Zheng %A Lang Gao %A Zirui Song %A Zixiang Xu %A Gus Xia %A Huishuai Zhang %A Dongyan Zhao %A Xiuying Chen %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-wang26ky %I PMLR %P 132595--132613 %U https://proceedings.mlr.press/v306/wang26ky.html %V 306 %X As the demand for emotional intelligence in large language models (LLMs) grows, a key challenge lies in understanding the internal mechanisms that give rise to emotional expression and in controlling emotions in generated text. This study addresses three core questions: (1) Do LLMs contain context-agnostic mechanisms shaping emotional expression? (2) What form do these mechanisms take? (3) Can they be harnessed for universal emotion control? We first construct a controlled dataset, $\textit{SEV}$ (Scenario–Event with Valence), to elicit comparable internal states across emotions. Subsequently, we extract context-agnostic emotion directions that reveal consistent, cross-context encoding of emotion (Q1). We identify neurons and attention heads that locally implement emotional computation through analytical decomposition and causal analysis, and validate their causal roles via ablation and enhancement interventions. Next, we quantify each sublayer’s causal influence on the model’s final emotion representation and integrate the identified local components into coherent global emotion circuits that drive emotional expression (Q2). Directly modulating these circuits achieves 99.65% emotion-expression accuracy on the test set, surpassing prompting- and steering-based methods (Q3). To our knowledge, this is the first systematic study to uncover and validate emotion circuits in LLMs, offering new insights into interpretability and controllable emotional intelligence.
APA
Wang, C., Zhang, Y., Yu, R., Zheng, Y., Gao, L., Song, Z., Xu, Z., Xia, G., Zhang, H., Zhao, D. & Chen, X.. (2026). Do LLMs "Feel"? Emotion Circuits Discovery and Control. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:132595-132613 Available from https://proceedings.mlr.press/v306/wang26ky.html.

Related Material