Emergence of Hierarchical Emotion Representations in Large Language Models
Keywords: LLM, emotion
Abstract: As large language models (LLMs) increasingly power emotionally engaging conversational agents, understanding how they represent, predict, and potentially influence human emotions is critical for their ethical deployment in sensitive contexts. In this work, we reveal emergent hierarchical structures in LLMs' emotion representations, drawing inspiration from psychological theories of emotion. By analyzing probabilistic dependencies between emotional states in LLM outputs, we propose a method for extracting these hierarchies. Our results show that larger models, such as LLaMA 3.1 (405B parameters), develop more intricate emotion hierarchies, resembling human emotional differentiation from broad categories to finer states. Moreover, we find that stronger emotional modeling enhances persuasive abilities in synthetic negotiation tasks, with LLMs that more accurately predict counterparts' emotions achieving better outcomes. Additionally, we explore the effects of persona biases—such as gender and socioeconomic status—on emotion recognition, revealing that LLMs can misclassify emotions when processing minority personas, thus exposing underlying biases. This study contributes to both the scientific understanding of how LLMs represent emotions and the ethical challenges they pose, proposing a novel interdisciplinary perspective on the issue.
Supplementary Material: zip
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10257
Loading