Emotions Where Art Thou: Understanding and Characterizing the Emotional Latent Space of Large Language Models

This work investigates how large language models (LLMs) internally represent emotion by analyzing the geometry of their hidden-state space. The paper identifies a low-dimensional emotional manifold and shows that emotional representations are directionally encoded, distributed across layers, and aligned with interpretable dimensions. These structures are stable across depth and generalize to eight realworld emotion datasets spanning five languages. Cross-domain alignment yields low error and strong linear probe performance, indicating a universal emotional subspace. Within this space, internal emotion perception can be steered while preserving semantics using a learned intervention module, with especially strong control for basic emotions across languages. These findings reveal a consistent and manipulable affective geometry in LLMs and offer insight into how they internalize and process emotion. Recent work has also examined emotion manipulation and decoding. For instance, models have been used to map text to dimensional emotion ratings like valence-arousal-dominance (VAD) (Shah et al., 2023; Broekens et al., 2023) , or to generate emotionally inflected language on demand (Reichman et al., 2025) . LLMs have also been shown to be more likely to comply with emotionally framed requests (Vinay et al., 2024) . These studies also treat emotion primarily as a label or generation condition-not a latent internal representation.