Saved in:
Bibliographic Details
Main Authors: Sun, Lihao, Yan, Lewen, Lu, Xiaoya, Lee, Andrew, Zhang, Jie, Shao, Jing
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.03147
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914544587112448
author Sun, Lihao
Yan, Lewen
Lu, Xiaoya
Lee, Andrew
Zhang, Jie
Shao, Jing
author_facet Sun, Lihao
Yan, Lewen
Lu, Xiaoya
Lee, Andrew
Zhang, Jie
Shao, Jing
contents We show that emotion vectors in LLMs are organized by a two-dimensional valence-arousal (VA) subspace exhibiting circular geometry. Through principal component decomposition and ridge regression, we recover meaningful VA axes underlying emotion steering vectors whose projections correlate with human affect ratings across 44,728 words. Steering along these axes produces monotonic control over the affective properties of generated text, and further affords bidirectional control over multiple downstream behaviors (refusal and sycophancy) from a single subspace. These effects replicate across Llama-3.1-8B, Qwen3-8B, and Qwen3-14B. We propose lexical mediation to explain why these effects and prior emotionally framed controls work: refusal and compliance tokens occupy distinct VA regions, and VA steering directly modulates their emission probabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2604_03147
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control
Sun, Lihao
Yan, Lewen
Lu, Xiaoya
Lee, Andrew
Zhang, Jie
Shao, Jing
Computation and Language
Artificial Intelligence
Computers and Society
We show that emotion vectors in LLMs are organized by a two-dimensional valence-arousal (VA) subspace exhibiting circular geometry. Through principal component decomposition and ridge regression, we recover meaningful VA axes underlying emotion steering vectors whose projections correlate with human affect ratings across 44,728 words. Steering along these axes produces monotonic control over the affective properties of generated text, and further affords bidirectional control over multiple downstream behaviors (refusal and sycophancy) from a single subspace. These effects replicate across Llama-3.1-8B, Qwen3-8B, and Qwen3-14B. We propose lexical mediation to explain why these effects and prior emotionally framed controls work: refusal and compliance tokens occupy distinct VA regions, and VA steering directly modulates their emission probabilities.
title Valence-Arousal Subspace in LLMs: Circular Emotion Geometry and Multi-Behavioral Control
topic Computation and Language
Artificial Intelligence
Computers and Society
url https://arxiv.org/abs/2604.03147