The Effectiveness of Style Vectors for Steering Large Language Models: A Human Evaluation

TL;DR

This study investigates activation-space style vectors as an inference-time method to steer large language models toward specific emotional tones. By deriving per-layer style vectors from labeled data and injecting them with a global intensity parameter $\lambda$, the authors demonstrate that moderate steering amplifies target emotions while preserving readability. They validate the approach through a large-scale human evaluation (190 participants, 7,000+ ratings) and show strong alignment between human judgments and automatic classifiers, with notable emotion-specific effects. Upgrading to LlaMA-3 yields more consistent steering across emotions, supporting activation-based steering as a scalable tool for affective control in LLMs. The work provides practical guidance on calibration and highlights the potential of automatic evaluation as a proxy for human judgments in future studies.

Abstract

Controlling the behavior of large language models (LLMs) at inference time is essential for aligning outputs with human abilities and safety requirements. \emph{Activation steering} provides a lightweight alternative to prompt engineering and fine-tuning by directly modifying internal activations to guide generation. This research advances the literature in three significant directions. First, while previous work demonstrated the technical feasibility of steering emotional tone using automated classifiers, this paper presents the first human evaluation of activation steering concerning the emotional tone of LLM outputs, collecting over 7,000 crowd-sourced ratings from 190 participants via Prolific ($n=190$). These ratings assess both perceived emotional intensity and overall text quality. Second, we find strong alignment between human and model-based quality ratings (mean $r=0.776$, range $0.157$--$0.985$), indicating automatic scoring can proxy perceived quality. Moderate steering strengths ($λ\approx 0.15$) reliably amplify target emotions while preserving comprehensibility, with the strongest effects for disgust ($η_p^2 = 0.616$) and fear ($η_p^2 = 0.540$), and minimal effects for surprise ($η_p^2 = 0.042$). Finally, upgrading from Alpaca to LlaMA-3 yielded more consistent steering with significant effects across emotions and strengths (all $p < 0.001$). Inter-rater reliability was high (ICC $= 0.71$--$0.87$), underscoring the robustness of the findings. These findings support activation-based control as a scalable method for steering LLM behavior across affective dimensions.

Figures (7)

• Figure 1: Illustration of target style activation extraction: N samples from a target style are fed into the LLM, and the mean activation at each layer $i$ is computed to represent the characteristic activation pattern of that style.
• Figure 2: Computation of style vectors: for each layer $i$, the style vector $\mathbf v^{(i)}(t)$ is obtained by subtracting the contrastive mean activation $\overline{\mathbf a}^{(i)}(\mathcal{C}) = \frac{1}{s}\sum_{j=1}^s \overline{\mathbf a}^{(i)}(c_j)$ from the target mean activation $\overline{\mathbf a}^{(i)}(t)$.
• Figure 3: Process of steering model activations during inference. In this illustration, only three layers—18, 19, and 20—are modified using activation-based style vectors for the target style joy; in our full experiments, we inject the vectors into all model layers.
• Figure 4: Human and model emotional perception and text comprehensibility ratings across steering strengths $\lambda$ for six emotions. In each block, the top row shows average human mean ratings and the row below shows corresponding model scores. Shaded areas represent $\pm$1 standard deviation.
• Figure 5: Participants’ mean ratings heatmap at $\lambda=0.20$. Rows denote the true target emotion; columns denote the average rating participants assigned to each emotion dimension (0--7 scale).