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

Technical Summary: The Effectiveness of Style Vectors for Steering Large Language Models

Problem Statement
Controlling the behavior of Large Language Models (LLMs) at inference time is critical for aligning outputs with human safety requirements and communication goals. While prompt engineering and fine-tuning are established methods, they often lack precise stylistic control or require significant computational resources. Activation steering offers a lightweight alternative by directly modifying internal model activations to guide generation. However, prior research has primarily relied on automated classifiers or illustrative examples to demonstrate feasibility. A significant gap exists regarding whether humans can actually perceive nuanced steering in LLM outputs, particularly concerning emotional tone, and how such steering impacts text quality and comprehensibility.

Methodology
The study employs an activation-based style steering approach adapted from Konen et al. [2], utilizing the LlaMA-3-8B-Lexi-Uncensored model. The methodology consists of three primary phases:

1. Vector Construction: Style vectors are derived from the GoEmotions dataset (mapped to Ekman's six basic emotions: anger, disgust, fear, joy, sadness, surprise). For a target style $t$ and contrastive styles $c$, the mean activation vectors $\bar{a}^{(i)}$ are computed for each layer $i$. The style vector $v^{(i)}(t)$ is calculated as the difference between the target mean activation and the mean of contrastive activations:
   $$v^{(i)}(t) = \bar{a}^{(i)}(t) - \frac{1}{s}\sum_{j=1}^{s} \bar{a}^{(i)}(c_j)$$
2. Inference-Time Steering: During generation, the original activations $a^{(i)}(x)$ are modified by adding the scaled style vector:
   $$\hat{a}^{(i)}(x) = a^{(i)}(x) + \lambda \cdot v^{(i)}(t)$$
   Here, $\lambda$ acts as a scaling parameter controlling the intensity of the steering. The study applies these vectors across all model layers.
3. Evaluation:
   • Automatic Evaluation: A pre-trained DistilRoBERTa model classifies emotion intensity, and the LlaMA-3 model itself rates text comprehensibility on a scale of 1–10.
   • Human Evaluation: A within-subjects study ($N=190$) recruited via Prolific. Participants rated 7,000+ generated texts on perceived emotional intensity (0–7 scale) and text comprehensibility (0–7 scale) across 19 prompts, 6 emotions, and 8 steering strengths ($\lambda \in [0.00, 0.35]$).

Key Contributions
The paper advances the literature in three specific directions:

1. First Human Evaluation: It provides the first large-scale human evaluation of activation steering for emotional tone, moving beyond automated metrics to assess perceived intensity and text quality.
2. Steering Strength vs. Quality Analysis: It analyzes the trade-off between steering intensity and output quality, identifying that moderate strengths preserve fluency while higher strengths degrade coherence.
3. Model Architecture Comparison: It demonstrates that upgrading from Alpaca (in prior work) to LlaMA-3 yields more consistent and pronounced steering effects across emotions and strengths.

Results

• Emotional Intensity: Human ratings confirmed a clear increase in perceived emotional intensity for five of the six emotions (anger, disgust, fear, joy, sadness) as $\lambda$ increased. The strongest effects were observed for disgust ($\eta^2_p = 0.616$) and fear ($\eta^2_p = 0.540$). Surprise showed minimal steerability ($\eta^2_p = 0.042$), with no meaningful change in perceived intensity.
• Human-Model Alignment: There was a strong alignment between human and model-based quality ratings (mean Pearson $r = 0.776$). While human raters were more conservative in their upper-bound ratings, the trends in emotional amplification and the decline in comprehensibility at high $\lambda$ values were closely mirrored by the automatic classifiers.
• Text Quality: Comprehensibility remained stable for steering strengths up to $\lambda \approx 0.15$. Beyond this threshold, text quality (fluency and coherence) began to degrade significantly. The study notes that human raters also perceived "co-perception" of emotions (e.g., steering for anger also increased perceived disgust), reflecting the multidimensional nature of human emotion perception, which the single-dimension model classifiers did not fully capture.
• Reliability: Inter-rater reliability was high for aggregated judgments (ICC $\approx 0.71–0.87$) for most emotions, though individual rater consistency varied.

Significance
The paper claims that activation-based control is a scalable and effective method for steering LLM behavior across affective dimensions. It validates that:

• Activation steering can successfully amplify target emotions in a controllable manner without retraining.
• Moderate steering strengths ($\lambda \approx 0.15$) offer a viable balance between emotional expressiveness and text comprehensibility.
• Automatic scoring can serve as a reliable proxy for human-perceived quality, potentially reducing the cost of future evaluations.
• The technique is particularly effective for LlaMA-3 architectures, suggesting that model capability influences the efficacy of activation engineering.

The authors conclude that while the method is effective, it requires emotion-specific calibration to manage the trade-off between intensity and coherence, and future work should explore generalizability to multi-turn dialogues and more abstract stylistic attributes.