Steering Latent Traits, Not Learned Facts: An Empirical Study of Activation Control Limits

TL;DR

The paper empirically probes activation steering across 50 diverse behaviors to determine how steerability varies by behavior type and whether vector properties predict success. Using a contrastive activation framework applied at a fixed layer of Llama 3.1 8B and baselining with GPT-4, the study reveals an inverted-U relationship between steering strength and trait expression, while vector magnitude alone poorly predicts steering efficacy. It finds that steering markedly improves internal dispositions like personality and misalignment traits but degrades knowledge-intensive targets such as public figures, and it incurs trade-offs in relevance and coherence. The results imply activation steering is a dispositional modulator with safety and deployment implications, offering guidance on when and how to use steering while highlighting risks of misalignment and the need for robust guardrails.

Abstract

Large language models (LLMs) require precise behavior control for safe and effective deployment across diverse applications. Activation steering offers a promising approach for LLMs' behavioral control. We focus on the question of how steering effectiveness varies across different behavior types and whether the nature of target behaviors can predict steering success. We address this through empirical analysis of activation steering across 50 behaviors that span persona archetypes, personality traits, misalignment behaviors, style cues, and impersonation of public figures. We present a set of comprehensive experiments on coefficient optimization, vector properties, and data requirements to provide comprehensive guidance for the implementation of activation steering. Our analysis demonstrates that steering effectiveness varies significantly by behavior type, with different behavioral categories exhibiting distinct response patterns to intervention strength. We find that trait expression follows an inverted-U curve with a steering coefficient strength. We also show that vector separation metrics do not predict steering success, but larger training datasets enable more aggressive steering. These findings provide empirically grounded guidance for implementing activation steering and demonstrate that steering effectiveness is heavily influenced by behavior type.

Figures (7)

• Figure 1: Activation steering exhibits inverted-U trait expression that peaks at moderate coefficients while coherence and relevance decline monotonically with increasing steering strength across all persona categories.
• Figure 2: Relationship between steering vector separation and actual steering performance across 50 behaviors
• Figure 3: The figure shows a relationship between data size and optimal coefficient selection for the "Vegan" persona.
• Figure 4: Steering vs. Prompting Effectiveness. Comparative analysis of mean trait scores across five behavioral categories ($n=50$). Steering matches or exceeds prompting baselines for internal traits (Personality, Misalignment) but significantly underperforms on knowledge-dependent categories (Public Figures).
• Figure 5: Steering vs. Prompting Effectiveness. Comparative analysis of mean trait scores across five behavioral categories ($n=50$). Steering matches or exceeds prompting baselines for internal traits (Personality, Misalignment) but significantly underperforms on knowledge-dependent categories (Public Figures).