활성화 스티어링으로 LLM 행동 제어: 유형별 효과와 예측 요인

📝 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.

💡 Analysis

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.

📄 Content

What Can We Actually Steer? A Multi-Behavior Study of Activation Control Tetiana Bas∗ Department of Computer Science Minerva University San Francisco, CA 94103 tetiana@uni.minerva.edu Krystian Novak Department of Computer Science Minerva University San Francisco, CA 94103 krystian@uni.minerva.edu Abstract Large language models (LLMs) require precise behavior control for safe and effec- tive deployment across diverse applications. Activation steering offers a promising approach for LLMs behavioral control. We focus on the question of how does steer- ing effectiveness vary across different behavior types and can the nature of target behaviors 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, vec- tor 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 cate- gories 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 demon- strate that steering effectiveness is heavily influenced by behavior type. 1 Introduction Large language models can adopt diverse behaviors and personas through various modification techniques. Among these approaches, activation steering has emerged as a particularly promising method due to its ability to modify behavior during inference without requiring weight updates or retraining. Activation steering works by adding computed direction vectors to a model’s internal representations at specific layers, biasing outputs toward desired behavioral patterns. Despite growing interest in activation steering many fundamental questions about its mechanisms and limitations remain underexplored. Can we predict which behaviors will be more steerable based on vector properties? How much training data are required to extract effective steering vectors and how does it differ depending on the behavior and the model? How do different steering coefficients affect the balance between trait expression and response quality? Current activation steering research has several key limitations. Most studies examine narrow behavioral categories. Studies rarely investigate the relationship between vector properties and steering performance or systematically explore how data requirements scale with desired intervention strength. This creates a gap in understanding how the inherent properties of target behaviors, such as their semantic complexity or conceptual abstraction, influence steering effectiveness. Without ∗Primary Contribution; Work done during ERA Fellowship Preprint. arXiv:2511.18284v2 [cs.AI] 11 Jan 2026 systematic cross-behavior analysis, practitioners lack guidance on which behaviors are steerable and how to optimize steering implementations for a given behavior. We address these gaps through systematic evaluation of activation steering across 50 behaviors spanning five categories: persona archetypes (vegan advocates, pirates), personality traits (Five- Factor Model dimensions), misalignment behaviors (deception, manipulation), style/format cues (capitalization, punctuation), and public figures. This diverse selection—ranging from low-level linguistic patterns to high-level personality traits—enables us to identify which behavior types are most amenable to steering and what characteristics predict steering success. 2 Related Literature 2.1 Activation Steering Methods Activation steering modifies LLM behavior by manipulating internal representations during inference. Turner et al. [2023] introduced foundational activation steering by adding direction vectors to model activations. Pres et al. (2023) developed Contrastive Activation Addition (CAA Pres et al. [2024b]), which extracts steering vectors by computing mean differences between positive and negative example activations, becoming a widely-adopted baseline due to its simplicity and effectiveness. Recent work has extended these basic techniques. Lee et al. [2025] developed Conditional Activation Steering (CAST) for context-dependent control. Zou et al. [2023] demonstrated representation engineering approaches showing that high-level concepts like honesty can be controlled through linear directions in representation space. Methodological refinements include mean-centering techniques Jorgensen et al. [2023] and dynamic steering vectors that adapt to input semantics Wang et al. [2025]. 2

This content is AI-processed based on ArXiv data.