Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment

TL;DR

This paper introduces W2S, an adaptive framework for selecting the optimal layer for steering in LLMs based on input, significantly improving alignment performance over fixed-layer methods.

Contribution

The paper proposes an input-dependent layer selection method for steering vectors, addressing the limitation of fixed-layer interventions in LLM alignment.

Findings

W2S outperforms fixed-layer baselines in various settings.

Optimal steering layers vary substantially across inputs.

Input-dependent control enhances LLM alignment effectiveness.

Abstract

Steering vectors have emerged as a lightweight and effective approach for aligning large language models (LLMs) at inference time, enabling modulation over model behaviors by shifting LLM representations towards a target behavior. However, existing methods typically apply steering vectors at a globally fixed layer, implicitly assuming that the optimal intervention layer is invariant across inputs. We argue that this assumption is fundamentally limited, as representations relevant to a target behavior can be encoded at different layers depending on the input. Theoretically, we show that different inputs can require steering at different layers to achieve alignment with a desirable model behavior. We also provide empirical evidence that the optimal steering layer varies substantially across inputs in practice. Motivated by these observations, we introduce Where to Steer (W2S), a framework that adaptively selects the intervention layer conditioned on the input, by learning a mapping from input embeddings to optimal steering layers. Across multiple LLMs and alignment behaviors, W2S consistently outperforms fixed-layer baselines, with improvements in both in-distribution and out-of-distribution settings. Our findings highlight the importance of input-dependent control in LLM alignment and demonstrate that adaptive layer selection is a key design dimension missing in the current methodology of steering vectors.