Steer2Adapt: Dynamically Composing Steering Vectors Elicits Efficient Adaptation of LLMs

TL;DR

STEER2ADAPT introduces a dynamic, compositional approach to activation steering, enabling efficient and flexible adaptation of large language models to diverse tasks by leveraging reusable concept subspaces.

Contribution

It proposes a novel framework that composes steering vectors from a low-dimensional semantic prior, improving adaptability and efficiency over static methods.

Findings

Achieves an average of 8.2% improvement across tasks.

Demonstrates data efficiency and stability in adaptation.

Works effectively in reasoning and safety domains.

Abstract

Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.