Summary
This study looks at how large language models can be guided to change their language style or behavior by adding special instructions called steering vectors. The authors found that these steering vectors work best when applied at specific layers inside the model, depending on the behavior they want to control, like language choice, being concise, or avoiding unwanted outputs. They also showed that adding two or three of these vectors together can combine different behaviors effectively if each is put at the right layer. Finally, they discovered that the steering vectors are roughly at right angles to each other inside the model, which helps them keep their separate effects when combined.
large language modelssteering vectorsinstruction tuningmodel layersbehavior controladditive compositionjailbreak behaviorconcisenessresidual streamgeometric orthogonality
Authors
Hyun Gu Kang, Daniil Gurgurov, Tanja Baeumel, Josef van Genabith, Simon Ostermann
Abstract
This study examines the compositionality of steering vectors for language and behavioral control in large language models. Focusing on language, jailbreak, and conciseness, we investigate whether additive, training-free composition of attribute steering vectors can preserve the intended steering effect of each attribute, across four instruction-tuned models from two model families and two size scales. We find that single-attribute steering is reliable for all three attributes, but only within an appropriate combination of intervention layer and steering strength, with abstract behaviors (jailbreak, conciseness) favoring middle layers and language favoring earlier layers. We show that additive composition of two attribute vectors succeeds in steering both attributes simultaneously when each is injected at its own best-performing layer, and that this partially extends to three simultaneously composed attributes, addressing an inconsistency left open by prior work on training-free composition. We further analyze the geometric properties of these steering vectors, finding that they are approximately orthogonal in the residual stream, consistent with their compositional behavior.