Compositional Multilingual and Behavioral Attribute Steering
The paper studies whether simple additive steering vectors can simultaneously control language choice and behaviors like jailbreakiness and conciseness in instruction-tuned LLMs, finding that each attribute works reliably only when applied at suitable layers and strengths. Its nice contribution is showing that training-free composition works much better when each attribute is injected at its own optimal layer, with near-orthogonality in residual space explaining why, though the evidence is limited to a few attributes, model families, and up to mostly two or three-way compositions.
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.