Skip to main content Skip to main navigation

Publication

Compositional Multilingual and Behavioral Attribute Steering

Hyun Gu Kang; Daniil Gurgurov; Tanja Bäumel; Josef van Genabith; Simon Ostermann
In: Dana Arad; Yonatan Belinkov; Hanjie Chen; Nils Feldhus; Aaron Mueller; Hosein Mohebbi; Gabriele Sarti; Martin Tutek (Hrsg.). Proceedings of the 9th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP), Budapest, Hungary, Pages 655-676, ISBN 979-8-89176-476-7, Association for Computational Linguistics, 2026.

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.

Projects

More links