Steering multiple language and behavior traits in large language models
Compositional Multilingual and Behavioral Attribute Steering
Computation and Language
Summary
This study looks at how to control different qualities in large language models, like the language they use, how concise they are, or whether they avoid harmful instructions. The authors test whether combining simple control signals for these qualities works without extra training. They find that controlling one trait at a time works well if put in the right model layer, and combining two or three traits works reasonably when each is applied where it works best. They also show that these control signals behave like independent directions in the model's inner workings, helping explain why combining them can be effective.
What this means in practice
- •For ai product developers: Implement multi-trait control on language model outputs by injecting steering controls at optimal model layers without retraining.$Commercial implications: Enables development of customizable AI assistants that balance language style and safety without costly retraining.
- •For natural language processing engineers: Test additive combinations of steering vectors to improve control over multiple language attributes in instruction-tuned models.
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.