Understanding (Un)Reliability of Steering Vectors in Language Models
This work addresses the problem of unreliable control methods for language models, which is incremental as it builds on existing steering vector techniques to identify factors affecting their performance.
The paper investigates the unreliability of steering vectors in controlling language model behavior, finding that while all tested prompt types yield a net positive effect, they show high variance and sometimes opposite effects, with no clear best type, and that higher cosine similarity between activation differences and better separation of activations predict more effective steering.
Steering vectors are a lightweight method to control language model behavior by adding a learned bias to the activations at inference time. Although steering demonstrates promising performance, recent work shows that it can be unreliable or even counterproductive in some cases. This paper studies the influence of prompt types and the geometry of activation differences on steering reliability. First, we find that all seven prompt types used in our experiments produce a net positive steering effect, but exhibit high variance across samples, and often give an effect opposite of the desired one. No prompt type clearly outperforms the others, and yet the steering vectors resulting from the different prompt types often differ directionally (as measured by cosine similarity). Second, we show that higher cosine similarity between training set activation differences predicts more effective steering. Finally, we observe that datasets where positive and negative activations are better separated are more steerable. Our results suggest that vector steering is unreliable when the target behavior is not represented by a coherent direction.