What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
This provides mechanistic insights into an efficient model alignment technique for LLMs, though it is an incremental study focused on refusal behavior.
The researchers investigated how steering vectors affect large language models' internal mechanisms, specifically focusing on refusal behavior, and discovered that steering primarily interacts with attention mechanisms through the OV circuit while largely ignoring the QK circuit, with freezing attention scores reducing performance by only 8.75%. They also showed steering vectors can be sparsified by 90-99% while retaining most performance.
Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works-- specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit-- freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.