Reproduce “Disentangling Steering Vectors” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Takeru Hiramatsu, Kyohei Atarashi, Koh Takeuchi, Hisashi Kashima
Researchers at Kyoto University developed Steering Vector Dissection, a framework leveraging Sparse Autoencoders (SAEs) to decompose composite Large Language Model (LLM) steering vectors into semantically consistent basis vectors. The approach provides fine-grained control over LLM behavior, evidenced by extracting distinct semantic concepts and selectively removing 'rejection' components from 'unsafe' vectors, which increased jailbreak rates on a benchmark by up to 28.57% for specific model-layer pairs.
Key insights
- Conventional steering vectors are not monolithic but are composite objects comprising multiple underlying, entangled semantic and stylistic behavioral directions.
- Sparse Autoencoders can effectively decompose these composite steering vectors into a set of semantically consistent, distinct basis vectors, each representing a purer individual concept.
- The extracted basis vectors are behaviorally meaningful, allowing for targeted manipulation and the selective removal of specific sub-behaviors (e.g., 'rejection' components from 'unsafe' vectors).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.