Textual Steering Vectors Can Improve Visual Understanding in Multimodal LLMs
Authors: Woody Haosheng Gan, Deqing Fu, Julian Asilis, Ollie Liu, Dani Yogatama, Vatsal Sharan, Robin Jia, Willie Neiswanger
Date: 2026-04-16
Summary
Shows that textual steering vectors — a technique from interpretability — can improve visual understanding in multimodal LLMs. Bridges text-domain interpretability with vision capabilities.
Key Claims
- Steering vectors derived from text can transfer to improve visual processing
- Cross-modal transfer of representation engineering techniques
- Practical method to improve multimodal model performance without retraining
Concepts
- Mechanistic Interpretability
- AI Character & Personality — steering vectors relate to persona vectors
See Also
- representation-engineering-mistral — representation engineering approaches