Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Gan, Woody Haosheng, Fu, Deqing, Asilis, Julian, Liu, Ollie, Yogatama, Dani, Sharan, Vatsal, Jia, Robin, Neiswanger, Willie
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866908371469205504
author Gan, Woody Haosheng
Fu, Deqing
Asilis, Julian
Liu, Ollie
Yogatama, Dani
Sharan, Vatsal
Jia, Robin
Neiswanger, Willie
author_facet Gan, Woody Haosheng
Fu, Deqing
Asilis, Julian
Liu, Ollie
Yogatama, Dani
Sharan, Vatsal
Jia, Robin
Neiswanger, Willie
contents Steering methods have emerged as effective and targeted tools for guiding large language models' (LLMs) behavior without modifying their parameters. Multimodal large language models (MLLMs), however, do not currently enjoy the same suite of techniques, due in part to their recency and architectural diversity. Inspired by this gap, we investigate whether MLLMs can be steered using vectors derived from their text-only LLM backbone, via sparse autoencoders (SAEs), mean shift, and linear probing. We find that text-derived steering consistently enhances multimodal accuracy across diverse MLLM architectures and visual tasks. In particular, mean shift boosts spatial relationship accuracy on CV-Bench by up to +7.3% and counting accuracy by up to +3.3%, outperforming prompting and exhibiting strong generalization to out-of-distribution datasets. These results highlight textual steering vectors as a powerful, efficient mechanism for enhancing grounding in MLLMs with minimal additional data collection and computational overhead.
format Preprint
id arxiv_https___arxiv_org_abs_2505_14071
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models
Gan, Woody Haosheng
Fu, Deqing
Asilis, Julian
Liu, Ollie
Yogatama, Dani
Sharan, Vatsal
Jia, Robin
Neiswanger, Willie
Machine Learning
Computation and Language
Computer Vision and Pattern Recognition
Steering methods have emerged as effective and targeted tools for guiding large language models' (LLMs) behavior without modifying their parameters. Multimodal large language models (MLLMs), however, do not currently enjoy the same suite of techniques, due in part to their recency and architectural diversity. Inspired by this gap, we investigate whether MLLMs can be steered using vectors derived from their text-only LLM backbone, via sparse autoencoders (SAEs), mean shift, and linear probing. We find that text-derived steering consistently enhances multimodal accuracy across diverse MLLM architectures and visual tasks. In particular, mean shift boosts spatial relationship accuracy on CV-Bench by up to +7.3% and counting accuracy by up to +3.3%, outperforming prompting and exhibiting strong generalization to out-of-distribution datasets. These results highlight textual steering vectors as a powerful, efficient mechanism for enhancing grounding in MLLMs with minimal additional data collection and computational overhead.
title Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models
topic Machine Learning
Computation and Language
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2505.14071