Automating Steering for Safe Multimodal Large Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wu, Lyucheng, Wang, Mengru, Xu, Ziwen, Cao, Tri, Oo, Nay, Hooi, Bryan, Deng, Shumin
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909801353576448
author Wu, Lyucheng
Wang, Mengru
Xu, Ziwen
Cao, Tri
Oo, Nay
Hooi, Bryan
Deng, Shumin
author_facet Wu, Lyucheng
Wang, Mengru
Xu, Ziwen
Cao, Tri
Oo, Nay
Hooi, Bryan
Deng, Shumin
contents Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
format Preprint
id arxiv_https___arxiv_org_abs_2507_13255
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Automating Steering for Safe Multimodal Large Language Models
Wu, Lyucheng
Wang, Mengru
Xu, Ziwen
Cao, Tri
Oo, Nay
Hooi, Bryan
Deng, Shumin
Computation and Language
Artificial Intelligence
Information Retrieval
Machine Learning
Multimedia
Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
title Automating Steering for Safe Multimodal Large Language Models
topic Computation and Language
Artificial Intelligence
Information Retrieval
Machine Learning
Multimedia
url https://arxiv.org/abs/2507.13255