Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models

TL;DR

This paper introduces a training-free method called Dynamic Multimodal Activation Steering to reduce hallucinations in large vision-language models by dynamically selecting relevant attention heads based on semantic context.

Contribution

It reveals distinct attention head roles for truthfulness and perception, and proposes a novel, context-aware, activation steering technique for hallucination mitigation.

Findings

Significant performance improvements over state-of-the-art methods

Effective in diverse models and datasets

Enhances truthfulness and visual perception capabilities

Abstract

Large Vision-Language Models (LVLMs) exhibit outstanding performance on vision-language tasks but struggle with hallucination problems. Through in-depth analysis of LVLM activation patterns, we reveal two key findings: 1) truthfulness and visual perception capabilities predominantly engage different subsets of attention heads within the model architecture; and 2) truthfulness steering vectors vary significantly across different semantic contexts. Based on these observations, we propose Dynamic Multimodal Activation Steering, a training-free approach for hallucination mitigation. Our method constructs a semantic-based truthfulness steering vector database and computes visual perception steering vectors, enabling context-aware interventions during inference by dynamically selecting the most relevant steering vectors based on input semantic similarity and applying them to the most influential attention heads. We conduct comprehensive experiments across multiple models and datasets, demonstrating that our approach significantly enhances model performance, outperforming existing state-of-the-art methods.