Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models
This addresses hallucination issues in vision-language models, which is a critical problem for improving reliability in applications like image captioning or visual QA, but it is incremental as it builds on existing steering techniques.
The paper tackles hallucination problems in Large Vision-Language Models by proposing Dynamic Multimodal Activation Steering, a training-free method that uses context-aware interventions based on activation patterns, resulting in significant performance enhancements that outperform existing state-of-the-art methods.
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.