Attention Hijackers: Detect and Disentangle Attention Hijacking in LVLMs for Hallucination Mitigation

TL;DR

The paper tackles LVLM hallucinations by revealing Attention Hijacking: instruction tokens can distort visual attention during decoding, producing hallucinations. It introduces AID, a training-free framework with Attention Hijackers Detection, Attention Disentanglement, and Re-Disentanglement to identify and neutralize hijacker influence while preserving beneficial instruction guidance. Across multiple benchmarks (POPE, CHAIR, MME, MM-Vet, MMBench) and backbones, AID markedly reduces hallucinations without extra training and shows robust generalization and ablations supporting its components. The approach enhances reliability of multimodal generation and suggests combining with retrieval-based methods for further factual accuracy improvements.

Abstract

Despite their success, Large Vision-Language Models (LVLMs) remain vulnerable to hallucinations. While existing studies attribute the cause of hallucinations to insufficient visual attention to image tokens, our findings indicate that hallucinations also arise from interference from instruction tokens during decoding. Intuitively, certain instruction tokens continuously distort LVLMs' visual perception during decoding, hijacking their visual attention toward less discriminative visual regions. This distortion prevents them integrating broader contextual information from images, ultimately leading to hallucinations. We term this phenomenon 'Attention Hijacking', where disruptive instruction tokens act as 'Attention Hijackers'. To address this, we propose a novel, training-free strategy namely Attention HIjackers Detection and Disentanglement (AID), designed to isolate the influence of Hijackers, enabling LVLMs to rely on their context-aware intrinsic attention map. Specifically, AID consists of three components: First, Attention Hijackers Detection identifies Attention Hijackers by calculating instruction-driven visual salience. Next, Attention Disentanglement mechanism is proposed to mask the visual attention of these identified Hijackers, and thereby mitigate their disruptive influence on subsequent tokens. Finally, Re-Disentanglement recalculates the balance between instruction-driven and image-driven visual salience to avoid over-masking effects. Extensive experiments demonstrate that AID significantly reduces hallucination across various LVLMs on several benchmarks.

Figures (8)

• Figure 1: Upper part: shows Attention Hijacking, where the Hijacker’s erroneous attention map disrupts generated tokens, causing hallucinations. Lower part: illustrates how our proposed AID isolates Hijacker's influence, allowing the LVLM to rely on its context-aware attention map, thus reducing hallucinations.
• Figure 2: The proportion of Hijacking-Induced hallucinations across different LVLMs.
• Figure 3: Illustration of Attention Hijacking. The Attention Hijacker (orange curve) shows a visual attention distribution more aligned with the hallucination token than with the image token (blue curve), indicated by consistently higher similarity values.
• Figure 4: Attention deficiencies of Visual Encoder (ViT). Vision Transformer overemphasis background regions with high-norm outlier tokens, resulting in hallucinations though Attention Hijacking.
• Figure 5: The framework of our AID. The proposed framework consists of three key components: Attention Hijackers Detection, which identifies hijackers by calculating instruction-driven visual salience; Attention Disentanglement, which masks the visual attention of identified hijackers to isolates the influence of Hijackers; and Re-Disentanglement, which adjusts the balance between instruction-driven and image-driven salience to prevent over-masking and ensure effective Disentanglement.