paper-with-me

홈 › Papers

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

2025-03-11 · Beitao Chen, Xinyu Lyu, Lianli Gao, Jingkuan Song, Heng Tao Shen

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.

📄 PDF Abstract BibTeX arXiv:2503.08216

Code (0)

등록된 구현이 없습니다.

Tasks

AttributeDisentanglementHallucination

Methods 이 논문이 사용한 방법론

Softmax The Softmax output function transforms a previous layer's output into a vector of probabilities. It is commonly used for multiclass classification. Given an input vector $x$…
Attention 설명 없음

Similar Papers 제목 키워드 기반

Attention Hijacking in Trojan Transformers

2022-08-09 · Weimin Lyu, Songzhu Zheng, Tengfei Ma, Haibin Ling 외

Trojan attacks pose a severe threat to AI systems. Recent works on Transformer models received explosive popularity and the self-attentions are now indisputable. This raises a central question: Can we reveal the Trojans …

Divergent discourse between protests and counter-protests: #BlackLivesMatter and #AllLivesMatter

2016-06-22 · Ryan J. Gallagher, Andrew J. Reagan, Christopher M. Danforth, Peter Sheridan Dodds

Since the shooting of Black teenager Michael Brown by White police officer Darren Wilson in Ferguson, Missouri, the protest hashtag #BlackLivesMatter has amplified critiques of extrajudicial killings of Black Americans. …

Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination

2026-05-11 · Yangneng Chen, Junlin Li, Weijun Yao, Xilai Ma 외 arxiv

Large Vision-Language Models (LVLMs) have achieved remarkable progress in multimodal tasks, yet their reliability is persistently undermined by hallucinations-generating text that contradicts visual input. Recent studies…

Attention Hijacking: Response Manipulation Across Queries in Vision-Language Models

2026-05-17 · Zhiqiang Wang, Dongrui Liu, Yan Li, Zonghao Ying 외 arxiv

Existing adversarial attacks on vision-language models (VLMs) can steer model outputs toward attacker-specified target responses, but their effectiveness often degrades when the same perturbed input is paired with differ…

Response GenerationAdversarial Attack

Structure-Aware Robust Fine-Tuning: Defending Vision-Language-Action Robots Against Physical Attention Hijacking

2026-08-04 · Jinquan Zhang, Dongfu Yin, Run Yang, Yufeng Yan 외 arxiv

Vision-Language-Action (VLA) policies promise general robotic manipulation, but their robustness against physical-world attacks remains fragile. In particular, we show that physically realizable adversarial patches can r…