Guardado en:
Detalles Bibliográficos
Autores principales: Li, Zhaoxu, Kong, Chenqi, Bao, Peijun, Xia, Song, Tu, Yi, Yu, Yi, Jiang, Xinghao, Jiang, Xudong
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:https://arxiv.org/abs/2602.09825
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866915788551618560
author Li, Zhaoxu
Kong, Chenqi
Bao, Peijun
Xia, Song
Tu, Yi
Yu, Yi
Jiang, Xinghao
Jiang, Xudong
author_facet Li, Zhaoxu
Kong, Chenqi
Bao, Peijun
Xia, Song
Tu, Yi
Yu, Yi
Jiang, Xinghao
Jiang, Xudong
contents Hallucinations in Large Vision-Language Models (LVLMs) pose significant security and reliability risks in real-world applications. Inspired by the observation that humans are more error-prone when uncertain or hesitant, we investigate how instability in a model 's internal knowledge contributes to LVLM hallucinations. We conduct extensive empirical analyses from three perspectives, namely attention heads, model layers, and decoding tokens, and identify three key hallucination patterns: (i) visual activation drift across attention heads, (ii) pronounced knowledge fluctuations across layers, and (iii) visual focus distraction between neighboring output tokens. Building on these findings, we propose Stability-Aware Knowledge-Enhanced Decoding (SAKED), which introduces a layer-wise Knowledge Stability Score (KSS) to quantify knowledge stability throughout the model. By contrasting the most stability-aware and stability-agnostic layers, SAKED suppresses decoding noise and dynamically leverages the most reliable internal knowledge for faithful token generation. Moreover, SAKED is training-free and can be seamlessly integrated into different architectures. Extensive experiments demonstrate that SAKED achieves state-of-the-art performance for hallucination mitigation on various models, tasks, and benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2602_09825
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SAKED: Mitigating Hallucination in Large Vision-Language Models via Stability-Aware Knowledge Enhanced Decoding
Li, Zhaoxu
Kong, Chenqi
Bao, Peijun
Xia, Song
Tu, Yi
Yu, Yi
Jiang, Xinghao
Jiang, Xudong
Computer Vision and Pattern Recognition
Hallucinations in Large Vision-Language Models (LVLMs) pose significant security and reliability risks in real-world applications. Inspired by the observation that humans are more error-prone when uncertain or hesitant, we investigate how instability in a model 's internal knowledge contributes to LVLM hallucinations. We conduct extensive empirical analyses from three perspectives, namely attention heads, model layers, and decoding tokens, and identify three key hallucination patterns: (i) visual activation drift across attention heads, (ii) pronounced knowledge fluctuations across layers, and (iii) visual focus distraction between neighboring output tokens. Building on these findings, we propose Stability-Aware Knowledge-Enhanced Decoding (SAKED), which introduces a layer-wise Knowledge Stability Score (KSS) to quantify knowledge stability throughout the model. By contrasting the most stability-aware and stability-agnostic layers, SAKED suppresses decoding noise and dynamically leverages the most reliable internal knowledge for faithful token generation. Moreover, SAKED is training-free and can be seamlessly integrated into different architectures. Extensive experiments demonstrate that SAKED achieves state-of-the-art performance for hallucination mitigation on various models, tasks, and benchmarks.
title SAKED: Mitigating Hallucination in Large Vision-Language Models via Stability-Aware Knowledge Enhanced Decoding
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2602.09825