Can-Jie Liu

RVSD: Retrieval Vision Sparse Decoding for Mitigating Visual Hallucinations in Large Vision-Language Models

RVSD is proposed, a training-free and plug-and-play decoding framework that unifies token sparsification and Semantic-Space Visual Retrieval (SSVR) within a single decoding pass, and asemantics-directed token selection is introduced within RVSD, which selectively sparsifies redundant tokens while preserving critical vi...