Large Vision-Language Models (LVLMs) face a tug-of-war between powerful linguistic priors and visual evidence, often leading to \emph{semantic drift}: a progressive detachment from the input image that can abruptly emerge at specific decoding steps. Through a token-level diagnosis, we show that hallucination is frequently triggered not by the absence of grounded candidates, but by a failure of selection -- the model chooses a linguistically convenient yet visually unfaithful token even when better grounded alternatives exist. Motivated by this insight, we propose \textbf{D}ynamic \textbf{L}ogits \textbf{C}alibration (DLC), a training-free decoding framework that introduces a lightweight visual referee to intervene exactly when drift happens. At each step, DLC performs a dual-aspect grounding check on top-k candidates: (1) it assesses the intrinsic visual relevance of a candidate token and (2) its contextual visual coherence. These signals are evaluated against an adaptive historical baseline to compute a relative visual advantage, which is then used to dynamically calibrate logits and favor grounded tokens. Extensive experiments on CHAIR, POPE, SHR, GPT-4o evaluation, and MME demonstrate that DLC consistently reduces hallucinations across multiple LVLMs while preserving response quality. Further analyses validate robustness to different vision backbones and demonstrate a favorable trade-off between output quality and computational cost as the candidate pool size varies. Code will be released on https://github.com/JiaheChen2002/DLC.
@article{arxiv.2506.21509,
title = {Curing Semantic Drift: A Dynamic Approach to Grounding Generation in Large Vision-Language Models},
author = {Jiahe Chen and Jiaying He and Qiyuan Chen and Qian Shao and Jiahe Ying and Hongxia Xu and Jintai Chen and Jianwei Zheng and Jian Wu},
journal= {arXiv preprint arXiv:2506.21509},
year = {2026}
}