STAR: Mitigating hallucinations in vision-language models through object-centric prompting

Large Vision-Language Models (LVLMs) have achieved remarkable success in visual understanding tasks, yet persistent object hallucinations remain a critical bottleneck for their reliable deployment. Existing mitigation strategies often suffer from high computational costs or the inherent resolution constraints of visual encoders. We propose a lightweight and training-free Spatial Truth-Anchored Recognition (STAR) framework designed to suppress hallucinations by leveraging CNN-based object detection priors. Our approach systematically integrates structured object-level evidence—including category labels, spatial coordinates, and confidence scores—directly into the reasoning prompt. The framework operates through a synergistic two-stage mechanism: global context alignment to establish robust visual anchors, and adaptive local verification via high-resolution cropping for low-confidence regions. Extensive experiments on the MME, POPE, and CHAIR benchmarks demonstrate that STAR consistently outperforms state-of-the-art decoding-time interventions across various model scales, including Gemma-3 and Qwen-3.5-VL. Our results provide empirical evidence that explicit object-centric anchors effectively counteract the dominance of statistical language priors, offering a generalizable and efficient solution for enhancing the factual grounding of LVLMs.