Mitigating Hallucinations in LVLMs via Summary-Guided Decoding

Large Vision-Language Models (LVLMs) have demonstrated impressive performance on multimodal tasks. However, they struggle with object hallucinations due to over-reliance on learned textual patterns, ignoring the provided image. To address this issue, we first investigate language priors in LVLMs.We observe two key findings: (1) Even when predicting image-related part-of-speech (POS) tokens, models increasingly rely on linguistic priors as the token sequences grow, thereby amplifying hallucinations. (2) Methods that directly control LVLM's output distribution to mitigate language priors can lead to a degradation in text quality or exacerbate hallucinations.Based on these insights, we propose Summary-Guided Decoding (SGD). This method naturally encourages the model to focus more on the image information, with control over only the image-related POS tokens for preserving text quality.Through experiments, we demonstrate that SGD achieves state-of-the-art performance on object hallucination benchmarks. Furthermore, while existing methods show a trade-off between precision and recall, SGD proves to be Pareto optimal in this respect.Lastly, we show that while existing methods suffer from text quality degradation due to such trade-offs, SGD preserves text quality to the maximum extent possible.This paper not only focuses on preventing object hallucination but also presents analysis and solutions aimed at maintaining the original properties of LVLMs.