Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decoding

TL;DR

The paper tackles object hallucinations in large vision-language models caused by over-reliance on language priors. It analyzes how language priors manifest across token types and lengths and shows that longer sequences intensify priors, worsening hallucinations. Introducing Summary-Guided Decoding (SumGD), it shortens conditioning input via summaries and selectively applies this to image-related POS tokens, preserving text quality by leaving language-modeling tokens intact; two variants, SumGD-S and SumGD-D, are explored. Empirical results on CHAIR and related metrics reveal SumGD achieves state-of-the-art reductions in object hallucination while maintaining or improving text quality, and shows favorable Pareto optimality in the precision-recall trade-off across model sizes and token lengths. The work offers a principled approach to decouple visual grounding from language priors, with practical impact for deploying LVLMs in real-world captioning and description tasks.

Abstract

Large Vision-Language Models (LVLMs) demonstrate impressive capabilities in generating detailed and coherent responses from visual inputs. However, they are prone to generate hallucinations due to an over-reliance on language priors. To address this issue, we investigate the language priors in LVLMs and make two key observations: (1) Even when predicting the tokens associated with image-related part-of-speech (POS), models increasingly rely on linguistic priors as the token sequences grow, thereby amplifying hallucinations. (2) Methods that directly calibrate LVLM's output distribution to mitigate language priors can lead to a degradation in text quality or even exacerbate hallucinations. Based on these findings, we propose a novel method, Summary-Guided Decoding (SumGD). This method naturally encourages the model to focus more on image information by reducing the text context through summaries, while controlling only the image-related POS tokens to maintain text quality. Through experiments, we demonstrate that SumGD achieves state-of-the-art performance on object hallucination benchmarks. Furthermore, in terms of the trade-off between precision and recall, SumGD achieves Pareto optimality among the existing methods. Lastly, we observe that although existing methods struggle to balance the reduction of object hallucinations with maintaining text quality, SumGD demonstrates robustness in handling this challenge.

Figures (11)

• Figure 1: An example of LVLMs' hallucination. LVLMs hallucinate due to their over-reliance on previously generated text. The red fonts represent the hallucinatory content.
• Figure 2: (Top) The average JSD between the LVLM and the LLM for each POS category up to 32 tokens. (Bottom) The average JSD between the LVLM and the LLM for each POS category across intervals, with each interval consisting of 32 tokens.
• Figure 3: (Left) Attention weights of image tokens and text tokens at each decoding step (or token length). (Right) Object hallucination ratio at each generated token position.
• Figure 4: Illustration of our Summary-Guided Decoding.
• Figure 5: (Left) A position closer to the top-left indicates an optimal balance between factuality and recall. (Right) Trade-off between generated token length and hallucination (lower is better).