Residual Decoding: Mitigating Hallucinations in Large Vision-Language Models via History-Aware Residual Guidance

TL;DR

This paper introduces Residual Decoding, a training-free technique that leverages historical information within large vision-language models to reduce hallucinations and improve visual grounding without additional training.

Contribution

The paper presents a novel, training-free residual decoding method that utilizes internal model mechanisms to mitigate hallucinations in LVLMs, enhancing their reliability and grounding accuracy.

Findings

ResDec significantly reduces hallucinations in LVLMs.

ResDec improves visual grounding and object recognition accuracy.

ResDec performs well across multiple LVLM benchmarks.

Abstract

Large Vision-Language Models (LVLMs) can reason from image-text inputs and perform well in various multimodal tasks. Despite this success, they are affected by language priors and often produce hallucinations. Hallucinations denote generated content that is grammatically and syntactically coherent, yet bears no match or direct relevance to visual input. To address this problem, we propose Residual Decoding (ResDec). It is a novel training-free method that uses historical information to aid decoding. The method relies on the internal implicit reasoning mechanism and token logits evolution mechanism of LVLMs to correct biases. Extensive experiments demonstrate that ResDec effectively suppresses hallucinations induced by language priors, significantly improves visual grounding, and reduces object hallucinations. In addition to mitigating hallucinations, ResDec also performs exceptionally well on comprehensive LVLM benchmarks, highlighting its broad applicability.