SAKED: Mitigating Hallucination in Large Vision-Language Models via Stability-Aware Knowledge Enhanced Decoding

TL;DR

This paper introduces SAKED, a training-free decoding method that reduces hallucinations in large vision-language models by assessing and leveraging internal knowledge stability, leading to more reliable outputs.

Contribution

The paper presents a novel stability-aware decoding approach that quantifies knowledge stability within models to mitigate hallucinations without additional training.

Findings

SAKED significantly reduces hallucinations across multiple models and tasks.

It achieves state-of-the-art performance in hallucination mitigation.

The method is model-agnostic and seamlessly integrates into existing architectures.

Abstract

Hallucinations in Large Vision-Language Models (LVLMs) pose significant security and reliability risks in real-world applications. Inspired by the observation that humans are more error-prone when uncertain or hesitant, we investigate how instability in a model 's internal knowledge contributes to LVLM hallucinations. We conduct extensive empirical analyses from three perspectives, namely attention heads, model layers, and decoding tokens, and identify three key hallucination patterns: (i) visual activation drift across attention heads, (ii) pronounced knowledge fluctuations across layers, and (iii) visual focus distraction between neighboring output tokens. Building on these findings, we propose Stability-Aware Knowledge-Enhanced Decoding (SAKED), which introduces a layer-wise Knowledge Stability Score (KSS) to quantify knowledge stability throughout the model. By contrasting the most stability-aware and stability-agnostic layers, SAKED suppresses decoding noise and dynamically leverages the most reliable internal knowledge for faithful token generation. Moreover, SAKED is training-free and can be seamlessly integrated into different architectures. Extensive experiments demonstrate that SAKED achieves state-of-the-art performance for hallucination mitigation on various models, tasks, and benchmarks.