RePlus: Revisiting the Role of Language Priors in Large Vision-Language Models' Hallucinations

🌐 Project Page | 📖 arXiv | Github

Introduction

We propose a benchmark called RePlus to revisit the role of language priors in Large Vision-Language Models' hallucinations. RePlus stands out from existing LVLM's hallucination benchmarks in the following major aspects:

1. We are the first study to explore the positive role of language priors, filling a gap in the field's research;
2. To demonstrate the positive effects of language priors, each image-question pair in RePlus is characterized by ambiguity and challenge.

RePlus Benchmark

Our goal is to assess the role of language priors in LVLMs in aiding them to avoid hallucinations. Therefore, we divide our questions into two categories:

• Hard to answer without language priors, where we select questions with poor image quality that makes it difficult for models to interpret the visual information accurately, aiming to show that language priors can aid understanding when visual information is ambiguous.

• Impossible to answer without language priors, where we select questions that can't be answered based solely on visual information even with a powerful vision encoder, aiming to show the indispensable role of language priors in providing requisite external knowledge. It can be further divided into two groups:

  • External-Context Comprehension: Visual input is of high quality, yet the comprehension of such visuals is contingent upon the background knowledge provided by the contextual input.
  • Internal-Knowledge Inference: The model must draw upon its extensive pre-trained knowledge base to provide an answer.

Citation