FaithScore: Fine-grained Evaluations of Hallucinations in Large Vision-Language Models

TL;DR

FaithScore is a novel, reference-free metric that evaluates the faithfulness of vision-language model answers by verifying atomic image facts, correlating well with human judgment, and revealing current models' hallucination issues.

Contribution

We propose FaithScore, a fine-grained, reference-free metric for assessing faithfulness in vision-language models, along with benchmark datasets and analysis of current model hallucinations.

Findings

FaithScore correlates highly with human judgments.

Current LVLMs often generate unfaithful hallucinated content.

Our datasets enable systematic evaluation of hallucinations.

Abstract

We introduce FaithScore (Faithfulness to Atomic Image Facts Score), a reference-free and fine-grained evaluation metric that measures the faithfulness of the generated free-form answers from large vision-language models (LVLMs). The FaithScore evaluation first identifies sub-sentences containing descriptive statements that need to be verified, then extracts a comprehensive list of atomic facts from these sub-sentences, and finally conducts consistency verification between fine-grained atomic facts and the input image. Meta-evaluation demonstrates that our metric highly correlates with human judgments of faithfulness. We collect two benchmark datasets (i.e. LLaVA-1k and MSCOCO-Cap) for evaluating LVLMs instruction-following hallucinations. We measure hallucinations in state-of-the-art LVLMs with FaithScore on the datasets. Results reveal that current systems are prone to generate hallucinated content unfaithful to the image, which leaves room for future improvements. We hope our metric FaithScore can help evaluate future LVLMs in terms of faithfulness and provide insightful advice for enhancing LVLMs' faithfulness.