ARGUS: Hallucination and Omission Evaluation in Video-LLMs

TL;DR

ARGUS is a new benchmark for evaluating Video-LLMs on freeform video captioning, specifically measuring hallucinations and omissions to better understand their true performance.

Contribution

The paper introduces ARGUS, a benchmark that assesses hallucination and omission in Video-LLMs during freeform captioning tasks, addressing limitations of existing multiple-choice benchmarks.

Findings

Video-LLMs hallucinate more in freeform tasks than in multiple-choice.

ARGUS effectively quantifies hallucination and omission rates.

Benchmark reveals significant gaps in current Video-LLM capabilities.

Abstract

Video large language models have not yet been widely deployed, largely due to their tendency to hallucinate. Typical benchmarks for Video-LLMs rely simply on multiple-choice questions. Unfortunately, VideoLLMs hallucinate far more aggressively on freeform text generation tasks like video captioning than they do on multiple choice verification tasks. To address this weakness, we propose ARGUS, a VideoLLM benchmark that measures freeform video captioning performance. By comparing VideoLLM outputs to human ground truth captions, ARGUS quantifies dual metrics. First, we measure the rate of hallucinations in the form of incorrect statements about video content or temporal relationships. Second, we measure the rate at which the model omits important descriptive details. Together, these dual metrics form a comprehensive view of video captioning performance.