[Paper Review] VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models

Yuxuan Wang, Yueqian Wang+3 more|arXiv (Cornell University)|Jun 24, 2024

Schizophrenia research and treatment4 citations

TL;DR

This paper introduces VideoHallucer, the first comprehensive benchmark for evaluating intrinsic and extrinsic hallucinations in large video-language models (LVLMs). Using an adversarial binary VideoQA setup with paired basic and hallucinated questions, it reveals that current LVLMs suffer from widespread hallucinations, especially in detecting extrinsic factual hallucinations, and that scaling data and parameters offers limited improvement for such cases. A self-PEP framework further enhances hallucination resistance by 5.38% on average.

ABSTRACT

Recent advancements in Multimodal Large Language Models (MLLMs) have extended their capabilities to video understanding. Yet, these models are often plagued by "hallucinations", where irrelevant or nonsensical content is generated, deviating from the actual video context. This work introduces VideoHallucer, the first comprehensive benchmark for hallucination detection in large video-language models (LVLMs). VideoHallucer categorizes hallucinations into two main types: intrinsic and extrinsic, offering further subcategories for detailed analysis, including object-relation, temporal, semantic detail, extrinsic factual, and extrinsic non-factual hallucinations. We adopt an adversarial binary VideoQA method for comprehensive evaluation, where pairs of basic and hallucinated questions are crafted strategically. By evaluating eleven LVLMs on VideoHallucer, we reveal that i) the majority of current models exhibit significant issues with hallucinations; ii) while scaling datasets and parameters improves models' ability to detect basic visual cues and counterfactuals, it provides limited benefit for detecting extrinsic factual hallucinations; iii) existing models are more adept at detecting facts than identifying hallucinations. As a byproduct, these analyses further instruct the development of our self-PEP framework, achieving an average of 5.38% improvement in hallucination resistance across all model architectures.

Motivation & Objective

• To address the lack of comprehensive evaluation for hallucinations in large video-language models (LVLMs), especially those arising from dynamic content like actions and events.
• To establish a clear taxonomy of hallucinations in LVLMs, distinguishing between intrinsic (contradicting video content) and extrinsic (unverifiable from video) types, with subcategories for detailed analysis.
• To develop a robust, adversarial binary VideoQA evaluation framework that minimizes language bias and improves reliability in hallucination detection.
• To empirically analyze the hallucination vulnerabilities of 11 state-of-the-art LVLMs and identify the limitations of scaling data and model parameters.
• To guide the development of improved defense mechanisms, such as the self-PEP framework, through insights from the benchmark’s findings.

Proposed method

• Proposes a two-tiered hallucination taxonomy: intrinsic hallucinations (inconsistent with video) and extrinsic hallucinations (unverifiable from video), with five subtypes: object-relation, temporal, semantic detail, extrinsic factual, and extrinsic non-factual.
• Designs an adversarial binary VideoQA evaluation protocol where each video is paired with a basic question and a hallucinated version, ensuring balanced 'yes' and 'no' responses to reduce language bias.
• Constructs high-quality, human-annotated question pairs with detailed explanation annotations to improve clarity and reduce misinterpretation during evaluation.
• Employs a binary classification setup where models must distinguish between basic (correct) and hallucinated questions, measuring accuracy on both types.
• Introduces the self-PEP framework, a self-supervised data augmentation and fine-tuning method that improves hallucination resistance by learning from model predictions on adversarial examples.
• Evaluates 11 LVLMs across all VideoHallucer categories, including intrinsic and extrinsic hallucinations, using standardized metrics and ablation studies on scaling effects.

Experimental results

Research questions

• RQ1 To what extent do current large video-language models exhibit intrinsic and extrinsic hallucinations when reasoning over dynamic video content?
• RQ2 How does scaling model size and training data impact the detection of different hallucination types, particularly extrinsic factual hallucinations?
• RQ3 Why do models perform better at recognizing facts than detecting hallucinations, and what are the underlying causes of this discrepancy?
• RQ4 Can a self-supervised framework like self-PEP effectively improve hallucination resistance across diverse LVLM architectures?
• RQ5 How do human-level performance and model-level performance compare in detecting hallucinated questions across various hallucination subtypes?

Key findings

• A majority of current LVLMs exhibit significant hallucination issues, with performance gaps between humans and models being substantial across all VideoHallucer settings.
• Scaling model parameters and training data improves detection of basic visual cues and counterfactual hallucinations but provides limited benefit for detecting extrinsic factual hallucinations.
• Models are consistently more proficient at recognizing facts than detecting hallucinations, indicating a fundamental asymmetry in their reasoning capabilities.
• The self-PEP framework achieves an average of 5.38% improvement in hallucination resistance across all evaluated model architectures, demonstrating its effectiveness.
• Gemini-1.5-Pro shows the highest overall accuracy on hallucinated questions (64% on extrinsic factual), but still falls short of human-level performance.
• LLaMA-VID and VideoLaVIT show strong baseline performance on basic questions (98% and 97.5% respectively), but struggle with hallucinated ones (2.5% and 6% accuracy), highlighting vulnerability to hallucination.