[Paper Review] EventHallusion: Diagnosing Event Hallucinations in Video LLMs

Jiacheng Zhang, Yang Jiao+6 more|arXiv (Cornell University)|Sep 25, 2024

Physical Unclonable Functions (PUFs) and Hardware Security4 citations

TL;DR

This paper introduces EventHallusion, a new benchmark to diagnose event hallucinations in Video LLMs caused by model priors, and proposes Temporal Contrastive Decoding (TCD), a lightweight method that improves reasoning by contrasting original and temporally disrupted video inputs during decoding. TCD significantly reduces hallucinations in open-source VideoLLMs, especially on rare or misleading event scenarios, while closed-source models show stronger inherent robustness.

ABSTRACT

Recently, Multimodal Large Language Models (MLLMs) have made significant progress in the video comprehension field. Despite remarkable content reasoning and instruction following capabilities they demonstrated, the hallucination problem of these VideoLLMs is less explored compared with its counterpart in the image domain. To mitigate this gap, we propose EventHallusion, a novel benchmark that focuses on assessing the VideoLLMs' hallucination toward event, the crux of video analysis. From a hallucination attribution perspective, our EventHallusion benchmark is curated to assess a VideoLLM's susceptibility toward language priors and vision-language biases. On the other hand, we also propose a simple yet effective method, called Temporal Contrastive Decoding (TCD), to tackle the hallucination problems of VideoLLMs. The proposed TCD method rectifies the model's bias toward its priors during the decoding stage by comparing the original video with a modified version, in which temporal cues are disrupted. Through comprehensive evaluation of eight open-source and two closed-source VideoLLMs on the proposed EventHallusion benchmark, we observe that the open-source models suffer significantly from hallucination problems, whereas the closed-source ones perform markedly better. By further equipping open-source VideoLLMs with the proposed TCD approach, evident performance improvements are achieved across most metrics in the EventHallusion benchmark. Our codes and benchmark data are available at https://github.com/Stevetich/EventHallusion.

Motivation & Objective

• To address the lack of systematic evaluation for event hallucinations in Video LLMs, particularly those driven by language and visual priors.
• To develop a benchmark that specifically targets hallucinations in dynamic video events, rather than static image or temporal order hallucinations.
• To investigate why open-source VideoLLMs are more prone to event hallucinations than closed-source models.
• To propose a plug-and-play method that mitigates hallucinations without fine-tuning.
• To evaluate the effectiveness of the proposed method across diverse VideoLLM architectures and event types.

Proposed method

• EventHallusion is constructed by manually collecting videos with rare or counter-intuitive events and annotating questions that trigger model priors, such as assuming 'riding' when a person carries a bike.
• The benchmark is divided into three categories: 'Entire Rare Events', 'Interleaved Common-Rare Events', and 'Misleading Questions', to probe different hallucination triggers.
• Temporal Contrastive Decoding (TCD) compares the model's output for the original video with that of a temporally disrupted version, where frame order is shuffled to break temporal coherence.
• During autoregressive decoding, TCD uses a contrastive loss that encourages the model to generate consistent predictions across both inputs, reducing reliance on spurious priors.
• The method uses a weighted combination of logits from original and distorted inputs, with hyperparameters α=1 and β=0.1, to stabilize the decoding process.
• TCD is applied as a lightweight inference-time method, requiring no model fine-tuning or architectural changes.

Experimental results

Research questions

• RQ1 How do existing VideoLLMs perform on event hallucinations when confronted with rare or counter-intuitive video events?
• RQ2 To what extent are open-source VideoLLMs vulnerable to hallucinations compared to closed-source models on event comprehension?
• RQ3 Can a simple, non-fine-tuned method like TCD effectively reduce event hallucinations across diverse VideoLLM architectures?
• RQ4 How does the number of sampled frames in the distorted video affect the performance of TCD?
• RQ5 Why do models struggle more with interleaved common-rare events than with entirely rare events?

Key findings

• Open-source VideoLLMs exhibit severe hallucination issues on EventHallusion, especially on 'Interleaved' and 'Misleading' categories, with accuracy dropping significantly when no explicit cues are provided.
• Closed-source VideoLLMs demonstrate markedly better robustness, maintaining high accuracy across all categories, suggesting stronger intrinsic resistance to prior-based errors.
• TCD achieves consistent and significant improvements on all three tested VideoLLM backbones (LLaVA-NeXT-Video, VideoChat2, Vila), particularly on 'Interleaved' and 'Misleading' cases.
• The performance of TCD degrades when the distorted video uses fewer than 4 frames, indicating that excessive temporal disruption harms the contrastive signal.
• TCD shows minimal improvement on 'Entire Rare Events', suggesting that such scenarios remain challenging due to the lack of contextual cues and strong priors.
• On the binary classification task, TCD improves overall accuracy by up to 15.6% on LLaVA-NeXT-Video, confirming its effectiveness in reducing hallucinatory event predictions.