LongVQUBench: Benchmarking Long-Term Video Quality Understanding of Vision-Language Models

Evaluating video quality in long-form content is a major challenge for modern Large Vision-Language Models (LVLMs). While existing benchmarks focus on short clips or semantic understanding, they fail to address the complexities of long-term perceptual fidelity, such as cumulative degradation and temporal coherence. To address this, the authors introduce LongVQUBench, a comprehensive benchmark containing over 1,200 diverse videos and 1,500 questions designed to test perceptual reasoning across extended durations.

The benchmark employs a hierarchical evaluation framework consisting of three levels: Local Event Quality Understanding (LQU) for localized distortions, Cross-Event Quality Reasoning (CQR) for integrating multiple events, and Global Quality Understanding (GQU) for holistic evaluation. To probe fine-grained sensitivity, the authors also implement a Needle Distortion Question-Answering (NDQA) paradigm, where specific spatial or temporal artifacts are sparsely inserted into long videos.

The study evaluated 14 state-of-the-art LVLMs, including proprietary, open-source, and agentic models. The results demonstrate a clear performance gap: models perform relatively well on local event detection but struggle significantly with global quality reasoning. As reasoning depth and video length increase, accuracy consistently declines. Furthermore, the authors found that simply increasing the number of sampled frames yields diminishing returns, suggesting that current models lack the architectural mechanisms required for effective long-range temporal integration.

LongVQUBench establishes a critical foundation for moving beyond simple semantic video understanding toward human-level perceptual comprehension. By exposing the limitations of current models in temporal localization and distortion attribution, this work provides a roadmap for developing more robust, explainable, and perceptually aware vision-language systems. It highlights that future progress requires better mechanisms for modeling cumulative perceptual changes rather than just increasing input frame counts.