PlaylistEval: Can Video-Language Judges Be Trusted at Day Scale and Beyond?

2026-09-28

AI Summary

This study addresses the unknown reliability of video-language models on ultra-long videos and the absence of corresponding evaluation benchmarks. To this end, it proposes an automated agent framework coupled with a causal data augmentation method. This framework generates hundred-hour-scale cross-modal retrieval question-answer pairs without manual annotation, enforcing multimodal cross-clip reasoning by controlling answer discrepancies through causal degradation. Based on this approach, a 630-sample evaluation benchmark is constructed, achieving 93% agreement with human judgments. Experiments reveal that among 17 state-of-the-art models, the best-performing one attains only 75.4% accuracy, with performance degrading significantly as video duration increases.

Abstract

Video-language models are increasingly used as judges of video understanding, both for evaluating model outputs and for training reward models. Whether their judgments remain reliable when the evidence is buried in day-long videos has yet to be established. Existing benchmarks cannot answer this. Their videos are typically only a few minutes long, many answer pairs can be separated from the transcript alone, and collecting human judgments does not scale to ultra-long videos. We introduce PlaylistEval, an agentic framework that builds video-language judge benchmarks over 100-hour playlist collection without human annotation. It automatically generates questions with paired answers whose differences are controlled by causal degradation, so that every pair demands retrieval across the collection. The resulting benchmark contains 630 pairs across seven domains spanning both static and dynamic knowledge, and on a stratified subset of 152 pairs it agrees with human judgments 93.0% of the time (IAA 0.781). Evaluating 17 omnimodal and multimodal models from eight families reveals that frontier judges reach only 75.4% pairwise accuracy, while open-source judge models perform far behind. We further show that both retrieval and final judgment depend on using multiple modalities, and that judge accuracy degrades as the playlist set grows. We release our pipeline, benchmark, and evaluation code at https://playlisteval.github.io.

Problem

• Video-Language Judges
• Ultra-long Videos
• Benchmark Evaluation
• Video Understanding

Innovation

• Agentic Framework
• Video-Language Judges
• Causal Degradation
• Ultra-long Video Benchmark
• Multi-modal Retrieval