• TimeJudge introduces a zero-shot framework that recasts temporal error detection as binary question pairs, eliminating the need for task-specific fine-tuning.
• TEDBench provides a rigorously constructed benchmark with 381 videos, 1524 captions, and 3048 QA pairs, covering four complexity levels for fine-grained temporal error evaluation.
• Comprehensive evaluations show that TimeJudge consistently improves recall and F1-score across multiple state-of-the-art video-LLMs for temporal consistency assessment.
• The approach is generalizable, scalable, and training-free, making it suitable for real-world evaluation of video captions in multimodal systems.