VlogReward Evaluates Vlog Editing with Multi-Dimensional Feedback

The growing popularity of vlogs has created a demand for automated tools to assess and refine video editing. However, evaluating vlogs is inherently subjective and lacks standardized criteria, datasets, and effective reward models. This research addresses these gaps by defining a comprehensive, multi-dimensional vlog evaluation framework, developed with input from professional creators, which categorizes assessment into six key dimensions: Creativity, Consistency, Concept Design, Cinematography, Narration, and Pacing. To support this, the researchers curated a large dataset of 100,000 vlog edits and a benchmark called VRMBench. They then introduce VlogReward, a robust reward model capable of providing both detailed multi-dimensional scores and practical feedback for iterative improvements. Technically, VlogReward enhances the Group Relative Policy Optimization (GRPO) framework with an adjustable inter-group comparison reward, which helps the model better distinguish between varying quality edits. This system achieves state-of-the-art results, significantly outperforming leading Multimodal Large Language Models (MLLMs) such as GPT-5 and Gemini-3-Pro.