WeVisCut Trains a Single Agent for Long-Form Video Editing
The WeVisCut paper introduces a single-agent system that turns narration into long-form video by combining tool execution with policy learning and shared production state. Its four-stage pipeline generated more than 130,000 training examples from 3,750 finished videos, which the authors used to train a 9B policy and introduce a 100-task benchmark with five metrics. The authors report that WeVisCut outperformed its base model and task-matched native editors on the proposed metrics, matched Claude Opus 5 in comparable performance, and used nearly 33 times fewer generated tokens per successful output minute.
The work combines a concrete editing agent, training-data pipeline, and benchmark in one long-form video-production system, while its reported token-efficiency result offers a potentially useful direction for reducing agentic media-production costs.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.38352 (cs)
[Submitted on 28 Oct 2026]
Title:WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video
Authors:Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu, Jun He, Chong Sun, Chen Li, Jing Lyu, Erkun Yang, Cheng Deng
View a PDF of the paper titled WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video, by Chenyu Mu and 10 other authors