WeVisCut — learning an editing policy from finished videos

TL;DR

WeVisCut combines a long-form script-to-video agent harness with a trained 9B policy. The authors report performance comparable to Claude Opus 5 in a shared tool environment, using nearly 33x fewer generated tokens per successful output minute. (source)

Authors & Org

Chenyu Mu, Wenming Weng, Qinrui Wang, Xingsong Ye, Hao Yu, Jun He, Chong Sun, Chen Li, Jing Lyu, Erkun Yang and Cheng Deng. Affiliations are not established in the captured abstract record. First submitted 2026-10-28. (source)

Method

The harness maintains shared production state so tool feedback informs later decisions and repairs. A four-stage data pipeline combines editing inversion with verified replay, converting decisions recovered from finished videos into executable supervision. A single policy learns to coordinate production and recover from failures within the same harness. (source)

Results

From 3,750 finished videos, the pipeline yields over 130K training examples. Evaluation uses 100 tasks and five metrics covering editing quality and delivery reliability. The authors report improvements over the base model and task-matched native editors on their metrics; the abstract gives no individual metric scores or uncertainty estimates. The token comparison is normalized by successful output minutes, not wall-clock time or total production cost. (source)

Significance

For Agents (LLM Agents), the useful idea is recovering and replaying the decisions behind an artifact before policy training. VibeVideoBench: evaluating finished video delivery provides a related delivery-focused evaluation lens, but its benchmark is not the one used here. (source) (related benchmark)

Open Questions

The abstract leaves tool costs, render time, training-video rights, generalization to other editors and metric details unresolved. Full-paper results, code and generated videos were not independently reviewed. (source)

Cite

• arXiv:2610.38352.
• Captured abstract.