vivo AI Lab and CUHK-Shenzhen introduce RefineCut, a framework that uses a deterministic verifier to provide verifiable rewards, training an 8B open-source model for on-device video editing. The model surpasses some teacher models and matches DeepSeek-V4-Pro under the same closed-loop protocol, accepted at EMNLP 2026 Main Conference.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
The vivo AI Lab and CUHK-Shenzhen team propose the RefineCut framework, enabling an 8B open-source model to self-evolve in a local closed loop through a verifier, surpassing closed-source models like GPT-5.4 on editing decision tasks, opening a new path for on-device creative agents.
Video generation models excel at synthesizing pixels from scratch, but the other half of real creation—how to select, order, and sync clips from existing footage while meeting constraints like duration, content, and music alignment—is essentially a constraint planning problem. Current mainstream approaches use closed-source large models with prompts and manual workflows as the 'editing brain,' but this freezes decision strategies in prompts, ties cost, latency, and privacy to cloud APIs, and yields unstable outputs.
Phone manufacturers need small editing agents that run locally and learn continuously. The RefineCut framework, proposed by vivo AI Lab and CUHK-Shenzhen, addresses this need by turning editing plans into checkable objects and using a deterministic verifier for self-evolution.
This article's information primarily comes from QbitAI's report on the vivo AI Lab and CUHK-Shenzhen research team, making it a secondary source. The paper has been accepted at EMNLP 2026 main conference, but specific experimental details and scores should be verified against the original paper.
Primary report
Primary source