Skip to content
QbitAI · WeChat

Kuaishou Kling Proposes VLM-as-Teacher for Test-Time Video Reasoning Optimization

快手可灵提出VLM-as-Teacher:用测试时在线优化,让视频生成模型学会按规则推理

City University and Kuaishou Kling proposed VLM-as-Teacher, which uses VLM feedback to optimize VGM LoRA at test time, reporting a 16.7-point average gain and raising VBVR-Bench from 0.666 to 0.781.

Why it matters: Kuaishou Kling and CityU use a vision-language model as a test-time teacher, feeding LoRA updates to a video generator so it follows text rules while it generates.

Read the original ↗Export Markdown