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: HKR-H/K/R all pass: the story has a novel test-time VLM-teacher hook, concrete VBVR-Bench gains from 0.666 to 0.781, and clear resonance around controllable video generation. It is a strong research release, not a must-write model launch.