Can vision models teach video AI to reason better?

Video reasoning requires AI to follow complex rules through generated visual sequences, but current models fail at fine-grained instruction following. This work flips the paradigm: instead of using vision-language models to plan, use them as evaluators that give real-time feedback during video generation. A lightweight LoRA module optimizes on VLM-extracted rewards at test time, pushing performance well beyond what either component achieves alone on symbolic and general reasoning benchmarks.