VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization

Researchers propose using vision-language models as test-time teachers, not direct solvers, for video reasoning.

The method extracts task rules and progress signals from VLMs, then turns them into differentiable rewards.

These rewards guide online updates to a lightweight LoRA module in a video generation model.

The approach delivers strong gains over prior VLM-as-solver and Best-of-N methods on symbolic and general benchmarks.