Fréchet Forcing holds MiniWorld-1B stable past 4,000 frames

A fine-tuning recipe called Fréchet Forcing keeps the MiniWorld-1B video world model producing coherent rollouts "beyond 4,000 frames," despite being pretrained on 253-frame clips. The arXiv preprint, from Yian Zhao, Yu Yan and co-authors, says the entire fine-tune runs on 381-frame clips "for at most 36 GPU-hours."

The angle is what the method does not need. Standard long-horizon training leans on a teacher model via distribution matching distillation, but the authors argue "the teacher becomes a bottleneck: an external video model is agnostic to the actions, whereas the model's own pretrained network drifts in the same manner as the student." Their replacement is three Fréchet-distance losses, all computed against short real clips: generative anchoring, context fidelity, and reality matching. "Because all three targets are derived from short real clips without comparing a rollout against a long recorded future, training requires no teacher, no online critic and no long videos," the abstract states.

The demonstrated domains are narrow. Results come from MiniWorld-1B on robot manipulation and camera-controlled exploration, and the abstract publishes no per-frame fidelity numbers or failure-point characterisation past the 4,000-frame mark.