Streaming video world models generate future frames chunk by chunk, using their own previous outputs and a stream of actions as context. When trained only on short clips, these models tend to drift once a rollout extends beyond the training horizon, because prediction errors feed back into the model's inputs and accumulate.
The arXiv paper introduces Frechet Forcing, a training method designed to mitigate this drift. The abstract suggests the approach specifically targets the re-entering prediction error that destabilises long-horizon rollouts, though the truncated abstract does not detail the mechanism.
As the source is only the abstract, the exact workings of Frechet Forcing are not described here. The paper is available on arXiv under the identifier 2610.25666.