When generation becomes a rollout
A convincing short clip does not guarantee a coherent long sequence. Reusing generated history can accumulate errors in motion, identity, and response to controls.
From visible errors to internal state
This research studies causal text-to-video and action-conditioned generation across increasing rollout lengths. Probes examine VAE latents, transformer hidden states, and key-value histories. Comparisons between online, refreshed, and recomputed histories help separate visible drift from memory effects.
Partition-consistency post-training explores how to reduce errors across rollout boundaries while preserving short-horizon quality.
Reading the illustration
The upper row shows an idealized controlled trajectory; the lower row shows a deliberately constructed accumulating offset. Both are synthetic drawings. They illustrate the evaluation question, not outputs or measured improvements from a video model.