In early June 2026, Yonghao Yu et al. proposed Video-Mirai (arXiv 2606.03971), directly addressing a long-neglected pain point in streaming autoregressive video diffusion: long-range drift.
Traditional causal video generators can only use the "past" to supervise their own representations at each step, but every emitted segment becomes a commitment that subsequent segments must inherit. The paper names this contradiction the "representation-level planning gap": the latent state that perfectly explains the current segment does not necessarily preserve the key signals needed for long-range consistency — identity, layout, motion. RhymeFlow tunes scheduling, LongLive-RAG adds retrieval, Video-Mirai takes a different angle — using the "future" as a supervision signal.
The method is clean: the causal generator does a normal forward rollout, a frozen foresight encoder reads the full segment in a non-causal way to produce a semantic target, and a lightweight predictor distills this stop-gradient target back into the causal state. The prediction target is the representation, not the generator input; the encoder and predictor are thrown away at inference, the original architecture, single-step FLOPs and KV-cache behavior are completely unchanged, and it's a zero-intrusion for latency-sensitive serving stacks.
On the effect, 5-second VBench pushes the Causal-Forcing baseline from 83.8 to 84.6; the 30-second super-training-length rollout improves the most — subject consistency 84.9→88.5, background consistency 90.2→91.9. Ablations point to the future-conditioned target as the key ingredient, and probe analysis also shows that future frames become easier to decode from the current features.
Video-Mirai's engineering significance lies in proving "online inference must be causal, offline representation supervision does not have to be" — in line with REPA-style predictor alignment and JEPA-style latent prediction. For autoregressive video teams, this is an almost zero-cost training-time add-on, worth porting to the long-range-consistency polishing of production models like Wan and Kling.