news

History Forcing posts 79.82 on VBench-Long at 60 seconds, 0.60 points above Rolling Forcing

The 1.3B student also reports 85.07 VBench quality at 15 seconds, with the state-of-the-art claim scoped to other 1.3B models rather than to the field.

By Priya Shenoy · 4 November 2026

A 1.3-billion-parameter autoregressive video model trained with a new distillation recipe reports a VBench quality score of 85.07 for 15-second generation and a VBench-Long aggregate of 79.82 at 60 seconds. The paper calls that state of the art, and the comparison set it is state of the art within is other 1.3B students, not the commercial systems most people see output from.

The method, History Forcing, comes from a preprint posted as arXiv 2611.02782. It targets a failure that anyone who has generated a long clip will recognize: the authors write that in few-step autoregressive video diffusion, "maintaining consistency over long rollouts remains challenging, especially when subjects or scenes reappear after going out of sight." A room looks different when the camera comes back to it, or an object that left frame does not return.

The training change is to score history separately from generation

History Forcing is described as a unified on-policy distillation framework that decouples historical supervision from generative supervision. In practice the teacher is asked for predictions with and without informative history, and the residual between those two predictions is weighted and added to the score or velocity prediction used as the distillation target. The student is then aligned to that corrected target, which is how long-range historical constraints get pushed into a model that only ever generates a few steps at a time.

The efficiency claim is the part worth keeping: the student is trained on rollouts of at most 15 seconds and the paper reports that it extrapolates directly to 60-second generation. Training cost does not scale with the horizon being advertised.

The winning margin is 0.60 points on an automated metric

On VBench-Long for 60-second generation, the paper reports the highest aggregate score among the compared 1.3B students at 79.82, exceeding Rolling Forcing by 0.60 points. VBench aggregates automated dimension scores such as subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality and imaging quality, so a six-tenths gap is a difference between automated estimators, not a demonstrated difference in what a viewer would pick.

Cross-paper comparison in this lineage is also harder than the shared metric name suggests. Prospective Forcing, another recent entry, evaluates 30-second videos on 200 Movie Gen Video Bench prompts and aggregates six VBench-Long quality dimensions by dividing the weighted sum by 5.5. History Forcing reports at 60 seconds. Same benchmark family, different horizon, different prompt pool, and the numbers are not interchangeable on that basis. The same caution applies to step-count scaling results, where eight times the sampling steps bought 0.06 VBench points in NVIDIA's AnyFlow checkpoints.

What a 60-second consistency score does and does not tell you

Scene continuity is the hardest thing to score automatically, because the failures are semantic. A character's jacket changing color between shots, a sign rewriting itself, a room rearranging after a cut: these register weakly in frame-level consistency metrics and strongly with a human watching. Our own multi-shot prompts exist for exactly that reason, and models that score closely on aggregate benchmarks often separate clearly once a shot change is involved.

None of this makes 79.82 a bad number. It makes it a number with a narrow jurisdiction: one parameter class, one automated suite, one horizon, against a set of academic baselines that mostly share a base checkpoint. Open weights for models like Wan 3.0 are what let this research line run at all, and they are also why its leaderboards are populated almost entirely by variants of the same underlying system.

Sources: arXiv 2611.02782, History Forcing full text, arXiv 2610.14654.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.