Zhejiang, Microsoft History Forcing Hits 79.82 on VBench-Long
TL;DR
- History Forcing is a distillation framework that strengthens historical supervision so few-step autoregressive video models stay consistent across long rollouts.
- The 1.3B student model scores 85.07 on VBench at 15 seconds and 79.82 on VBench-Long at 60 seconds, approaching the 5B LongLive-2.0.
- On out-of-sight tasks, Evolution Quality rises from 2.35 to 2.70 on PlayWorld and Object Consistency from 58.43 to 62.18 on MemoBench.
A 1.3B-parameter autoregressive video model, trained with a new distillation recipe, posts a VBench quality score of 85.07 on 15-second clips and a VBench-Long aggregate of 79.82 at 60 seconds, according to a preprint from researchers at Zhejiang University, Microsoft Research, Monash University and Adelaide University.
The method, History Forcing, targets a specific failure mode. "Few-step autoregressive video diffusion enables efficient generation, yet maintaining consistency over long rollouts remains challenging, especially when subjects or scenes reappear after going out of sight," the authors write. The framework is pitched as an on-policy distillation that pulls more signal out of past frames and into the student's training target.
On out-of-sight benchmarks the paper reports Evolution Quality rising from 2.35 to 2.70 on PlayWorld and Object Consistency climbing from 58.43 to 62.18 on MemoBench. The student "outperforms the compared 1.3B models, including those distilled from 14B teachers, and approaches the 5B LongLive-2.0 model on VBench-Long," the authors claim.
It is the second paper in the 'Forcing' lineage of long-video distillation work we have logged this week, following Wave Forcing on Monday. The abstract does not name the teacher model used for distillation.
Originally reported by arxiv.org
Read the original article →Original headline: Zhejiang, Microsoft Research's History Forcing Keeps Video World Models Coherent Over Long Horizons