NTH

History Forcing: Distilling History-Dependent Continuations for Long-Horizon Video World Models

AuthorsHangyi Kuang, Chengzhi Wang, Yefei He, Weijie Wang, Yuke Zhao, Zhuokun Chen, Zisu Huang, Yifan Yang, Xiaoxuan He, Feng Chen, Bohan Zhuang

AffiliationsZhejiang University · History Forcing for long-horizon video world models. Left: Long-video generation. Our method maintains appearance consistency and scene continuity, while the baseline exhibits appearance drift and cha · Microsoft Research · Nov 2026 out of sight visible 60 s 30 s 45 s revisit 0s 15 s Mountain texture drift Body proportion drift Lost unobserved object Baseline ⋯ Baseline ⋯ 0s 10 s 20 s 30 s 40 s 50 s 60 s 0s 1.5 s 4.5 · Monash University · Adelaide University Long-video Generation Out-of-sight Generation Maintain long-horizon consistency across extended sequences Preserve unobserved changes when the camera returns

November 9, 2026 2 min read
Watch on YouTube
The one-line take

History Forcing teaches video world models to use past context more effectively, helping scenes stay coherent even when subjects disappear and later return.

Key results

85.07
VBench quality score

15-second video generation.

79.82
VBench-Long aggregate score

60-second generation by a 1.3B model.

2.70
PlayWorld Evolution Quality

Improved from 2.35 with History Forcing.

62.18
MemoBench Object Consistency

Improved from 58.43 with History Forcing.

2.31%
Training-time increase

Additional time per long-video training step.

What the paper found

History Forcing targets a key weakness in long-horizon video generation: models often lose track of subjects, scenes, and events as they generate or look away. Instead of distilling directly from a history-conditioned teacher, it compares that teacher’s predictions for the same noisy video chunk with and without informative history, then adds the weighted difference—the context residual—to the reference score or velocity target. This preserves the bidirectional teacher’s visual-quality guidance while strengthening historical consistency. For long videos, the method uses earlier student-generated chunks as history; for out-of-sight world modeling, it uses a fixed-camera teacher rollout to track scene changes while the student’s camera moves away. Tested with Wan2.1-T2V-1.3B and DreamX-World-5B, it scores 85.07 on VBench for 15-second videos and 79.82 on VBench-Long for 60-second videos, among the best results for 1.3B models. On PlayWorld, Evolution Quality rises from 2.35 to 2.70; on MemoBench, Object Consistency improves from 58.43 to 62.18. The long-video training-time increase is 2.31%, showing that the added historical supervision comes with modest overhead.

Original abstract

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. Despite progress in fidelity and rollout robustness, existing distillation methods do not fully exploit richer historical information to constrain the student's current prediction. To address this gap, we introduce \textbf{History Forcing}, a unified on-policy distillation framework that strengthens historical supervision for few-step autoregressive video generators. The framework uses the residual between teacher predictions with and without informative historical context to decouple this supervision. Such a historical context can be obtained using earlier generated chunks with less accumulated drift or a fixed-view teacher rollout that keeps the scene visible while the student looks away. The residual is then weighted and added to the score or velocity prediction used for distillation, forming a corrected target. By aligning the student with this corrected target through distillation, History Forcing incorporates long-range historical constraints into autoregressive generation. Our method achieves state-of-the-art performance among 1.3B models, with a VBench quality score of 85.07 for 15s video generation and a VBench-Long aggregate score of 79.82 for 60s generation. For out-of-sight world modeling, evaluations further demonstrate more plausible scene evolution and more consistent states when the camera returns, with Evolution Quality improving from 2.35 to 2.70 on PlayWorld and Object Consistency from 58.43 to 62.18 on MemoBench.

Read the original paper

More in World Models

Browse all 59 papers →
01World Model

Duo: A Reciprocal World Model for Full-Duplex Interaction

Hengyuan Cao, Biao Gong, Guoshan Lu, Qiyuan Zhang, Lin Liu, Yiding Tian, Yilei Jiang, Yuxin Ma, Chengzhang Wu, Kun Chen, Yiqian Xie, Ting Liu, Xin Zheng, Xijun Gu, Tianren Ma, Jiancheng Pan, Chi Zhang, Xuzhe Zheng, Siyi Li, Zhenzhong Lan

Duo aims to create an always-on audiovisual assistant that can perceive, reason, and respond continuously, including during interruptions.

Read analysis
02World Model

PanDWM: Panoramic 4D Driving World Models

Jiajun Sun, Zeyu Zhang, Zeyu Ren, Guangfeng Jiang, Yang Zhao, Hao Tang

PanDWM generates a continuous 360-degree driving view and reconstructs it into a fast 4D scene.

Read analysis
03World Model

WorldRuler: Interactive World Modeling under Configurable Rules

Mingliang Zhai, Chuanhao Li, Jiaming Tan, Zhen Li, Yuwei Wu, Yunde Jia, Zian Meng, Yuanyang Yin, Kaipeng Zhang

WorldRuler aims to make generated interactive worlds obey customizable rules, even after saving and restoring the world.

Read analysis