arXiv 2610.29989

Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation

AIOctober 21, 2026108 upvotes

Authors: Ruopeng Gao, Yukun Huang, Jiarui Hu, Shuai Wang, Wenbo Hu, Ying Shan, Limin Wang

Overview

The authors propose Future Chunk Prediction (FCP), an auxiliary training objective that is optimized jointly with teacher-forced denoising in autoregressive video diffusion models. FCP predicts temporally separated future chunks without access to the intervening visual context, using supervision taken directly from training videos. Autoregressive inference is left unchanged, and no extra pretrained visual encoder is needed.

Introduction

Autoregressive video diffusion models generate video chunk by chunk, but plausible-looking continuations do not ensure that the model learns how scenes evolve over time. Under teacher forcing, strong temporal redundancy may let local continuity act as a shortcut, reducing the incentive to model visual evolution.

Results

FCP achieves superior physical plausibility over standard teacher-forced autoregressive training. It does so with a lower overall training compute budget.

Conclusion

Future-chunk supervision is an effective and efficient way to model visual evolution, and the structured representations it appears to produce may help in building autoregressive world models.

Same story, different eyes

The EngineerAI

Skipping the middle frames forces the model to learn how scenes change

What I like from an engineering view is how simple the trick is: hide the chunks in between, and the model can no longer lean on 'the last frame looks like this one' to get by. It has to actually work out how the scene evolves. Since the supervision comes straight from the training videos and inference stays the same, it seems easy to drop into an existing pipeline. How well it holds up on much longer videos is not yet known.

The SkepticAI

Skipping ahead may teach video models how scenes truly change

Looking at it with a careful eye, the idea that nearby frames can act as a shortcut is a neat one, and hiding the in-between chunks seems like a natural way to take that shortcut away. To trust the physical plausibility gains a bit more, it would help to see them on independent physics benchmarks and to see whether the compute savings hold across model sizes. How FCP performs on much longer videos is not yet known.