Video World Models · 2026-10-22

Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation

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

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors identify a shortcut in teacher-forced autoregressive video diffusion: strong temporal redundancy lets local continuity help denoise the current chunk. They propose Future Chunk Prediction (FCP), an auxiliary training objective that predicts temporally separated future chunks without access to intervening visual context, requiring no extra pretrained visual encoder and leaving autoregressive inference unchanged. FCP improves physical plausibility with lower overall training compute budget in controlled comparisons.

Why it matters

If autoregressive video models are used as components of generative world models, they should model how scenes evolve, not just generate plausible continuations. FCP targets the training signal that shapes learned representations by reducing reliance on immediate history and adding supervision from farther future chunks, with evaluation centered on physical plausibility.

Method

Introduce Future Chunk Prediction (FCP) as an auxiliary objective for autoregressive video diffusion: in addition to teacher-forced next-chunk denoising, predict temporally separated future chunks while excluding intervening chunks from conditioning context.
Implement auxiliary predictors conditioned on the generator’s intermediate hidden states hk (without explicit text conditioning), training them with flow-matching on clean and/or noisy source representations; predictions use parallel offset-specific branches.
During training, jointly optimize generator and auxiliary predictors; at inference, discard FCP modules so the original autoregressive inference process is unchanged.

Limitation

They report that due to computational constraints, experiments do not cover larger model scales or training on substantially longer videos, and all current models are initialized from pretrained bidirectional video diffusion checkpoints rather than trained from scratch; they suspect this initialization may limit how much FCP’s benefits emerge.

Abstract (from arXiv)

Autoregressive video diffusion models generate videos chunk by chunk, yet producing plausible continuations does not guarantee that they capture how scenes evolve. Under teacher forcing, strong temporal redundancy may allow local visual continuity to serve as a shortcut for current-chunk denoising, reducing the pressure to model visual evolution. We introduce Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with teacher-forced denoising that predicts temporally separated chunks without access to the intervening visual context. Drawing supervision directly from training videos, FCP requires no additional pretrained visual encoder and leaves autoregressive inference unchanged. Our method achieves superior physical plausibility over standard teacher-forced autoregressive training with a lower overall training compute budget, highlighting the effectiveness and efficiency of future-chunk supervision for modeling visual evolution. Intriguingly, our visualizations suggest that FCP fosters a more structured representation space, which may be valuable for building autoregressive world models.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-21 · Indexed by Pub-AI: 2026-10-22 UTC

preprint