Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
Ruopeng Gao, Yukun Huang, Jiarui Hu, Shuai Wang, Wenbo Hu, Ying Shan, Limin Wang
Nanjing University · ARC Lab, Tencent IEG · Zhejiang University
Published October 21, 2026 · on the feed October 22, 2026 · ▲ 172 · arXiv 2610.29989
The abstract
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.
The Brief
1:55A short podcast summary of the paper. Written and voiced by AI.
Transcript
Looking Ahead. Modeling Future Evolution in Autoregressive Video Generation. By Ruopeng Gao and six others, from Nanjing University, ARC Lab at Tencent, and one other institution. Published on the twenty-first of October. Its code repository is worked out: about seventy source files, with a readme and a licence.
Autoregressive video diffusion models build a video chunk by chunk. Producing a plausible next chunk, though, does not guarantee the model has captured how a scene evolves. Under teacher forcing, consecutive chunks are strongly redundant, so local visual continuity can act as a shortcut for denoising the current chunk. The model can lean on what it just saw, and feel less pressure to learn how things change. Continuations can look right without the model having learned the dynamics behind them.
Future Chunk Prediction, or FCP, targets that shortcut with an auxiliary objective. Alongside teacher-forced denoising, the model is also asked to predict chunks that lie further ahead in time, without access to the visual context in between. With nothing adjacent to copy from, the only way to succeed at that task is to model the evolution itself.
Two properties keep it light. The supervision comes directly from the training videos, so FCP needs no additional pretrained visual encoder. And it is optimized jointly with the usual objective, leaving autoregressive inference unchanged. Generation at test time works exactly as before. Only what the model was asked to learn has changed.
Compared with standard teacher-forced autoregressive training, the method achieves better physical plausibility, and does so on a lower overall training compute budget.
Better and cheaper together is the efficiency claim the authors make for future-chunk supervision. They add one more observation, framed with care: their visualizations suggest FCP fosters a more structured representation space, which may be valuable for building autoregressive world models. Asking a model to look ahead, in other words, may shape what it learns as well as what it generates.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A video model that looks right may not understand anything. Here is the trap.
You are training a video model that makes clips chunk by chunk, each chunk denoised given the true chunks before it.
The road ahead
- 1 · The problem ~7 min A video model trained on true past chunks can make plausible clips by copying local continuity, without learning how scenes evolve.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Add a training task that predicts chunks further in the future, with the chunks in between hidden, so copying cannot work.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train the generator with teacher-forced denoising plus parallel future-chunk predictors, optimize both together, then drop the predictors at inference.
- 6 · The experiment ~6 min You’ll guess: Compared with standard teacher-forced autoregressive training, what did the future-chunk method cost in overall training compute?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.