Paper review · October 23, 2026
Looking Ahead: Modeling Future Changes in Autoregressive Video Generation
Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
By adding an auxiliary objective (FCP) that predicts distant future chunks only during training, the authors improved the physical plausibility of autoregressive video generation without increasing inference cost.
Problem
Autoregressive video diffusion generates the next chunk conditioned on the previous chunks. Under teacher forcing training, adjacent chunks are very similar, so the model can lower the loss through a shortcut of extending the immediately preceding frames without learning how the scene changes.
Method
FCP (Future Chunk Prediction) is an auxiliary loss that uses Flow Matching to predict the future chunk Δ chunks ahead, looking only at the intermediate representation of the current chunk and without seeing the chunks in between. Each offset has its own lightweight predictor made of 3 DiT blocks, connected in parallel to block shortcuts that would leak information between predictors. No text conditioning is given to the predictors, so they depend only on the generator's representation. The auxiliary loss is backpropagated into the generator, and all predictors are removed at inference. No external pretrained encoder is needed.
Results
- On Physics-IQ Verified image-to-video generation, Wan2.2-5B + FCP reached an IQ of 33.8, the highest among the autoregressive models compared. MAGI-1-4.5B scored 31.1 and Helios-Base 26.9.
- Wan2.1-1.3B + FCP reached an IQ of 30.9, higher than Causal Forcing (12.1) and Causal Forcing++ (21.8) at the same 1.3B scale.
- Wan2.2-5B + FCP (33.8) also scored higher than the bidirectional Wan2.2-5B (24.9) and Wan2.2-14B (32.2). However, these comparisons do not use the same training conditions.
- The authors reported that, with the same data exposure and training budget, physical plausibility improved over standard teacher forcing while total training compute was lower. In visualizations, they also observed a tendency for the representation space to become more structured.
Limitations
Evaluation focused mainly on physical plausibility benchmarks and did not comprehensively assess scene-change modeling overall. The implementation is limited to the image-to-video setting and Wan-family backbones. The training data consists of videos about 5 seconds long. The more structured representation is an observation based on visualizations, with limited quantitative validation. Some details of the controlled comparison tables were cut off in the provided text.
Why it matters
The paper shows that adding a single training objective, without changing the inference pipeline, can improve a video model's physical understanding. Because it scales with video data alone and needs no external encoder, it is a practical direction for training world models.
Key terms
- Teacher Forcing
- A method in which, during training, the model predicts the next part conditioned on the ground-truth history instead of on its own generated output.
- Autoregressive video diffusion
- A method that splits video into chunks of several frames and generates them one after another with diffusion, each conditioned on the previous chunks.
- Flow Matching
- A generative model training method that learns the velocity field along paths connecting noise and data, in order to turn noise into data.
Original
Authors: Ruopeng Gao, Yukun Huang, Jiarui Hu, Shuai Wang, Wenbo Hu, Ying Shan, Limin Wang · Posted 2026-10-21 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.