Next Forcing: Causal World Modeling with Multi-Chunk Prediction

Gangwei Xu Qihang Zhang Jiaming Zhou Xing Zhu Yujun Shen Xin Yang Yinghao Xu

Convergence and Success Rate on RoboTwin
2.3x faster convergence at 50 FPS
94.1 / 93.5% Clean / Random success rate
2x inference acceleration

We introduce Next Forcing, a multi-chunk prediction framework for causal world modeling. By supervising future video chunks through chained MCP modules, Next Forcing reduces the myopic supervision of autoregressive world models and improves both training convergence and inference efficiency. At inference time, MCP modules predict upcoming chunks in parallel, reducing sequential generation cost and accelerating rollout.

How It Works

Next Forcing extends causal world modeling from one-step prediction to multi-chunk prediction: chained MCP modules expose the backbone to multiple future chunks during training while keeping generation causal.

Multi-chunk prediction

The main model denoises the current chunk, while chained MCP modules predict future chunks (next1, next2, ...) using features from the main model, providing dense temporal supervision during training and enabling parallel chunk prediction at inference.

Architecture diagram of Next Forcing with chained multi-chunk prediction modules.

Video Demos on RoboTwin

The first demo shows faster training convergence. The second shows MCP-accelerated inference, where future chunks are predicted in parallel to reduce rollout cost.

PhyWorld Benchmark

On physical reasoning videos, Next Forcing produces more consistent dynamics than LingBot-VA under the same causal setup.

General Video Comparison

We evaluate pure video generation after removing the action stream. Next Forcing consistently achieves lower FVD than LingBot-VA throughout training, and the qualitative comparisons below show stronger temporal consistency.

FVD (↓) Across Training Steps

Citation

@article{nextforcing,
  title={Next Forcing: Causal World Modeling with Multi-Chunk Prediction},
  author={Gangwei Xu and Qihang Zhang and Jiaming Zhou and Xing Zhu and Yujun Shen and Xin Yang and Yinghao Xu},
  journal={},
  year={2026}
}