Next Forcing:
Causal World Modeling with Multi-Chunk Prediction

Gangwei Xu^1,2   Qihang Zhang¹   Jiaming Zhou^1,4   Xing Zhu¹   Yujun Shen¹   Xin Yang²   Yinghao Xu^3,1
1Robbyant, Ant Group   ²HUST   ³HKUST   ⁴HKUST (GZ)

Overview

Next Forcing tackles the myopic supervision problem in autoregressive video world models: next-chunk denoising often learns local appearance shortcuts instead of long-range dynamics, especially at high frame rates.

By training lightweight Multi-Chunk Prediction (MCP) modules to predict multiple future chunks, Next Forcing provides denser temporal supervision, achieves faster and more stable convergence across frame rates, sets new state-of-the-art results on RoboTwin, and enables 2x inference acceleration via parallel chunk generation.

Highlights

• Multi-Chunk Prediction (MCP): auxiliary modules predict next^1, next^2, and next^3 chunks to provide long-range temporal supervision beyond the current chunk.
• Faster and stable training: Next Forcing converges faster and reaches higher success rates across frame rates, with the strongest gains at high FPS where appearance shortcuts are most severe.
• LLM-style inference acceleration: the MCP module can be retained at inference to predict the next chunk in parallel with the current chunk, similar in spirit to parallel/speculative decoding in LLMs.

Method

During training, the main model denoises the current chunk, while lightweight MCP modules predict multiple future chunks through a causal chain. These future prediction losses provide dense temporal supervision to the backbone and encourage the model to learn long-range dynamics instead of local appearance shortcuts.

The same trained checkpoint supports two inference modes:

• Zero-overhead mode: remove MCP modules and run the main model exactly like the baseline.
• MCP-accelerated mode: keep the first MCP module so one autoregressive step produces both the current chunk and the next chunk.

Results

Training Convergence

Next Forcing converges faster than LingBot-VA across frame rates. The gain is most pronounced at 50 fps: on the Random setting, Next Forcing reaches LingBot-VA's 45k-step accuracy at only 20k steps, corresponding to 2.3x faster convergence.

Final RoboTwin Accuracy

Next Forcing achieves the best average success rate on the RoboTwin benchmark across 50 bimanual manipulation tasks.

Setting	X-VLA	pi_0	pi_0.5	Motus	Being-H0.7	Fast-WAM	LingBot-VA	Next Forcing
Clean	72.9	65.9	82.7	88.7	90.2	91.9	92.9	94.1
Random	72.8	58.4	76.8	87.0	89.6	91.8	91.5	93.5

Inference Acceleration

MCP-accelerated inference predicts the next video chunk in parallel with the current chunk, reducing sequential video denoising cost while preserving comparable accuracy.

Inference Mode	12 fps Clean	12 fps Random	25 fps Clean	25 fps Random	50 fps Clean	50 fps Random
Standard	94.1	93.5	92.6	91.4	91.8	90.5
MCP-accelerated (2x)	93.5	90.6	91.0	89.8	92.2	91.3

PhyWorld

On PhyWorld, Next Forcing improves both video quality and physical consistency over LingBot-VA.

Method	FVD (↓)		Abnormal Ratio (↓)
	OOT	IT	OOT	IT
LingBot-VA	5.3	3.5	12%	3%
Next Forcing	4.7	3.2	8%	2%

General Video Pretraining

On 3.5M in-house general video clips, Next Forcing also improves pure video generation after removing the action stream.

At 50k training steps, Next Forcing reduces FVD by 58% on Test Set 1 (94 vs. 225) and by 52% on Test Set 2 (97 vs. 204). It also surpasses LingBot-VA's 50k-step FVD with only 10k training steps.

Project Status

• Project page and demos
• Paper
• Training and inference code
• Model checkpoints

Citation

@article{xu2026next,
  title={Next Forcing: Causal World Modeling with Multi-Chunk Prediction},
  author={Xu, Gangwei and Zhang, Qihang and Zhou, Jiaming and Zhu, Xing and Shen, Yujun and Yang, Xin and Xu, Yinghao},
  journal={arXiv preprint arXiv:2606.11187},
  year={2026}
}