Next Forcing: Multi-Chunk Prediction for Causal World Modeling (arXiv 2606.11187)

Forum topic · 小凯 · 2026-06-11

Summary

Next Forcing is a multi-chunk prediction (MCP) framework for causal world modeling, presented in arXiv paper 2606.11187 by Gangwei Xu and colleagues. Autoregressive video generation, a dominant paradigm for World Action Models (WAMs), suffers from slow training convergence, limited converged accuracy at high frame rates, and slow inference due to iterative video denoising. Inspired by multi-token prediction in large language models, Next Forcing augments a main model with lightweight auxiliary MCP modules that simultaneously denoise video chunks at multiple future time scales. These modules form a causal chain across prediction depths and leverage intermediate features from the main model's multi-layer fusion to predict future dynamics, providing explicit supervision about what comes next. The approach achieves faster training, higher accuracy, and approximately 2x faster inference, reaching state-of-the-art results on the RoboTwin benchmark with 94.1/93.5% scores. Released June 9, 2026, in the computer vision domain.

Paper Overview

Field: Computer Vision (CV)
Authors: Gangwei Xu, Qihang Zhang, Jiaming Zhou, Xing Zhu, Yujun Shen, Xin Yang, Yinghao Xu
Published: 2026-06-09
arXiv: 2606.11187

Abstract

Autoregressive video generation has emerged as a powerful paradigm for World Action Models (WAMs). However, existing approaches suffer from slow training convergence and limited converged accuracy, particularly at high frame rates, as the training supervision is confined to the current chunk without explicit signals about future dynamics; they also suffer from slow inference due to iterative video denoising.

Next Forcing is a multi-chunk prediction (MCP) framework for causal world modeling that enables faster training, higher accuracy, and accelerated inference. Inspired by multi-token prediction in large language models, Next Forcing introduces an MCP training objective that augments the main model with lightweight auxiliary MCP modules to simultaneously denoise video chunks across multiple future time scales.

The MCP modules form a causal chain across prediction depths, using intermediate features from the main model's multi-layer fusion to predict future dynamics.

Results

State-of-the-art performance on RoboTwin: 94.1 / 93.5%
2x inference speedup

---

*Auto-collected on 2026-06-11*