Next Forcing: Causal World Modeling with Multi-Chunk Prediction

arXiv 2026 63.1 method

TLDR

Next Forcing introduces multi-chunk prediction for causal world modeling, improving training convergence, accuracy, and inference speed in video generation.

Reasoning

The paper presents a novel multi-chunk prediction framework that addresses slow training and inference in autoregressive video world models, with strong empirical results at high frame rates. However, the abstract lacks details on limitations, such as scalability or generalization to diverse environments.

Read-first score

Read-first score 63.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.9496233

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,evaluation,metric,result

Topical relevance 29%

65.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 435.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Multi-chunk prediction (MCP) framework for causal world modeling, inspired by multi-token prediction in large language models.
Lightweight auxiliary MCP modules that simultaneously denoise video chunks at multiple future temporal horizons (next^1, next^2, next^3).
Causal chain across prediction depths using intermediate features fused from multiple layers of the main model, enabling near-future predictions to inform farther-future ones.
Dense multi-scale temporal supervision back to the main model during training.
Inference acceleration by retaining MCP modules to predict the next video chunk in parallel with the current one, achieving 2x speedup.

Methodology

Next Forcing augments a main autoregressive video generation model with lightweight auxiliary multi-chunk prediction (MCP) modules that denoise video chunks at multiple future horizons (next^1, next^2, next^3). These modules form a causal chain, leveraging intermediate features fused from multiple layers of the main model to predict future dynamics, providing dense multi-scale temporal supervision. Training uses the MCP objective, and evaluation is performed on RoboTwin, PhyWorld, and general video pretraining benchmarks, with baselines including LingBot-VA and metrics such as accuracy, convergence speed, and FVD.

Key Results

At 50 fps, Next Forcing achieves a 93.1% relative improvement over LingBot-VA at 5k training steps and 2.3x faster convergence, and establishes new state-of-the-art results on RoboTwin (94.1/93.5% on Clean/Random). It also achieves 2x inference acceleration, over 50% FVD reduction on general video pretraining, and significant improvements on PhyWorld.

Resources

Paper