Video World Models · 2026-10-14

LongForcing: Scaling Temporal Manifold Matching for Long-Horizon World Rollout

Ziyu Zhu, Zhaoxu Sun, Fan Jiang, Mengchao Wang, Junfeng Ni, Ting-Bing Xu, Zhicheng Liu, Mu Xu

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

LongForcing is a distillation method for long-horizon world rollout that scales teacher supervision along the temporal axis. It keeps short-horizon teacher-forcing and causal ODE-distillation, but replaces the final DMD real-score teacher with an extended-horizon bidirectional teacher that supervises longer student self-rollouts using the standard distribution-matching formulation.

Why it matters

The authors argue that short-horizon distillation leaves a supervision-horizon gap: delayed rollout failures beyond the training window are not directly supervised, so drift can progress during long closed-loop generation. LongForcing targets this by extending distribution-matching supervision horizon without changing the student architecture or increasing inference cost.

Method

Retain short-horizon teacher forcing and causal ODE distillation stages from Causal Forcing.
In final distribution-matching distillation, replace the real-score teacher with an extended-horizon bidirectional teacher that scores longer student self-rollouts.
Supervise longer contiguous portions of on-policy trajectories using the standard DMD objective, without modifying the student architecture or increasing inference cost.

Limitation

The finite-horizon and long-run guarantees they present are conditional on assumptions about teacher quality and student–teacher mismatch on the student’s start distribution, including controlled leakage, mismatch, and recovery for the blockwise extension.

Abstract (from arXiv)

Autoregressive video diffusion models have emerged as a foundation for long-horizon world rollout. However, when extended far beyond their training windows, these models suffer from progressive drift: visual quality deteriorates, motion becomes inconsistent, and object identity gradually collapses as generated frames are recursively reused as context. We study a supervision-horizon gap: short-horizon distillation leaves later rollout states without direct supervision. We show that matching short-window marginals need not determine long-horizon behavior, while longer on-policy trajectories expose delayed failures to training. This motivates temporal manifold matching at scale for long-horizon world models. Building on this insight, we introduce LongForcing, a distillation approach that scales teacher supervision to long student self-rollouts. LongForcing preserves the short-horizon teacher-forcing and causal ODE-distillation stages, while replacing the realscore teacher in the final distribution-matching stage with an extended-horizon bidirectional teacher. The long-horizon teacher supervises extended student self-rollouts using the standard distribution-matching formulation, without modifying the student architecture or increasing inference cost. We evaluate LongForcing on text-to-video generation and interactive action-conditioned world models, achieving a VBench Quality score of 85.52 and improving WorldRoamBench 60s action score by 23.9 points in first-person ablations. Qualitative stress tests further sample interactive rollouts up to 24 hours, far beyond the training horizon. These results suggest that scaling supervision along the temporal axis is a practical step toward long-horizon world models. The project website and code are at https://github.com/Fantasy-AMAP/LongForcing.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC

preprint