acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 real-score 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.

Primary area:
generative models
Keywords:
Long Video Generation, World Model, Autoregressive Rollout

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zWMxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the supervision-horizon gap in autoregressive video diffusion: students trained with short-horizon distribution matching receive insufficient supervision for failures emerging during extended rollouts. The authors distinguish between exposing delayed failures through longer self-rollouts and constraining cross-window dependencies through joint trajectory matching. LongForcing retains Causal Forcing’s short-horizon teacher-forcing and causal ODE-distillation stages, while extending the final DMD stage to longer student-generated trajectories supervised by an extended-horizon bidirectional teacher. Experiments cover text-to-video generation and action-conditioned world modeling, supplemented by finite-horizon theoretical analysis and qualitative stress tests.

Critical strengths
  • Relevant and practically important problem. Progressive degradation during autoregressive rollout is a substantial obstacle to streaming video generation and interactive world models. The distinction between rollout exposure and the temporal scope of teacher supervision provides a useful perspective on this problem.

  • Simple integration without additional inference overhead. The method preserves the student architecture and inference procedure, concentrating changes in training. This makes it potentially useful for existing few-step autoregressive generation pipelines.

  • Encouraging evidence beyond short-video benchmarks. The 60-second interactive ablation improves action score from 64.39 to 88.26 and strict accuracy from 42.89 to 85.16. The time-resolved drift analysis and human preference study provide complementary evidence that the longer-supervision configuration improves perceptual rollout quality.

Critical weaknesses
  • Limited methodological novelty relative to closely related work. The central intervention extends standard DMD to longer trajectories using a long-horizon teacher. Context Forcing already studies long-context teacher supervision of student self-rollouts, while Self-Forcing++ supervises segments sampled from long generated trajectories. LongForcing’s distinction of jointly scoring a longer, contiguously noised path is meaningful, but its independent benefit is not established through direct comparisons with these alternatives.

  • The horizon ablation does not isolate supervision length. Section 6.3 states that each text-to-video rollout length uses a separately trained teacher checkpoint. Consequently, changes in supervision horizon are confounded with teacher quality and potentially training resources. The results are also substantially non-monotonic: 30-second Total scores are 81.87, 78.02, 79.18, and 82.83 for increasing horizons. These results establish that the longest tested configuration performs best, but do not establish a reliable temporal scaling relationship. Compute-matched comparisons are also needed.

  • The theoretical analysis is insufficiently connected to the implemented objective. The drift guarantees depend on clean-path KL divergence, whereas training minimizes divergence between noise-perturbed distributions using estimated scores. The main text does not provide a quantitative bridge between these quantities or incorporate score-estimation error. Moreover, Proposition 4.3 assumes teacher recovery from degraded states and sufficiently small conditional student–teacher mismatch, without empirically validating these conditions. The analysis offers a useful conditional interpretation, but limited algorithm-specific guarantees.

  • The claimed advantage over local matching requires more precise qualification. Proposition 4.1 concerns short-window marginal distributions. It does not establish that local conditional matching with sufficiently informative history cannot constrain the joint trajectory distribution. This distinction is particularly relevant to continuation-conditioned methods such as Context Forcing. A controlled comparison is necessary to show that full-path scoring improves upon history-conditioned local supervision.

Questions for authors
  • Can you fix a single long-horizon teacher and compare supervision lengths under matched training budgets? How much of the reported improvement remains after controlling for teacher checkpoint quality?

  • Under identical student initialization and rollout exposure, how does joint long-path DMD compare with short-window scoring and long-history-conditioned continuation matching?

  • What gradient truncation or KV-cache detachment is used during training? How closely does the implemented update approximate the gradient of the stated joint-path objective?

  • What are the long-horizon teacher’s training data, additional training cost, and independently measured capabilities? Can it demonstrably recover from degraded student-generated histories?

Reviewer_ZeUGOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LongForcing extends the final distribution-matching stage of an autoregressive video distillation pipeline. The short teacher-forcing and causal ODE stages are retained, while a teacher adapted to longer sequences scores longer on-policy student rollouts. The paper motivates this choice through the distinction between local-window and joint-path matching, and provides conditional bounds on within-horizon drift and degraded-state occupancy. Experiments cover text-to-video generation and action-conditioned world models, with supervision horizons up to 321 frames, quantitative evaluation up to 60 seconds, and longer qualitative demonstrations.

Critical strengths
  1. The intervention is simple and addresses a concrete weakness of autoregressive distillation: later student states receive no direct supervision when training stops early. It preserves the student architecture and sampling procedure. The strongest evidence is the first-person comparison in Table 5, where the 321-frame configuration raises action score from 64.39 to 88.26 and strict accuracy from 42.89 to 85.16 on 60-second rollouts. These are substantial changes, rather than just a small movement in an aggregate image-quality score.

  2. Figure 3 provides useful evidence that the gains persist beyond short clips. At 59 seconds, the reported mean drift score falls from 67.46 to 31.91 when supervision increases from 81 to 321 frames. In the blinded human study, the longer-horizon model receives 72.0% overall preference versus 16.2% for the shorter-horizon model, with ties retained. Improvements in temporal coherence and motion naturalness make this more convincing than an image-quality score alone. The main theoretical discussion also distinguishes the ideal clean-path objective from the practical noisy DMD estimator and states that the guarantees depend on teacher quality and student–teacher mismatch.

Critical weaknesses
  1. The main results show that the full recipe helps, but do not establish how much comes specifically from longer joint supervision. Section 6.1 changes both the real-score teacher and the student rollout horizon relative to Causal Forcing. Table 4 is also markedly nonmonotonic: the 30-second Total score falls from 81.87 at 81 frames to 78.02 and 79.18 at 161 and 241 frames, before reaching 82.83 at 321. This makes the interpretation of supervision length as a scaling dimension less clear. A comparison using the same long-horizon teacher and student rollouts, with short-window versus full-path scoring, would help isolate the proposed mechanism. Reporting the training cost alongside these results would also make the tradeoff easier to judge.

  2. The main paper's comparisons leave the advantage over alternative approaches to drift uncertain. Table 3 compares against LongLive, Self Forcing, and Causal Forcing; LongForcing's Total is only 0.03 above LongLive, while its semantic score is lower (78.19 versus 81.17). Figure 3 compares two configurations of LongForcing rather than alternative methods. BAgger [Po et al., 2026] and History-Guided Video Diffusion [Song et al., 2025] are relevant omissions from both the discussion and experiments: the former learns from corrective trajectories, while the latter changes how history is used during generation. A comparable Diffusion Forcing model with and without History Guidance would be particularly useful for separating the benefit of longer supervision from history-noising and guidance. Comparisons should report motion and inference cost as well as visual degradation. The related-work discussion should also acknowledge the error analysis of [Wang et al., 2026] and FramePack’s treatment of forgetting and drift [Zhang et al., 2025]. Finally, Figure 4's snapshots up to 54 minutes are encouraging evidence of visual stability, but do not show whether objects, locations, or action consequences persist across long absences. The quantitative evidence supports improved rollout stability over the tested 30–60 seconds more directly than persistent world state over much longer periods.

Missing references: [Po et al., 2026] Ryan Po, Eric Ryan Chan, Changan Chen, and Gordon Wetzstein. "BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion Models." CVPR 2026.

[Song et al., 2025] Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. "History-Guided Video Diffusion." ICML 2025.

[Wang et al., 2026] Jing Wang, Fengzhuo Zhang, Xiaoli Li, Vincent Y. F. Tan, Tianyu Pang, Chao Du, Aixin Sun, and Zhuoran Yang. "Error Analyses of Auto-Regressive Video Diffusion Models." JMLR 2026.

[Zhang et al., 2025] Lvmin Zhang, Shengqu Cai, Muyang Li, Gordon Wetzstein, and Maneesh Agrawala. "Frame Context Packing and Drift Prevention in Next-Frame-Prediction Video Diffusion Models." NeurIPS 2025.

Questions for authors
  1. Assessment-changing control: Can the same 321-frame teacher and the same long student rollouts be used for both full-path DMD and short-window scoring, with comparable training compute? How much of the gain remains?

  2. Clarification: What are the measured GPU-hours and peak memory for each horizon, including teacher adaptation separately from DMD? Are the small text-to-video differences stable across training seeds or paired evaluation resampling?

  3. Assessment-changing comparison: How does LongForcing compare with BAgger or another corrective-training method, and with an appropriate History Guidance baseline, on the same long text-to-video prompts? Please report time-resolved quality, motion/dynamic degree, and inference cost. A matched-backbone Diffusion Forcing model with and without History Guidance would help separate the benefit of extended supervision from history-noising and guidance. A current MiniMax model could additionally contextualize short-clip perceptual quality, provided the exact version and protocol are specified.

  4. Clarification: Does the Figure 4 evaluation include a departure-and-return or delayed-action test of retained world state? Please clarify whether these trajectories use resets or reconditioning.

Reviewer_qAUSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies visual and action drift in autoregressive video diffusion models beyond their training horizon and proposes LongForcing. It retains Causal Forcing's short-horizon teacher-forcing and causal ODE-distillation stages, while replacing the real-score teacher in the final distribution-matching distillation (DMD) stage with an extended-horizon bidirectional teacher. This teacher jointly scores longer student-generated trajectories. The student architecture and inference sampling procedure remain unchanged; the additional computation occurs during training.

The theoretical analysis demonstrates that local-window marginals need not identify a complete path distribution, and provides finite-horizon drift bounds conditional on ideal clean-path KL, with extensions that depend on teacher leakage, student-teacher mismatch, and recovery probabilities.

Experiments cover text-to-video generation and action-conditioned world models. Compared with Causal Forcing, five-second VBench Quality increases from 84.59 to 85.52, and 30-second VBench-Long Total increases from 81.05 to 82.83. For the 1.3B first-person world model, the 60-second action score increases from 64.39 to 88.26 between the 81-frame and 321-frame supervision configurations. The paper also reports time-resolved drift measurements and a blinded preference study with 20 evaluators. Rollouts of up to 24 hours are explicitly qualitative stress tests; quantitative evaluation mainly covers 30-60 seconds.

Critical strengths
  1. The training change improves rollout performance while preserving the student inference procedure. Both generation regimes report gains. In particular, the 1.3B first-person model gains 23.87 action-score points on 60-second rollouts, while strict accuracy increases from 42.89 to 85.16. This suggests practical value without adding inference operations.

  2. Several forms of evaluation support the observed improvement, and the theoretical scope is stated carefully. The paper reports aggregate gains on short and 30-second text-to-video benchmarks. The 59-second drift score decreases from 67.46 to 31.91 between the 81-frame and 321-frame configurations. In the blinded study of 20 shared cases rated by 20 evaluators, the 321-frame configuration receives 72.0% overall preference. The analysis explicitly distinguishes finite-horizon guarantees from conditional longer-run extensions.

Critical weaknesses
  1. The horizon ablation changes teacher checkpoints as well as supervision length. Section 6.3 and Appendix A.1 state that the 81-, 161-, 241-, and 321-frame text-to-video configurations use separately trained corresponding teachers. The comparison therefore changes student rollout length, teacher adaptation, and training computation together. Table 4 is also non-monotonic: 30-second Total falls from 81.87 at 81 frames to 78.02 at 161 frames before reaching 82.83 at 321 frames. A fixed long-horizon teacher supervising different student rollout lengths, together with compute-matched controls, would better establish the contribution of supervision coverage itself.

  2. Direct empirical comparisons with the closest long-horizon supervision approaches are insufficient. The related work discusses Context Forcing's long-context teacher and Self-Forcing++'s local matching, but the main text-to-video tables do not compare these approaches. There is also no controlled comparison between local-window scoring and full-path scoring under the same teacher and long student rollout. Proposition 4.1 establishes a general non-identifiability result, but does not directly demonstrate that local matching is inadequate on the evaluated video tasks. This gap limits the evidence for the distinct contribution of full-path scoring over existing long-video training methods.

Questions for authors

The following clarification and control requests could change the assessment:

  1. With a fixed 321-frame teacher checkpoint, do the main gains remain when only the student DMD rollout length changes among 81, 161, 241, and 321 frames? Please also clarify whether the teacher is fixed in the first-person world-model horizon ablation.
  2. Under the same teacher, student rollout, and matched scoring budget, how does local short-window scoring compare with joint full-path scoring?
  3. What are the total teacher-adaptation-plus-DMD compute costs and the training-seed or paired evaluation uncertainty for the horizon configurations? In particular, 30-second Total increases only from LongLive's 82.80 to 82.83, so uncertainty information would help establish this relative advantage.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.