Zhejiang University Nanjing University Wuli RTP-LLM Alibaba

Adversarial Gradient Forcing

for Robust Few-Step Autoregressive Video Generation

Chenxi Xia1,2,♯Yixuan Xu2Shan Liu2Jiapeng Wang2Maohua Li2,3
Hanlin Tang2Kan Liu2Tao Lan2Weiyi Lu2,†Xinggao Liu1,§

1 Zhejiang University2 Alibaba Group3 Nanjing University

♯ Work done during internship at Alibaba† Project lead§ Corresponding author

arXiv Hugging Face Code

Watch the AGF method overview.

Explore comparisons
TL;DR

AGF mitigates error accumulation in few-step autoregressive video generation by training models to withstand adversarial history perturbations. Trained on only five-second rollouts, AGF sustains high-quality video generation for four minutes.

Abstract

Autoregressive video generation models generate videos sequentially, with each new block conditioned on previously generated blocks. By training on the model's own rollouts, Self Forcing and its follow-up methods substantially mitigate exposure bias caused by the mismatch between training- and inference-time history distributions. However, these models still suffer from error accumulation during long-horizon generation. We identify a key underlying limitation: the model remains highly sensitive to errors in its generated history, allowing small deviations to be amplified through subsequent generations. We refer to this underexplored problem as the local history robustness gap.

To address this gap, we introduce Adversarial Gradient Forcing (AGF), a framework designed to improve the robustness of few-step autoregressive video models to perturbations and accumulated errors in their own generated histories. AGF uses distribution matching gradients from future block predictions to select adversarial perturbations of self-generated histories. It then jointly trains the generator on the original and perturbed histories to improve the robustness of future generation. Extensive experiments show that AGF achieves the highest VBench Total score on 60-second video generation and the best visual quality on four-minute video generation. Notably, AGF achieves such long-horizon robustness while requiring only 5-second rollouts for training.

Motivation

Every generated block becomes context for the next one. Yet existing models remain highly sensitive to errors in their generated history, allowing small deviations to be amplified through subsequent generations. We refer to this sensitivity as the local history robustness gap.

The controlled example below keeps the model and future sampling randomness fixed. A one-time perturbation to one history block causes artifacts to emerge earlier and grow more severe. AGF targets this sensitivity directly by training on histories selected for their effect on future generation.

A controlled comparison of Self-Forcing with an unperturbed history and with a single history perturbation.
A single history perturbation can alter the entire continuation. The lower rollout differs only by the intervention to its generated history.

Adversarial Gradient Forcing

AGF learns to continue from imperfect context by constructing small history perturbations that are especially harmful to future predictions. Distribution matching gradients provide the signal for selecting these perturbations. The model then learns from the clean history and its adversarial counterpart together.

The AGF pipeline: sample on-policy trajectories, select adversarial histories using future DMD gradients, and jointly train on clean and perturbed histories.

  1. On-policy rollout

    The generator produces a five-second rollout using its own predictions as history. The trajectory is recorded without retaining the full computation graph, and one history block is selected for perturbation.

  2. Adversarial history selection

    DMD gradients from blocks after the selected history block identify a direction that worsens the subsequent generation. AGF takes a bounded, RMS-normalized step in this direction, producing a challenging history for the model to learn from.

  3. Joint distillation

    With the perturbation fixed, the generator is trained jointly on the clean and adversarial histories. During early training, a prediction consistency regularizer aligns the first-step velocity predictions under the two histories, helping preserve global scene structure. The regularizer is then removed for the remaining training. At inference, AGF retains the original four-step sampler and adds no perturbation-selection procedure.

Long-horizon performance

We evaluate six methods at 5 seconds, 60 seconds and 240 seconds using 944, 240 and 120 prompts, respectively. AGF remains competitive on short videos and achieves the highest 60-second VBench Total and four-minute Quality 5. Human evaluation also favors AGF, with the best mean rank across 240 ranking tasks.

Table 1 · VBench evaluation. Best and second-best results are highlighted; higher is better.
Method5 seconds60 seconds240 seconds
Total ↑Quality ↑Semantic ↑Total ↑Quality 5 ↑
Self-Forcing82.2182.9566.0179.5682.47
Causal Forcing82.9483.8263.8879.8381.79
LongLive80.3482.1170.5979.8183.62
SGF82.0883.5664.0979.6784.07
OPSD-V81.9583.1667.6780.0682.84
AGF (ours)82.0384.6768.6681.4784.82

Quality 5 (Q5 in the paper) averages subject consistency, background consistency, motion smoothness, aesthetic quality and imaging quality. It excludes Dynamic Degree and is not the official VBench overall score.

Quality across the full rollout

AGF maintains its initial Quality 5 quality across four minutes; the five baselines decline over time.
Quality 5 change relative to the first 20-second segment, over 120 prompts. Shading: pointwise 95% bootstrap confidence intervals. AGF stays closest to its initial quality throughout the four-minute rollout, ending at +0.03 percentage points; all five baselines decline by 0.96–2.61 points.

Human preference · mean rank

AGF has the lowest, best mean rank at 2.23, followed by LongLive 2.85, OPSD-V 3.24, SGF 3.32, Self-Forcing 4.43 and Causal Forcing 4.93.
Mean rank across 240 ranking tasks over 120 prompts. Lower is better; AGF receives the best mean rank (2.23), followed by LongLive (2.85).

Robustness to history errors

We test whether robustness learned during training transfers to temporary errors at inference. All six methods are evaluated on 120 prompts. Either mild progressive blur or an additive error is applied to newly generated history blocks from 30 to 60 seconds. The intervention then stops, and generation continues to four minutes. The saved video frames themselves are not edited.

Paper Figure 15: Quality 5 trajectories under temporary blur and additive history errors. All methods lose quality during the intervention; AGF recovers closest to its initial quality.
Figure 15 · Generation stability under temporary history perturbations. Mean local Quality 5 change from each video’s first 20 seconds. Gray shading marks the 30–60-second intervention; colored bands are pointwise 95% bootstrap intervals from 2,000 prompt resamples.

All methods lose quality during the intervention and recover after it ends, but AGF returns closest to its initial level. Over 90–240 seconds, AGF’s mean Quality 5 change is −0.17 percentage points after blur and −0.13 after additive error. The corresponding ranges for the five baselines are −2.10 to −1.03 and −2.31 to −1.01 percentage points. These trajectories reflect both natural quality drift and the effects of the temporary intervention.

Cite this work

@misc{xia2026adversarial,
  title         = {Adversarial Gradient Forcing for Robust Few-Step Autoregressive Video Generation},
  author        = {Chenxi Xia and Yixuan Xu and Shan Liu and Jiapeng Wang
                   and Maohua Li and Hanlin Tang and Kan Liu and Tao Lan
                   and Weiyi Lu and Xinggao Liu},
  year          = {2026},
  eprint        = {2610.14553},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.14553}
}