AGF trains on 5-second clips, tops VBench at 60-second video

A new arxiv preprint opens with the problem it is trying to solve: "Autoregressive video generation models generate videos sequentially, with each new block conditioned on previously generated blocks." The ten authors, led by Chenxi Xia, argue that recent Self Forcing methods closed the training-versus-inference mismatch but left a different crack open. "The model remains highly sensitive to errors in its generated history," they write, so small deviations are amplified as a clip gets longer. They give the crack a name: "the local history robustness gap."

Their fix is called Adversarial Gradient Forcing, or AGF. In the authors' words: "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."

The outcome the paper leads with is "the highest VBench Total score on 60-second video generation and the best visual quality on four-minute video generation." The economic claim is in the sentence after: "AGF achieves such long-horizon robustness while requiring only 5-second rollouts for training."

The abstract names no baseline model, no competitor by name, and no compute figure; it is a one-page claim, not a benchmark table. Two of the researchers in our Who's Who tracker had posted the link by the time we picked it up.