Pathwise Test-Time Correction for Autoregressive Long Video Generation
Xunzhi Xiang, Zixuan Duan, Guiyu Zhang, Haiyu Zhang, Zhe Gao, Junta Wu, Shaofeng Zhang, Tengfei Wang, Qi Fan, Chunchao Guo · 2026
Read Pathwise Test-Time Correction for Autoregressive Long Video Generation by Xunzhi Xiang, Zixuan Duan, Guiyu Zhang, Haiyu Zhang, Zhe Gao, Junta Wu, Shaofeng Zhang, Tengfei Wang, Qi Fan, Chunchao Guo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05871.txt.
Description
Short video generators stumble on long clips, because small errors pile up frame after frame until the scene drifts. Xunzhi Xiang and 9 co-authors introduce Test Time Correction, a training free method for distilled autoregressive diffusion models, the fast models used for real time synthesis. Existing test time optimization works on images or short clips but fails on extended sequences, thanks to unstable reward landscapes and touchy distilled parameters.
The correction uses the initial frame as a stable reference anchor. Intermediate stochastic states along the sampling path are calibrated against that anchor as generation proceeds. Experiments show the method plugs into various distilled models with negligible overhead. Generation lengths grow while quality matches resource intensive training based methods on thirty second benchmarks. The authors frame anchoring, not optimizing, as the right test time move for long video. The work targets video generation researchers who need longer coherent output without retraining.
Catalog record
| CREATOR | Xunzhi Xiang, Zixuan Duan, Guiyu Zhang, Haiyu Zhang, Zhe Gao, Junta Wu, Shaofeng Zhang, Tengfei Wang, Qi Fan, Chunchao Guo |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-05871