Tempo-Zero

01Program videos

Every training video below comes from executing one program of the library. The trace that renders the frames also records the events, so the question, its ground-truth option and the interval of the deciding event come with the video. Play a clip: the gold bands are execution intervals, and the ground-truth option lights up once the deciding event has passed.

02Method

Tempo-Zero closes a loop of executable environments, execution-verified self-play and environment evolution. A single vision-language policy plays both roles, and every label it learns from comes from program execution.

03Main results

Training uses five rounds of LoRA self-play on rendered program videos only. The final checkpoint is evaluated zero-shot on twelve real-world benchmarks. Bold marks the best result per column within each backbone, underline marks the second best at 4B and 8B, and the small numbers give the change over the base model.

Temporal grounding (%, Avg. averages the mIoU of the first three benchmarks and GQA@0.3)

Tempo-Zero improves every grounding metric over the base model at 4B and 8B and gives the best result in 26 of their 28 metrics. Training never sees a real-world interval annotation, so these gains come from localization learned on rendered videos.

Long-video understanding and video reasoning (official accuracy, %)

Tempo-Zero improves over the base models on all eight benchmarks at 4B, 8B, and 30B-A3B. At 4B and 8B, where all self-play baselines are evaluated, it gives the best result in 13 of the 16 comparisons across eight benchmarks and two backbones.

04BibTeX

@misc{zhu2026tempozero,
  title  = {Tempo-Zero: Recursive Self-Play in Executable Worlds for Video Reasoning},
  author = {Zhu, Jingmin and Zou, Yanyan and Li, Xiuyu and Bailey, James and
            Phung, Dinh and Yu, Tao and Hou, Yunyun and Liu, Jun and Zhou, Jun and
            Wang, Sirui and Duan, Nan and Liu, Jiachen and Ke, Qiuhong},
  year   = {2026}
}