Evaluation criteria remain unchanged; supplementary information on the evaluation settings
TL;DR. Except for the approach announced in the first notice, all metrics remain consistent with their original definitions. This announcement provides supplementary information on three evaluation settings: the previously announced Motion Quality score constraint, a Ground Truth video resolution of 640 × 480, and a frame-count upper bound based on the corresponding Ground Truth sequence. Each team will need to submit a new final ZIP package.
WorldArena was created to provide a unified and reproducible evaluation framework for world models that understand actions, predict changes in environmental state, and support embodied tasks. Track 1 evaluates video generation in embodied-intelligence scenarios.
This announcement provides supplementary information on the following three evaluation settings:
- Motion Quality score constraint. The mitigation for Dynamic Degree, Flow Score, and Motion Smoothness described in the Motion Quality evaluation update remains in effect.
- Ground Truth resolution. Ground Truth videos used in the final evaluation have a resolution of 640 × 480. A submitted-video resolution of 640 × 480 or higher is recommended.
- Frame-count upper bound. The actual frame count of each submitted video does not exceed that of the Ground Truth video corresponding to that video, whose frame count represents the length of the action sequence. Any frame count from 0 up to this upper bound is accepted; no fixed frame count is required.
Final evaluation protocol: except for the approach announced in the first notice, all metrics remain consistent with their original definitions. The final protocol therefore consists of the originally published evaluation version together with the approach announced in the Motion Quality Evaluation Update.
To avoid ambiguity between materials submitted at different stages, intermediate evaluation results, and the final protocol, all teams are asked to submit a new final ZIP package. This gives each team an opportunity to clearly confirm the exact set of videos it wishes to enter into the final evaluation, establishes a single and traceable final-submission record, prevents earlier materials from being used by mistake, and better protects the interests of participating teams and the fairness of the evaluation.
The previous leaderboard will remain publicly available as an interim evaluation result. It will not receive further updates, while the final leaderboard will be produced under the protocol described in this announcement.
评价标准保持不变，评测设置补充说明
简要结论。 除第一次公告公布的方案外，所有指标均与最初的指标定义保持一致。本公告对三项评测设置作补充说明：此前公布的 Motion Quality 分数限制方案、Ground Truth 视频分辨率为 640 × 480，以及按照对应 Ground Truth 动作序列设置视频帧数上限。每个团队均需要重新提交最终 ZIP 压缩包。
WorldArena 建立的初衷，是通过一套统一、可复现的评测体系，选出能够理解动作、预测环境状态变化，并真正服务于具身任务的世界模型。Track 1 重点评估世界模型在具身智能场景下的视频生成能力。
本公告对以下三项评测设置作补充说明：
- Motion Quality 分数限制。 继续采用此前《Motion Quality 指标调整说明》中公布的方案，对 Dynamic Degree、Flow Score 和 Motion Smoothness 三项指标进行处理。
- Ground Truth 分辨率。 最终评测使用的 Ground Truth 视频分辨率为 640 × 480，提交视频推荐采用 640 × 480 或以上分辨率。
- 视频帧数上限。 每个提交视频的实际帧数不超过该视频所对应的 Ground Truth 视频帧数，该帧数代表相应动作序列的长度。0 至该上限范围内的帧数均可接受，不要求使用固定帧数。
最终评测方案：除第一次公告公布的方案外，所有指标均与最初的指标定义保持一致。最终采用的评测方案为最初公开的评测版本，并结合《Motion Quality 指标调整说明》中公布的处理方案。
为避免不同阶段的提交材料、中间评测结果与最终评测规则相互混淆，本轮统一请各团队重新提交最终 ZIP 压缩包。这样既能让每个团队明确确认希望用于最终评测的完整视频集合，也能为最终评测建立唯一、清晰且可追溯的提交依据，避免误用此前材料，更好地保障参赛团队的权益与评测公平。
此前的榜单将继续公开保留，可作为阶段性评测结果。该榜单不再继续更新，最终榜单将按照本公告所述方案产生。