HappyWorld-Bench
HappyWorld-Bench finds reliability gaps in video, spatial, and embodied world models under interaction.
HappyWorld-Bench evaluates whether generated worlds stay reliable under exploration, interaction, and modification, using six hierarchical capabilities across video, spatial, and embodied tracks. It includes 1,138 video prompts, 300 spatial scenes, and 254 embodied test cases, with HappyWorld-Arena producing human A/B Elo ratings alongside automated behavioral metrics. Evaluations of 14 video, 9 spatial, and 8 embodied systems show consistency failures: spatial models top out at 70.14% placement accuracy and 73.33% edit execution, and embodied models struggle to preserve state across multi-step actions.
- Six capabilities span construction through unified world modeling.
- Tracks cover 1,138 video prompts, 300 spatial scenes, and 254 embodied cases.
- Arena Elo ratings complement automated behavioral-correctness metrics.
- 14 video, 9 spatial, and 8 embodied systems were evaluated.
- Spatial models reach at best 70.14% placement and 73.33% edit accuracy.
- Products
- HappyWorld-BenchHappyWorld-Arena
From the original article
1 min read at huggingface.co
Evaluating world models requires assessing both the quality of the worlds they generate and their consistency and responsiveness under exploration, interaction, and modification. We introduce HappyWorld-Bench, a comprehensive benchmark that evaluates whether generated worlds remain reliable as agents interact with them.
By Zhiqi Bai, Junai Cai, Yixin Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.