PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning

Published in arXiv preprint, 2026

Links: arXiv ยท PDF

PRISM evaluates programmatic video generation through code, focusing on whether models can produce spatially correct animated outputs rather than merely executable programs.

Highlights

  • Provides 10,372 human-calibrated instruction-code pairs across English and Chinese prompts.
  • Covers real-world knowledge visualization scenarios spanning hundreds of subject categories.
  • Introduces a funnel-style evaluation framework for executability, spatial correctness, dynamic visual complexity, and temporal activity.

Recommended citation: Qiran Zhang, Yuheng Wang, Runde Yang, Lin Wu, Jingru Fan, Shu Yao, Jie Zhang, Tianle Zhou, Huatao Li, Ruijie Shi, Yihan Li, and Chen Qian. (2026). "PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning." arXiv:2605.19382.
Download Paper