World-Representation-Lab/World-Embedding-Dynamics-Retrieval

World Embedding Dynamics Retrieval

World Embedding Dynamics Retrieval is the dynamics retrieval split of the World Embedding Benchmark. It contains 1,900 simulation videos from 19 physics families. The family shards are loaded together as the default configuration.

Usage

from datasets import load_dataset

dataset = load_dataset(
    "World-Representation-Lab/World-Embedding-Dynamics-Retrieval",
    split="test",
)

Fields

• query_id: unique text-query identifier.
• case_id: simulation case identifier.
• raw_text: structured physical description.
• parsed_text: natural-language retrieval caption.
• video: simulation video.

The benchmark uses parsed_text for text-video retrieval.

Citation

@misc{liu2026worldembeddingbenchmark,
  title         = {World Embedding Benchmark},
  author        = {Yiqi Liu and Ruifeng Yuan and Yang Wang and Long Li and Fengyu Cai and Hou Pong Chan and Jialin Yu and Hao Zhang and Chenghua Lin and Chenghao Xiao},
  year          = {2026},
  eprint        = {2610.03632},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.03632}
}