CoWorld-Reasoner-Dev

Overview

CoWorld-Reasoner-Dev contains the 120 development examples used for harness co-evolution in CoWorld Reasoner, with 244 images:

• MMSI-Bench: 30 examples
• SpinBench: 30 examples
• CausalSpatial: 30 examples
• PerceptionBench: 30 examples

These examples are excluded from the corresponding benchmark evaluations. The train split contains the development data, not the evaluation sets.

Data Fields

Each example contains:

• sample_id: the example identifier.
• benchmark: the source benchmark.
• task_type: the task category.
• answer_type: the answer format.
• question: the question, including options when present.
• options: structured options when available.
• images: an ordered list of images embedded in the Parquet file.
• image_paths: the corresponding relative paths used by the code.
• answer: the reference answer.
• score_spec: the scoring method.

Quick Start

Load the dataset

from datasets import load_dataset

ds = load_dataset("ModalityDance/CoWorld-Reasoner-Dev", split="train")
example = ds[0]

print(example["question"])
print(example["answer"])
print("Number of images:", len(example["images"]))
example["images"][0].show()

Use with CoWorld Reasoner

After installing the code repository, run:

python scripts/prepare/download.py

This saves questions and images to data/dev120/ and reference answers to scorer_bundle/data/dev120_evaluator_only.jsonl.

Acknowledgements

The examples are drawn from MMSI-Bench, SpinBench, CausalSpatial, and PerceptionBench. The source datasets retain their original licenses.

Citation

Please cite our paper as:

@misc{cheng2026coworldreasoner,
  title  = {CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness},
  author = {Dongjie Cheng and Ruifeng Yuan and Ran Wei and Ning Song and Yongqi Li and Wenjie Li},
  year   = {2026},
  eprint = {2610.21910},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url    = {https://arxiv.org/abs/2610.21910}
}