VersaCoT-10K: A Multi-Domain Visual Chain-of-Thought Dataset

VersaCoT-10K is a high-quality, multi-domain visual Chain-of-Thought (CoT) dataset introduced in From Stages to States: Unified Iterative Reasoning for MLLMs.

It contains 10,000 single-image multiple-choice visual question-answering samples spanning eight diverse reasoning domains. Each sample provides visual inputs, questions, candidate options, ground-truth answers, and high-quality CoT annotations, supporting both multimodal reasoning and reasoning-state refinement.

Dataset Overview

Property	Description
Total Samples	10,000
Training / Validation / Test	7,000 / 1,000 / 2,000
Reasoning Domains	8 (1,250 samples each)
Task Format	Single-image multiple-choice VQA
Annotations	Image, question, options, caption, answer, and CoT

Eight Reasoning Domains:

1. Medical Reasoning (PMC-VQA)
2. Driving and Safety (DrivingVQA)
3. Visual Commonsense (Visual7W)
4. Scientific Diagrams (AI2D)
5. Spatial Perception (CV-Bench)
6. Geometry and Mathematics (GeoQA)
7. Chart Computation (ChartQA)
8. Multidisciplinary Knowledge (MMMU)

Construction and Quality Control

VersaCoT-10K is constructed by standardizing samples from existing public benchmarks. CoT annotations are generated using Gemini-3-Flash and undergo rule-based filtering and multimodal quality assessment using Gemini-3.5-Flash, ensuring visual grounding, reasoning correctness, and answer consistency.

Related Resources

• Paper: From Stages to States: Unified Iterative Reasoning for MLLMs
• Code: ReState GitHub Repository

Citation

If you find VersaCoT-10K useful in your research, please cite:

@article{yu2026stages,
  title={From Stages to States: Unified Iterative Reasoning for MLLMs},
  author={Yu, Chuang and Ma, Xuefei and Feng, Yuanhao and Zeng, Wuxiong and Wang, Bo and Cai, Minghong and Wu, Dongming and Zhou, Fei and Tang, Chen and Tong, Wenwen and others},
  journal={arXiv preprint arXiv:2610.15620},
  year={2026}
}