YuChuang1205/VersaCoT-10K-RAD-Version

VersaCoT-10K-RAD is the rationale-augmented version of VersaCoT-10K, developed for multimodal Chain-of-Thought (MCoT) reasoning and reasoning-state refinement. It builds upon the Rationale Augmentation and Discrimination (RAD) paradigm introduced in MIND (ICML 2026) and is used in the ReState framework.

Dataset Overview

VersaCoT-10K contains 10,000 single-image multiple-choice VQA samples across eight visual reasoning domains:

• Medical Reasoning (PMC-VQA)
• Driving and Safety (DrivingVQA)
• Visual Commonsense (Visual7W)
• Scientific Diagrams (AI2D)
• Spatial Perception (CV-Bench)
• Geometry and Mathematics (GeoQA)
• Chart Computation (ChartQA)
• Multidisciplinary Knowledge (MMMU)

Each domain contains 1,250 samples, with the original dataset split into 7,000 training, 1,000 validation, and 2,000 test samples.

RAD Augmentation

Following the RAD paradigm proposed in MIND, VersaCoT-10K-RAD extends the original CoT annotations with:

• Positive Rationales (R⁺): Diverse reasoning chains that preserve correct reasoning semantics and conclusions.
• Hard Negative Rationales (R⁻): Plausible but logically incorrect reasoning chains that introduce challenging reasoning errors.
• Empty Reasoning States (φ): Initial states used to learn reasoning construction.

These augmented reasoning states provide supervision for reasoning construction, correct reasoning preservation, and erroneous reasoning correction, supporting unified reasoning-state learning and iterative self-refinement in ReState.

Related Resources

• ReState Paper: From Stages to States: Unified Iterative Reasoning for MLLMs
• MIND Paper (ICML 2026): Multi-rationale INtegrated Discriminative Reasoning Framework for Multi-modal Large Models
• Code: ReState GitHub
• Original Dataset: VersaCoT-10K
• RAD Dataset: VersaCoT-10K-RAD-Version

Citation

If you find this dataset useful in your research, please cite both ReState and MIND, which introduce VersaCoT-10K and the RAD augmentation paradigm, respectively.

ReState — Paper | Google Scholar

@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 Lu, Lewei and Yue, Xiangyu},
  journal={arXiv preprint arXiv:2610.15620},
  year={2026}
}

MIND (ICML 2026) — Paper | Google Scholar

@inproceedings{pmlr-v306-yu26bv,
  title={{MIND}: Multi-rationale {IN}tegrated Discriminative Reasoning Framework for Multi-modal Large Models},
  author={Yu, Chuang and Zhao, Jinmiao and Zhao, Mingxuan and Liu, Yunpeng and Shu, Xiujun and Feng, Yuanhao and Wang, Bo and Yue, Xiangyu},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  pages={151737--151756},
  year={2026},
  volume={306},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  url={https://proceedings.mlr.press/v306/yu26bv.html}
}