MRCL: Multimodal Reasoning Continual Learning

MRCL is a five-stage benchmark for studying catastrophic forgetting during continual post-training of vision-language models. It brings together recent, challenging, and reasoning-intensive multimodal datasets spanning medical understanding, navigation and planning, geometry, visual-spatial reasoning, and financial chart analysis.

The benchmark is introduced in RL Forgets! Towards Continual Policy Optimization.

Training and evaluation code is available in the CPO github repository.

Benchmark overview

MRCL is organized as a fixed continual-learning sequence:

1. MedBookVQA
2. Navigation
3. We-Math2.0
4. Puzzle
5. FinMME

Each stage provides a training split for adaptation and a held-out test split for measuring both acquisition of the current task and retention of previously learned tasks.

Stage	Task	Capability	Train	Test	Answer format
1	MedBookVQA	Medical visual understanding	~1.4K	~0.6K	Multiple choice
2	Navigation	Visual navigation and planning	~1.5K	~0.5K	Direction sequence
3	We-Math2.0	Multimodal mathematical reasoning	~5.7K	~3.7K	Numerical and multiple choice
4	Puzzle	Visual-spatial reasoning	~6.0K	~0.4K	Multiple choice
5	FinMME	Financial chart understanding	~8.0K	~1.0K	Numerical and multiple-answer choice

The use of datasets released in 2025 or later reduces the risk that benchmark tasks were already present in the pretraining data of recent models.

Split construction

• MedBookVQA: uses the Symptom Recognition and Surgery Operation subsets. Samples with parsing failures are removed, and the remaining data are randomly divided into training and test sets with a 7:3 ratio.
• Navigation: is derived from ThinkMorph. Samples are manually filtered to retain problems with a unique shortest path. Problems requiring 2-10 steps are split 7:3; those requiring 11-15 steps are reserved for testing.
• We-Math2.0: includes Plane Geometry, Analytic Geometry, and Solid Geometry from We-Math2.0-standard, together with difficulty-based splits from We-Math2.0-pro. Ground-truth mathematical expressions are normalized to standard LaTeX.
• Puzzle: uses the Jigsaw Assembly subset of ThinkMorph for training and the VisPuzzle subset for testing.
• FinMME: includes numerical questions and multiple-answer questions; numerical records provide a tolerance for chart-reading uncertainty.

Data format

The repository follows this layout:

MRCL/
├── MedBookVQA/
├── Navigation/
├── We-Math2/
├── Puzzle/
└── FinMME/

<TASK>/
├── images/
└── jsons/
    ├── train/data.json
    └── test/data.json

Each split is a JSON array in the following format:

{
  "image": "example.png",
  "conversations": [
    {
      "from": "human",
      "value": "<image>\nQuestion text"
    },
    {
      "from": "gpt",
      "value": "Ground truth"
    }
  ]
}

Image files are resolved relative to the corresponding <TASK>/images/ directory. FinMME numerical samples additionally contain a tolerance field.

Evaluation

Use the task-specific scoring rules reported with MRCL:

• MedBookVQA and Puzzle: exact-match accuracy.
• Navigation: a prediction is correct only when the complete direction sequence matches the reference shortest path.
• We-Math2.0: multi-blank answers receive partial credit; each correct blank contributes equally.
• FinMME: an incomplete but otherwise correct subset receives 0.5 credit for multiple-answer questions; any incorrect selection receives 0. Numerical answers are accepted when they fall within the provided tolerance.

For continual evaluation, let the model train on the five configurations in the specified order and evaluate all tasks seen so far after each stage. The paper reports Mean Finetune Accuracy (MFT), Mean Final Accuracy (MFN), and Mean Task Accuracy (MTA).

Citation

@article{luo2026rl,
    title={RL Forgets! Towards Continual Policy Optimization},
    author={Luo, Mao-Lin and Wang, Zhe-Xu and Zhou, Zi-Hao and Ye, Bo and Zhao, Jian and Zhang, Min-Ling and Wei, Tong},
    journal={arXiv preprint arXiv:2607.04364},
    year={2026}
}
@inproceedings{
    gu2026thinkmorph,
    title={ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning},
    author={Jiawei Gu and Yunzhuo Hao and Huichen Will Wang and Linjie Li and Michael Qizhe Shieh and Yejin Choi and Ranjay Krishna and Yu Cheng},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}
@article{yip2025medbookvqa,
    title={MedBookVQA: A Systematic and Comprehensive Medical Benchmark Derived from Open-Access Book},
    author={Yip, Sau Lai and He, Sunan and Nie, Yuxiang and Chan, Shu Pui and Ye, Yilin and Lam, Sum Ying and Chen, Hao},
    journal={arXiv preprint arXiv:2506.00855},
    year={2025}
}
@inproceedings{
    qiao2026wemath,
    title={We-Math 2.0: A Versatile MathBook System for Incentivizing Visual Mathematical Reasoning},
    author={Runqi Qiao and Qiuna Tan and Peiqing Yang and Yanzi Wang and Xiaowan Wang and Enhui Wan and Guanting Dong and Shiqiang Lang and Sitong Zhou and Yida Xu and Yuchen Zeng and Jie Wang and Chong Sun and Chen Li and Honggang Zhang},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}
@inproceedings{luo2025finmme,
    title={Finmme: Benchmark dataset for financial multi-modal reasoning evaluation},
    author={Luo, Junyu and Kou, Zhizhuo and Yang, Liming and Luo, Xiao and Huang, Jinsheng and Xiao, Zhiping and Peng, Jingshu and Liu, Chengzhong and Ji, Jiaming and Liu, Xuanzhe and others},
    booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
    pages={29465--29489},
    year={2025}
}