State-of-the-Art: Models trained on this dataset achieve SOTA performance in their size class
🧠 Model Training
Based on the MMFineReason dataset, we train a family of multimodal reasoning models at 2B / 4B / 8B scales, all initialized from the corresponding Qwen3-VL-Instruct backbones and fine-tuned using a unified data-centric training recipe.
Each MMFineReason model is trained in two stages:
Supervised Fine-Tuning (SFT) on MMFineReason-1.8M-SFT, leveraging long-form, visually grounded Chain-of-Thought (CoT) annotations with an average length of 2,910 tokens.
Reinforcement Learning (RL) using GSPO, applied on MMFineReason-1.8M-RL to further improve reasoning reliability and generalization.
📊 Model Performance
Main Results
Main Benchmark Results
Comparison of MMFineReason models with state-of-the-art models.
MMFineReason-4B surpasses Qwen3-VL-8B-Thinking (73.9 vs 72.5), while MMFineReason-8B outperforms the larger Qwen3-VL-30B-A3B-Thinking (75.7 vs 74.5) and exceeds Gemini-2.5-Flash. On mathematical benchmarks, MFR-8B achieves 83.4% on DynaMath (vs Qwen3-VL-32B-Thinking's 82.0%) and 67.1% on MathVision, outperforming HoneyBee-8B and OMR-7B by 23-30 points. Despite minimal chart training data, MFR-8B generalizes well to CharXiv (90.8%) and RealWorldQA (75.6%).
SFT vs RL Training Analysis
SFT vs RL Results
Results comparing MFR-SFT and MFR-Thinking models against base Qwen3-VL variants.
SFT drives major gains in mathematical reasoning (e.g., MathVision: 53.9% → 67.6% for 8B). RL enhances generalization on understanding benchmarks (e.g., AI2D: 78.5% → 82.5% for 2B) while showing variance on math benchmarks.
1@article{lin2026mmfinereason,
2 title={MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods},
3 author={Lin, Honglin and Liu, Zheng and Zhu, Yun and Qin, Chonghan and Lin, Juekai and Shang, Xiaoran and He, Conghui and Zhang, Wentao and Wu, Lijun},
4 journal={arXiv preprint arXiv:2601.21821},
5 year={2026},
6 url={https://mmfinereason.github.io/}
7}
📄 License
This dataset is released under the Apache 2.0 License. Individual source datasets may have their own licenses.
🤝 Acknowledgments
We thank the creators of FineVision, MMR1, BMMR, Euclid30K, GameQA-140K, LLaVA-CoT, WeMath, ViRL39K, and others. We also thank the Qwen team for the powerful Qwen3-VL series models.
