MMR1 is a multimodal reasoning enhancement research project published at CVPR 2026, with the repository address https://github.com/LengSicong/MMR1，主要面向多模态推理模型的训练方法研究与资源开放。
[Project Background and Positioning] The project points out that existing progress in multimodal reasoning is limited by two aspects: first, the lack of open, large-scale, high-quality long chain-of-thought (CoT) data; second, unstable reinforcement learning fine-tuning, where standard GRPO often suffers from vanishing gradients under low reward variance. MMR1 is developed around these two points, providing multi-level open-source resources including methods, data, and models.
[Core Method: Variance-Aware Sampling (VAS)] The project proposes Variance-Aware Sampling (VAS), a data selection strategy guided by the Variance Promotion Score (VPS). VAS combines outcome variance and trajectory diversity to increase reward variance, stabilize policy optimization, and improve convergence. The method workflow is as follows: training starts from a dataset prompt pool, a random sampler provides uniform data coverage, and a weighted sampler prioritizes prompts with higher reward variance and more diverse trajectories based on VPS; the two sources form training batches according to a mixing ratio λ to balance exploration and coverage. The policy model generates rollouts, which are evaluated through rewards and used to update the policy; VPS scores are periodically re-estimated as the model improves, enabling dynamic adaptation. At the algorithm level, this includes: sampling multiple rollouts for each prompt to estimate pass rate, outcome variance (OVS), trajectory diversity (TDS), and VPS; refreshing statistics at specified intervals; constructing batches by mixing uniform random and VPS proportions; generating rollouts, computing the GRPO loss, and updating policy parameters.
[Open-Source Resources]
- MMR1-SFT (approximately 16M): a supervised fine-tuning dataset containing 16 million long CoT cold-start trajectories (generated by Gemini2.5 Pro/Flash), with short answers verified by GPT-4o.
- MMR1-RL (15k): a reinforcement learning dataset containing 15,000 question-answer pairs (GPT-4o).
- Model checkpoints: MMR1-3B-SFT, MMR1-3B-RL, MMR1-7B-SFT, and MMR1-7B-RL have been released; MMR1-32B-SFT and MMR1-32B-RL are marked as "On the way" (to be released soon). The dataset covers diverse domains such as mathematics, science, charts/graphs, document tables, and general understanding, including approximately 1.6 million mathematics samples and an additional approximately 37,000 samples from other domains, integrating MathVerse, ScienceQA,Public resources such as ChartQA, DocVQA, and GQA, and includes newly curated and self-collected data. The project also provides a fully reproducible end-to-end training pipeline. For training details, see TRAIN.md in the repository.
[Evaluation Results] The project is evaluated on math-related multimodal reasoning benchmarks such as MathVerse, MathVista, MathVision, LogicVista, and ChartQA. MMR1-7B-RL scores 58.4 on average, achieving a new SOTA among 7B-scale reasoning models; MMR1-3B-RL achieves 52.7 and remains competitive at a smaller scale.
[VAS Training Dynamics Analysis] In terms of training efficiency, VAS significantly amplifies gradient magnitudes relative to the baseline, alleviating vanishing gradients; a higher clipping ratio indicates that policy updates are closer to the trust region boundary; both full VAS (λ=1.0) and hybrid VAS–random sampling (λ=0.5) converge faster and achieve higher final accuracy than the baseline. In terms of VPS dynamics, the score distribution gradually shifts from relatively uniform early in training to concentration in the middle interval, indicating that the identification of persistently informative prompts is converging; the transition matrix shows that many prompts move across intervals over time, indicating that reward variance changes dynamically as the policy evolves, with broader transitions early on and more stable updates later.
[Qualitative Demonstration and Contribution Statement] The project provides qualitative examples from MathVerse, demonstrating the model's ability to analyze problems, plan structured solutions, reason step by step, verify results, and provide alternative solution paths. The project is still under active development, and contributions via pull request or issue are welcome.
[Dependencies and License] MMR1 is built on Qwen2.5VL, LLaMA-Factory, and EasyR1; see ACKNOWLEDGEMENT.md for the acknowledgements list. The project is released under the Apache 2.0 license, and the service is a research preview for non-commercial use only, subject to the Qwen model license, the terms of use for data generated by OpenAI and Gemini, and ShareGPT privacy practices.
[Limitations] The project is positioned as a research-use, non-commercial preview; the 32B RL checkpoint has not yet been released; performance data is based on the project's evaluation on the specified benchmarks, and specific use must comply with the corresponding licenses and terms.
