Paper 2511.16334
OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 31
- GitHub
- 165 stars
01 In brief
Summary
OpenMMReasoner introduces a fully transparent two-stage recipe for training large multimodal reasoning models (LMRMs), covering supervised fine-tuning (SFT) and reinforcement learning (RL).
The SFT stage uses an 874K-sample cold-start dataset built from 103K raw questions, distilled with a strong teacher model (Qwen3-VL-235B-Instruct) and scaled via ×8 answer sampling to 583K verified traces, then mixed with math data (MMR1, MiroMind-M1).
Key findings include that answer diversity, teacher selection, and cross-domain mixing improve performance, while over-filtering is detrimental.
The RL stage uses a 74K-sample dataset and compares GSPO, GRPO, and DAPO, finding GSPO most stable and efficient; a composite reward (accuracy + format) with λ_fmt=0.1 and ×16 rollouts yield best results.
The final model outperforms baselines, achieving an 11.6% average improvement over Qwen2.5-VL-7B-Instruct across nine benchmarks, with notable gains on MathVista (79.5), MathVerse (63.8), and WeMath (79.0).
Textual reasoning also improves alongside multimodal gains, and the model is more token-efficient than OpenVisionReasoner.
All code, data, and pipelines are open-sourced.
02 From the paper
Abstract
Recent advancements in large reasoning models have fueled growing interest in extending such capabilities to multimodal domains. However, despite notable progress in visual reasoning, the lack of transparent and reproducible data curation and training strategies remains a major barrier to scalable research. In this work, we introduce OpenMMReasoner, a fully transparent two-stage recipe for multimodal reasoning spanning supervised fine-tuning (SFT) and reinforcement learning (RL). In the SFT stage, we construct an 874K-sample cold-start dataset with rigorous step-by-step validation, providing a strong foundation for reasoning capabilities. The subsequent RL stage leverages a 74K-sample dataset across diverse domains to further sharpen and stabilize these abilities, resulting in a more robust and efficient learning process. Extensive evaluations demonstrate that our training recipe not only surpasses strong baselines but also highlights the critical role of data quality and training design in shaping multimodal reasoning performance. Notably, our method achieves a 11.6% improvement over the Qwen2.5-VL-7B-Instruct baseline across nine multimodal reasoning benchmarks, establishing a solid empirical foundation for future large-scale multimodal reasoning research. We open-sourced all our codes, pipeline, and data at https://github.com/EvolvingLMMs-Lab/OpenMMReasoner.