OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe

2025-11-20

AI Summary

Multimodal reasoning research is hindered by opaque data curation and non-reproducible training strategies. To address this, we propose an open, general two-stage training paradigm: (1) supervised fine-tuning (SFT) on a rigorously validated cold-start dataset comprising 874K samples; followed by (2) policy optimization via reinforcement learning (RL) using 74K cross-domain RL samples. Our approach integrates heterogeneous, multi-source data while explicitly co-designing data quality assurance and training methodology. Evaluated across nine major multimodal reasoning benchmarks, our method achieves a +11.6% average improvement over the Qwen2.5-VL-7B-Instruct baseline. Crucially, we fully open-source all code, datasets, and end-to-end training pipelines—establishing the first systematic empirical foundation and infrastructure for reproducible, scalable multimodal reasoning research.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

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.

Problem

• Developing transparent multimodal reasoning models with reproducible training strategies
• Addressing data quality and scalable training for visual reasoning capabilities
• Creating open recipe for multimodal reasoning spanning SFT and RL stages

Innovation

• Two-stage recipe with SFT and RL
• Cold-start dataset with step-by-step validation
• Domain-diverse RL dataset for robust learning