OpenMMReasoner: New Training Framework Enhances AI Multimodal Reasoning with Smaller, Smarter Datasets

Introduction to OpenMMReasoner

A team of researchers from MiroMind AI and several Chinese universities have unveiled OpenMMReasoner, an innovative training framework designed to elevate the multimodal reasoning abilities of large language models (LLMs). This framework employs a two-phase approach, starting with supervised fine-tuning (SFT) on a carefully curated dataset, followed by reinforcement learning (RL) to further enhance the model’s reasoning capabilities across both text and visual modalities.

Addressing Challenges in Multimodal Reasoning

Recent advancements in reinforcement learning with verifiable rewards (RLVR) have boosted LLMs’ capacity to generate chain-of-thought (CoT) reasoning, enabling better performance on complex tasks like mathematics and coding. Inspired by these successes, researchers extended RLVR methods to large multimodal models (LMMs), improving their ability to interpret and reason across different data types.

However, a significant obstacle has been the lack of transparency in training pipelines for multimodal reasoning models. Many existing studies provide limited details about dataset curation and training strategies, hindering reproducibility and deeper understanding of how reasoning is effectively built into these models.

OpenMMReasoner addresses this gap by offering a fully transparent, open-source training recipe, fostering reproducibility and clarity in model development.

The Two-Stage Training Approach of OpenMMReasoner

Stage One: Supervised Fine-Tuning (SFT)

The initial phase involves a three-step SFT pipeline. Researchers first gathered approximately 103,000 raw question-answer pairs from public datasets encompassing general visual question answering and reasoning tasks. They then performed a data distillation step, leveraging the powerful Qwen3-VL-235B-Instruct model to generate high-quality reasoning traces for selected questions, enhancing the dataset’s reasoning depth.

To further boost answer diversity, multiple verified reasoning paths were generated for each question, expanding the dataset to 583,000 samples. A final “domain mixing” step introduced mathematical reasoning data, culminating in an enriched SFT dataset totaling 874,000 examples. This meticulous curation was key to improving the model’s reasoning generality and robustness.

Stage Two: Reinforcement Learning (RL)

The second phase applies reinforcement learning on a smaller, 74,000-sample dataset drawn from scientific, mathematical, and puzzle domains. Training uses a composite reward function assessing both answer accuracy and output format consistency. To optimize efficiency, a penalty mechanism discourages “overthinking,” preventing the generation of excessively long reasoning sequences that often increase computational costs and response latency.

This RL stage fine-tunes the model to produce concise, accurate, and reliable reasoning outputs, making it practical for real-world enterprise applications.

Significant Benefits for Enterprises

Kaichen Zhang, co-author of the study, highlights the practical advantages of OpenMMReasoner for businesses aiming to avoid dependence on large closed AI systems. “A smaller open-source reasoning model allows enterprises to deploy locally, reduce latency, lower token costs associated with long chains of thought, maintain full control over their data, and fine-tune the model to specific downstream tasks,” Zhang explained.

He further suggests that companies with limited domain-specific data can enhance their datasets by increasing answer diversity and incorporating domain mixing strategies to develop models with both general-purpose reasoning and specialized expertise — all without requiring massive datasets.

Performance and Insights from OpenMMReasoner

By fine-tuning the open-source Qwen2.5-VL-7B-Instruct vision-language model using the OpenMMReasoner recipe, researchers achieved a model that consistently outperforms leading visual reasoning systems such as Open Vision Reasoner (OVR) across multiple benchmarks, including WeMath, MathVerse, and MathVista.

The SFT stage alone established a strong baseline with superior accuracy and data efficiency despite using a significantly smaller training corpus compared to other methods. The subsequent RL phase further enhanced stability and consistency, driving the model to state-of-the-art performance.

Interestingly, the model demonstrated a gradual transfer of reasoning skills from multimodal to purely textual tasks, suggesting that strengthening multimodal reasoning can also improve text-only mathematical capabilities. This cross-modal competence opens possibilities for extending these techniques to video and audio data in the future.

Furthermore, the research emphasizes the importance of token efficiency. While longer reasoning chains can boost accuracy, excessive token generation reduces efficiency and increases cost. Their findings show that limiting the “reasoning budget” can maintain or even improve accuracy while optimizing resource use — a critical factor for enterprise deployments.

Open Source and Transparency: Empowering Long-Term Independence

All components of the OpenMMReasoner workflow, including datasets, training scripts, and a trained 7B parameter model, are fully open source. This transparency enables enterprises to audit data sources, customize training pipelines for new domains, and avoid vendor lock-in.

“For business leaders concerned about hidden biases, opaque data sets, or reliance on single providers, this level of transparency is essential,” Zhang stated. “It empowers teams to validate and adapt models confidently, ensuring sustainable and independent AI development.”

Conclusion

OpenMMReasoner represents a significant step forward in AI multimodal reasoning by combining smaller, smarter datasets with a transparent, two-stage training process. It offers enterprises a practical and efficient path to develop robust AI systems capable of sophisticated reasoning across text and vision, with promising potential for future expansion into audio and video modalities.

Fonte: ver artigo original

Chrono

Chrono is the curious little reporter behind AI Chronicle — a compact, hyper-efficient robot designed to scan the digital world for the latest breakthroughs in artificial intelligence. Chrono’s mission is simple: find the truth, simplify the complex, and deliver daily AI news that anyone can understand.

More Posts