Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , , , , , , , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866911164619816960 |
|---|---|
| author | Wei, Yana Zhao, Liang Sun, Jianjian Lin, Kangheng Yin, Jisheng Hu, Jingcheng Zhang, Yinmin Yu, En Lv, Haoran Weng, Zejia Wang, Jia Han, Chunrui Peng, Yuang Han, Qi Ge, Zheng Zhang, Xiangyu Jiang, Daxin Patel, Vishal M. |
| author_facet | Wei, Yana Zhao, Liang Sun, Jianjian Lin, Kangheng Yin, Jisheng Hu, Jingcheng Zhang, Yinmin Yu, En Lv, Haoran Weng, Zejia Wang, Jia Han, Chunrui Peng, Yuang Han, Qi Ge, Zheng Zhang, Xiangyu Jiang, Daxin Patel, Vishal M. |
| contents | The remarkable reasoning capability of large language models (LLMs) stems from cognitive behaviors that emerge through reinforcement with verifiable rewards. This work investigates how to transfer this principle to Multimodal LLMs (MLLMs) to unlock advanced visual reasoning. We introduce a two-stage paradigm built on Qwen2.5-VL-7B: a massive linguistic cold-start fine-tuning, followed by multimodal reinforcement learning (RL) spanning nearly 1,000 steps, surpassing all previous open-source efforts in scale. This pioneering work reveals three fundamental insights: 1) Behavior transfer emerges surprisingly early in cold start due to linguistic mental imagery. 2) Cold start broadly memorizes visual behaviors, while RL critically discerns and scales up effective patterns. 3) Transfer strategically favors high-utility behaviors such as visual reflection. Our resulting model, Open-Vision-Reasoner (OVR), achieves state-of-the-art performance on a suite of reasoning benchmarks, including 95.3% on MATH500, 51.8% on MathVision and 54.6% on MathVerse. We release our model, data, and training dynamics to catalyze the development of more capable, behavior-aligned multimodal reasoners. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_05255 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning Wei, Yana Zhao, Liang Sun, Jianjian Lin, Kangheng Yin, Jisheng Hu, Jingcheng Zhang, Yinmin Yu, En Lv, Haoran Weng, Zejia Wang, Jia Han, Chunrui Peng, Yuang Han, Qi Ge, Zheng Zhang, Xiangyu Jiang, Daxin Patel, Vishal M. Computer Vision and Pattern Recognition Computation and Language The remarkable reasoning capability of large language models (LLMs) stems from cognitive behaviors that emerge through reinforcement with verifiable rewards. This work investigates how to transfer this principle to Multimodal LLMs (MLLMs) to unlock advanced visual reasoning. We introduce a two-stage paradigm built on Qwen2.5-VL-7B: a massive linguistic cold-start fine-tuning, followed by multimodal reinforcement learning (RL) spanning nearly 1,000 steps, surpassing all previous open-source efforts in scale. This pioneering work reveals three fundamental insights: 1) Behavior transfer emerges surprisingly early in cold start due to linguistic mental imagery. 2) Cold start broadly memorizes visual behaviors, while RL critically discerns and scales up effective patterns. 3) Transfer strategically favors high-utility behaviors such as visual reflection. Our resulting model, Open-Vision-Reasoner (OVR), achieves state-of-the-art performance on a suite of reasoning benchmarks, including 95.3% on MATH500, 51.8% on MathVision and 54.6% on MathVerse. We release our model, data, and training dynamics to catalyze the development of more capable, behavior-aligned multimodal reasoners. |
| title | Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning |
| topic | Computer Vision and Pattern Recognition Computation and Language |
| url | https://arxiv.org/abs/2507.05255 |