Abstract

The remarkable reasoning capability of large language models (LLMs) stemsfrom cognitive behaviors that emerge through reinforcement with verifiablerewards. This work investigates how to transfer this principle to MultimodalLLMs (MLLMs) to unlock advanced visual reasoning. We introduce a two-stageparadigm 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 workreveals three fundamental insights: 1) Behavior transfer emerges surprisinglyearly in cold start due to linguistic mental imagery. 2) Cold start broadlymemorizes visual behaviors, while RL critically discerns and scales upeffective patterns. 3) Transfer strategically favors high-utility behaviorssuch 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. Werelease our model, data, and training dynamics to catalyze the development ofmore capable, behavior-aligned multimodal reasoners.