Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning

TL;DR

This work tackles transferring linguistic cognitive behaviors to multimodal visual reasoning by proposing a two-stage training pipeline on Qwen2.5-VL-7B: a large-scale linguistic cold-start fine-tuning followed by extensive multimodal reinforcement learning with verifiable rewards. The resulting Open-Vision-Reasoner achieves state-of-the-art open-source performance on both language and multimodal math benchmarks, and demonstrates strong visual reasoning on datasets like MathVision and MathVerse. A core contribution is the systematic analysis of visual cognitive behaviors, showing they emerge early, become refined through RL, and transfer in strategically useful ways, while also addressing perception challenges and scalability limits. The work provides datasets, training dynamics, and behavioral insights to guide future cognitively aligned multimodal agents and scaling efforts in vision-language reasoning.

Abstract

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.

Figures (11)

• Figure 1: Performance comparison with state-of-the-art models on both textual (AIME 2024, AIME 2025 balunovic_srimatharena_2025, MATH500 hendrycksmath2021) and multimodal (MathVista lu2023mathvista, MathVision wang2024measuring, MathVerse zhang2024mathversedoesmultimodalllm) math reasoning benchmarks. Open Vision Reasoner (OVR) demonstrates superior results among open-source models and performs competitively with commercial counterparts.
• Figure 2: Multiple Cognitive Behaviors in a Single Response. This case shows triggered visual-specific cognitive behaviors like visual divide-and-conquer, reflection, goal-driven visual tracing, along with the linguistic behavior backtracking.
• Figure 3: Training Dynamics. (a) The cold-start stage shows a step-wise loss decrease. (b) In the RL stage, reward (purple, left axis) and average response length (orange, right axis) grow steadily, with sharp surges after each sequence length expansion.
• Figure 4: Performance Evolution on Reasoning Benchmarks. OVR demonstrates sustained and convergent growth across both linguistic and multi-modal benchmarks throughout the cold start (left) and RL (right) training phases.
• Figure 5: Multimodal Cognitive Behavior Analysis. (a) Emergence of visual reflection across the cold start and RL training steps. (b) Emergence and transfer rates of four visual cognitive behaviors across base models and training stages. Numerical values denote the language-to-vision transfer rates for each behavior.