Vision‑R1 Multimodal Reasoning Model Delivers Human‑Level Logic and Near‑OpenAI O1 Accuracy

Background and Motivation

Complex reasoning remains a bottleneck for large language models (LLMs), and the problem intensifies when visual modalities are added. Existing text‑only chain‑of‑thought (CoT) methods such as "Chain‑of‑Thought" and "Tree‑of‑Thought" improve performance on textual tasks but ignore visual information, leading to poor results on image‑text reasoning tasks such as geometry problems with diagrams.

Directly applying reinforcement learning (RL) to multimodal models also fails because of data scarcity and the tendency of models to generate overly long, error‑prone reasoning chains—a phenomenon we call the over‑thinking optimization problem .

Key Challenges

Data scarcity: High‑quality multimodal CoT data are extremely rare, and manual annotation is costly.

Optimization difficulty: RL on a cold‑started multimodal model either produces short, simplistic chains or, after prolonged training, generates redundant and incorrect steps.

Vision‑R1: Proposed Solution

We introduce Vision‑R1 , a multimodal reasoning LLM that tackles both challenges through three innovations:

Unsupervised high‑quality CoT data generation: Using a Modality Bridging pipeline, we first prompt an existing multimodal model to produce structured pseudo‑reasoning chains from image‑text pairs. These pseudo‑chains are then fed back to the model to generate detailed textual descriptions, effectively converting visual information into text that can be captured by a language model. The pipeline yields a 200K dataset called Vision‑R1‑cold , characterized by human‑like "questioning‑reflection" reasoning patterns.

Progressive Thinking Suppression Training (PTST): Inspired by human cognitive development, PTST constrains reasoning length in early training epochs, forcing the model to internalize core logic before gradually allowing longer chains. This mitigates over‑thinking while preserving accuracy.

Hard Formatting Result Reward Function (HFRRF): Combined with Group Relative Policy Optimization (GRPO), HFRRF rewards only those outputs that satisfy both correct format and correct answer, encouraging "key‑node verification" reasoning.

Experimental Evaluation

On the MathVista benchmark, the 7B Vision‑R1 model achieves 73.5% accuracy , only 0.4% behind OpenAI’s O1. Detailed sub‑task results are 80.3% (Geometry Reasoning), 79.0% (Algebraic Reasoning), and 83.2% (Geometry Problem Solving), all surpassing the baseline Qwen‑2.5‑VL‑7B.

Scaling to 32B/72B parameters further improves performance, and on the MM‑Math benchmark Vision‑R1‑7B ranks just behind the 10× larger Qwen‑2.5‑VL‑72B.

Data Quality Analysis

Compared with Mulberry (260K) and llava‑cot (100K), Vision‑R1‑cold shows a 3‑5× increase in cognitive elements such as questioning, reflection, and inspection. Fine‑tuning a Llama‑3.2‑11B‑V‑Instruct base model with Vision‑R1‑cold yields substantial gains on both general and mathematical benchmarks over traditional pseudo‑CoT datasets.

Training Dynamics

Thought compression effect: PTST shortens reasoning steps in early epochs while accuracy rises sharply.

Progressive generalization: As training proceeds, the model expands reasoning length without sacrificing core logic, achieving complex yet correct inference.

Ablation Study

We compare four training strategies:

Vision‑R1‑Zero (pure RL): Lacks high‑quality initialization, resulting in short, simplistic chains and limited accuracy.

Vision‑R1‑CI (cold‑start only): Generates long chains but with many redundant errors, hurting overall performance.

Vision‑R1‑Long (cold‑start + RL, no PTST): Optimization is unstable and accuracy fluctuates.

Vision‑R1 (cold‑start + RL + PTST): Dynamically adjusts reasoning depth, delivering the best balance of complexity and correctness.

The ablation confirms that the combination of cold‑start data and PTST yields the optimal efficiency‑accuracy trade‑off, offering a new paradigm for multimodal RL training.

Implications

Vision‑R1 demonstrates that a 7B model can match the reasoning ability of 70B‑plus commercial systems when equipped with high‑quality unsupervised CoT data and progressive training. The approach paves the way for multimodal models to transition from "perception‑reproduction" to "thought‑emergence," a crucial step toward Artificial General Intelligence.