Update README.md

README.md

@@ -10,28 +10,27 @@ base_model:
 ![title](./assets/title.png)
 
 <div align="center">
-<a href="https://arxiv.org/abs/2506.12000" target="blank">
+<a href="https://github.com/linkangheng/Open-Vision-Reasoner/tree/main/paper/Open-Vision-Reasoner.pdf" target="blank" style="margin-right: 10px;">
     <img alt="arXiv" src="https://img.shields.io/badge/arXiv-OVR-red?logo=arxiv" height="20" />
-</a> <a href="https://huggingface.co/ovr" target="blank">
+</a><a href="https://huggingface.co/collections/Kangheng/ovr-686646849f9b43daccbe2fe0" target="blank" style="margin-right: 10px;">
     <img alt="HF Model: OVR" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Model-OVR-fb8740?&logoColor=white" height="20" />
-</a> <a href="https://huggingface.co/datasets/ovr" target="blank">
-    <img alt="Dataset: OVR" src="https://img.shields.io/badge/%F0%9F%97%84%EF%B8%8F%20_Dataset-OVR-48b9d0?&logoColor=white" height="20" />
-</a> <a href="https://huggingface.co/spaces/ovr/demo" target="blank">
-    <img alt="Demo: OVR" src="https://img.shields.io/badge/%F0%9F%9A%80%20_Demo-OVR-9368AB?&logoColor=white" height="20" />
+</a><a href="" target="blank" style="margin-right: 10px;">
+    <img alt="Dataset: OVR" src="https://img.shields.io/badge/%F0%9F%97%84%EF%B8%8F%20_Dataset(coming)-OVR-48b9d0?&logoColor=white" height="20" />
 </a>
 </div>
 
+### Overview
 ![preview](./assets/preview.png)
 
-**OVR-7B-ColdStart** is the foundational checkpoint from the **Open Vision Reasoner (OVR)** project, trained on over 2 million text-based reasoning samples to learn powerful cognitive behaviors that naturally transfer to visual tasks.
+The remarkable reasoning capbility of Large Language Models (LLMs) stems from cognitive behaviors that emerge when reinforcing against verifiable rewards. This work investigates how to transfer this principle to Multimodal LLMs (MLLMs) to unlock **advanced visual reasoning**.
 
-This model demonstrates a key finding: **linguistic cognitive patterns learned from text can effectively generalize to multimodal reasoning**, serving as the foundation for the final visual reasoning model `OVR-7B-RL`.
+We introduce a two-stage paradigm built on Qwen2.5-VL-7B: a massive **text-only cold-start fine-tuning**, followed by **multimodal reinforcement learning** (RL) spanning nearly 1,000 steps—surpassing all prior open-source efforts in scale. This pioneering work reveals three fundamental insights: 
 
-### Model Highlights
+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. 
 
-- **Base Model:** Built on `Qwen2.5-VL-7B`.
-- **Training:** Pure text cold-start fine-tuning to establish core cognitive reasoning patterns
-- **Transfer Learning:** Cognitive behaviors acquired from language reasoning successfully transfer to visual tasks.
+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.
 
 ### Model Card
 
@@ -44,6 +43,17 @@ This model demonstrates a key finding: **linguistic cognitive patterns learned f
 ![performance-text](./assets/performance-text.png)
 ![performance-vision](./assets/performance-vision.png)
 
+### Training Dynamics and Performance Evolution
+
+<p align="center">
+  <img width="45%" src="assets/coldstart_dynamic.png">
+  <img width="45%" src="assets/rl_dynamics.png">
+</p>
+
+<p align="center">
+  <img width="100%" src="assets/performance.png">
+</p>
+
 ###  Model Deployment
 
 ```bash