🎯 Vision-Zero

Scalable VLM Self-Improvement via Strategic Gamified Self-Play

A domain-agnostic framework enabling VLM self-improvement through competitive visual games

📋 Table of Contents

• 🎯 Overview
• 📊 Performance Results
• 🚀 Quick Start
• 🤖 Models & Dataset
• 🛠️ Setup
• 💪 Training
• 📊 Evaluation
• 📄 Citation

---

🎯 Overview

Although reinforcement learning (RL) can effectively enhance the reasoning capabilities of vision–language models (VLMs), current methods remain heavily dependent on labor-intensive datasets that require extensive manual construction and verification, leading to extremely high training costs and consequently constraining the practical deployment of VLMs.

To address this challenge, we propose Vision-Zero, a domain-agnostic framework enabling VLM self-improvement through competitive visual games generated from arbitrary image pairs.

🏆 Achievement: Despite using label-free data, Vision-Zero achieves state-of-the-art performance on reasoning, chart question answering, and vision-centric understanding tasks, surpassing other annotation-based methods.

---

📊 Performance Results

🎖️ Key Achievements

Metric	Result	Significance
Training Data	0 human annotations	First zero-human-in-the-loop VLM training paradigm
vs SOTA (ViGaL)	+3.4% average	Outperforms best curated-data method
vs GPT-4o	Competitive	Matches/exceeds on chart & OCR tasks
InfoVQA	+4.2% gain	Largest improvement on chart understanding
MMVP	+2.7% gain	Enhanced visual perception capabilities
Generalization	3 domains	Works on CLEVR, Charts, Real-World images

📊 OCR and Chart Understanding Tasks

Vision-Zero demonstrates exceptional performance on OCR, chart, and document understanding tasks, with the Chart-trained variant showing particularly strong improvements:

Model	AI2D	ChartQA	TextVQA	DocVQA	InfoVQA	OCRBench	SEEDBench2	Average
Qwen2.5-VL-7B (Baseline)	84.7	86.1	85.5	94.8	82.3	88.3	70.4	84.6
+ Vision-Zero (CLEVR)	84.5	86.3	85.3	94.9	82.5	88.1	69.5	84.4
+ Vision-Zero (Chart)	85.8	87.2	86.4	95.9	86.5	89.0	70.9	85.9
+ Vision-Zero (Real-World)	84.8	86.3	85.4	95.2	82.3	88.5	69.8	84.6
GPT-4o	84.4	85.7	82.2	91.1	78.2	73.9	72.0	81.1

Key Insight: Vision-Zero (Chart) achieves +4.2% improvement on InfoVQA and outperforms GPT-4o on most benchmarks, demonstrating that label-free self-play can surpass both supervised baselines and proprietary models.

🎯 Vision-Centric Understanding Tasks

Vision-Zero shows consistent improvements on tasks requiring fine-grained visual reasoning:

Model	RealWorldQA	MMVP	MMStar	BLINK	MuirBench	CRPE	Average
Qwen2.5-VL-7B (Baseline)	68.1	76.8	64.6	55.2	58.2	76.4	66.6
+ Vision-Zero (CLEVR)	68.5	79.2	65.2	57.2	59.4	76.9	67.7
+ Vision-Zero (Chart)	68.2	77.9	64.7	56.1	58.6	76.2	66.9
+ Vision-Zero (Real-World)	68.5	79.5	65.8	57.5	59.8	77.0	68.0
GPT-4o	75.4	86.3	-	68.0	68.0	-	-

Improvements over baseline:

• MMVP: +2.7% (Real-World) - Better visual perception
• BLINK: +2.3% (Real-World) - Enhanced visual reasoning
• Average: +1.4% gain with zero human annotations

Remarkable Achievement: Vision-Zero (Real-World) achieves 68.0% average across vision-centric tasks, representing the best performance among all 7B open-source models trained with reasoning-focused methods.

🚀 Comparison with State-of-the-Art Methods

Vision-Zero outperforms other recent VLM reasoning methods trained on Qwen2.5-VL-7B:

Method	Training Data	AI2D	ChartQA	RealWorldQA	MMVP	Average
R1-OneVision-7B	Curated	82.2	-	58.0	61.3	-
MM-Eureka-Qwen-7B	Curated	84.1	77.3	66.1	74.3	75.5
VLAA-Thinker-7B	Curated	84.0	84.3	65.4	71.6	76.3
OpenVLThinker-7B	Curated	81.8	-	60.2	71.3	-
ViGaL-Snake+Rotation	Curated	84.5	79.9	66.5	74.6	76.4
Vision-Zero (CLEVR)	Label-Free	84.5	86.3	68.5	79.2	79.6
Vision-Zero (Chart)	Label-Free	85.8	87.2	68.2	77.9	79.8
Vision-Zero (Real-World)	Label-Free	84.8	86.3	68.5	79.5	79.8

Zero-Annotation Superiority: Despite using no human-labeled data, Vision-Zero variants achieve +3.4% average improvement over the best competing method (ViGaL), demonstrating the power of strategic self-play for VLM training.

---

✨ Key Features

🎮 Strategic Self-Play Framework

Vision-Zero trains VLMs in "Who Is the Spy"-style games, where the models engage in strategic reasoning and actions across multiple roles. Through interactive gameplay, models autonomously generate their training data without human annotation.

🖼️ Gameplay from Arbitrary Images

Unlike existing gamified frameworks, Vision-Zero can generate games from arbitrary images, thereby enhancing the model's reasoning ability across diverse domains and showing strong generalization to different tasks. We demonstrate this versatility using three distinct types of image datasets: CLEVR-based synthetic scenes, charts, and real-world images.

📈 Sustainable Performance Gain

We introduce Iterative Self-Play Policy Optimization (Iterative-SPO), a novel training algorithm that alternates between Self-Play and reinforcement learning with verifiable rewards (RLVR), mitigating the performance plateau often seen in self-play-only training and achieving sustained long-term improvements.

🎉 Current Release Status

Component	Status	Description
🤖 Models	✅ Available	Trained models on Qwen2.5-VL-7B, InternVL3-8B, InternVL3-14B
📊 CLEVR Dataset	✅ Available	Complete CLEVR-based training dataset
🛠️ Training Code	✅ Available	Full open-source training pipeline
📈 Chart Dataset	✅ Available	Chart-based dataset for enhanced reasoning
🌍 Real-World Dataset	✅ Available	Real-world image dataset for diverse scenarios

---

🚀 Quick Start

# 1. Clone the repository
git clone https://github.com/wangqinsi1/Vision-Zero.git
cd vision-zero

# 2. Set up environment
conda create -n vision-zero python=3.10
conda activate vision-zero
bash setup.sh

# 3. Download a Trained model
# Choose from available models in the table below

# 4. Start training or inference
bash run_scripts/run_grpo_vision_zero.sh

🤖 Models & Dataset

🔬 Trained Models

Model Family	Size	Dataset	HuggingFace Link
Qwen2.5-VL	7B	CLEVR
Qwen2.5-VL	7B	Chart
Qwen2.5-VL	7B	Real-World
InternVL3	8B	CLEVR
InternVL3	14B	CLEVR

📊 Datasets

Dataset Type	Description	Link
CLEVR-based	Synthetic scenes for logical reasoning
Chart-based	Chart data for document understanding
RealWorld	Real-world image data

🛠️ Setup

📢 Acknowledgment: This repo is based on vlm-r1 - thanks for their contribution!

Prerequisites

• Python 3.10+
• CUDA-compatible GPU (recommended)
• Conda or similar environment manager

Installation

# Create and activate environment
conda create -n vision-zero python=3.10
conda activate vision-zero

# Install dependencies
bash setup.sh

💪 Training

📋 Training Pipeline

Step 1: 📁 Prepare Dataset and Model

Download one of the available datasets or prepare your own:

• CLEVR-based: Available now ✅
• Chart-based: Available now ✅
• Real-World: Available now ✅

Configure your training setup in run_scripts/run_grpo_vision_zero.sh:

# Configuration variables
IMAGES_DIR=$IMAGES_DIR          # Path to your images
SCENES_DIR=$SCENES_DIR          # Path to scene descriptions  
MODEL=$MODEL                    # Base model to fine-tune
OUTPUT_BASE_DIR=$OUTPUT_DIR     # Output directory for checkpoints
RUN_NAME="your_run_name"        # Experiment name

Step 2: 🚀 Start Training

Launch the training process with customizable hyperparameters:

bash run_scripts/run_grpo_vision_zero.sh

💡 Tip: All hyperparameters can be modified directly in the script file.

Step 3: 📊 Evaluation

Evaluate your trained model on out-of-distribution tasks using VLMEvalKit:

# After training completes and checkpoint is saved
# Use VLMEvalKit for comprehensive evaluation

---

📊 Evaluation

We use VLMEvalKit for comprehensive model evaluation on out-of-distribution tasks, ensuring robust performance assessment across various benchmarks.

---

📄 Citation

If you find Vision-Zero useful in your research, please consider citing our paper:

@misc{wang2025visionzeroscalablevlmselfimprovement,
    title={Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play}, 
    author={Qinsi Wang and Bo Liu and Tianyi Zhou and Jing Shi and Yueqian Lin and Yiran Chen and Hai Helen Li and Kun Wan and Wentian Zhao},
    year={2025},
    eprint={2509.25541},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2509.25541}
}

---

🌟 Star this repo if you find it helpful!

Made with ❤️ by the Vision-Zero team