Installation • Training • Evaluation • Visualization • Checkpoints • Citation

MM-Zero is a self-evolving reinforcement learning framework that improves vision-language models (VLMs) without requiring any human-annotated image data. It co-evolves three specialized agents — Proposer, CodeGen, and Solver — in an iterative loop where each agent bootstraps training signal for the others.

Each self-play iteration trains three models in sequence:

1. Proposer — generates diverse visual reasoning questions
2. CodeGen — writes SVG code that renders into images for those questions
3. Solver — learns to answer the generated visual questions via GRPO

At iteration i, each model evolves from its version at iteration i−1, creating a curriculum that grows in difficulty and diversity over time. The SVG rendering pipeline (cairosvg → PNG) is fully deterministic and requires no external image sources.

Built on EasyR1 / veRL.

1. Create a conda environment

conda create -n mm-zero python=3.12
conda activate mm-zero

2. Clone and install

git clone https://github.com/zli12321/MM-Zero.git
cd MM-Zero
bash setup.sh

setup.sh installs PyTorch (CUDA 12.8), vLLM, flash-attention, and all dependencies. It will also prompt you to log in to Weights & Biases for experiment tracking.

• 8× GPUs (80 GB each recommended, e.g., A100/H100)
  • 2 GPUs for GRPO training + 6 GPUs for vLLM inference (Proposer/CodeGen phases)
  • All 8 GPUs for Solver GRPO training
• 40 GB GPUs are supported by setting GPU_MEM=40

Launch the full self-play pipeline with a single command:

## Qwen3-VL-8B-Instruct
bash ./scripts/main_svg.sh

## Qwen3-VL-4B-Instruct
bash ./scripts/main_qwen3vl_4b.sh

This runs the complete iterative loop (Proposer → CodeGen → Solver) for multiple iterations, starting from a base model. Each iteration builds on the previous one's checkpoints.

Key environment variables (all have sensible defaults):

Example with custom settings:

STORAGE_PATH=/my/experiment \
Base_model=Qwen/Qwen3-VL-8B-Instruct \
GPU_MEM=80 \
TRAIN_STEPS=20 \
bash MM-zero_final/scripts/main_svg.sh

The pipeline is resumable — it automatically detects existing checkpoints and skips completed phases.

After training, evaluate the solver checkpoints on 12 multimodal benchmarks:

STORAGE=/path/to/your/storage bash run_eval.sh

The STORAGE path should point to the same directory used during training (i.e., STORAGE_PATH). The script automatically discovers all solver checkpoints under STORAGE/models/ and evaluates them with 8-way data parallelism.

Evaluated benchmarks: MMSI, MathVerse, MathVision, MathVista, MM-Vet, MMMU-Pro (4-option), VisNumBench, MMMU-Pro (10-option), MMMU-Pro-Vision, HallusionBench, MMMU, ChartQA.

Evaluation outputs are saved to STORAGE/eval_responses/, including per-model accuracy breakdowns and an LLM judge pass using Qwen2.5-14B-Instruct.

Compare evaluation accuracy across training iterations vs. the base model:

python eval_accuracy_comparison.py STORAGE_PATH/eval_responses/llm_accuracy_summary.jsonl

To plot co-evolution metrics (difficulty, solvability, diversity, render success rate, etc.):

python plot_coevolution.py \
    --storage_dirs /path/to/your/storage \
    --model_name "Qwen3-VL-8B-Instruct"

Pre-trained checkpoints and full training logs for Qwen3-VL-8B-Instruct are available on Hugging Face:

The logs include all model checkpoints across iterations and evaluation results on all 12 benchmarks.

To download to a specific folder:

huggingface-cli download IntelligenceLab/MM-Zero-Logs --local-dir /path/to/your/storage

Then point STORAGE to that folder when running evaluation or visualization.

Self-Agent/
├── MM-zero_final/
│   ├── scripts/              # Training orchestration scripts
│   ├── proposal_generate/    # Proposer inference & data generation
│   ├── code_generate/        # CodeGen inference
│   ├── code_render/          # SVG → PNG rendering pipeline
│   ├── question_evaluate/    # Question quality evaluation
│   ├── reward_function/      # GRPO reward functions
│   ├── configs/              # Training configurations
│   └── data/                 # Data utilities
├── verl/                     # Modified veRL/EasyR1 training engine
├── eval_generate.py          # Benchmark evaluation inference
├── llm_judge_eval.py         # LLM-based judge evaluation
├── eval_accuracy_comparison.py  # Accuracy comparison & plotting
├── plot_coevolution.py       # Co-evolution metric visualization
├── run_eval.sh               # Full evaluation pipeline
└── setup.sh                  # Environment setup

This project is released under the Apache 2.0 License.

@misc{li2026mmzeroselfevolvingmultimodelvision,
      title={MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data}, 
      author={Zongxia Li and Hongyang Du and Chengsong Huang and Xiyang Wu and Lantao Yu and Yicheng He and Jing Xie and Xiaomin Wu and Zhichao Liu and Jiarui Zhang and Fuxiao Liu},
      year={2026},
      eprint={2603.09206},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.09206}, 
}

• EasyR1 and veRL for the RL training framework
• vLLM for efficient inference