EVE: Verifiable Self-Evolution of MLLMs via Executable Visual Transformations

Official implementation of EVE, a method enabling Multimodal Large Language Models (MLLMs) to self-evolve through executable visual transformations.

Overview

EVE introduces a Challenger-Solver dual-policy framework where two policies co-evolve:

• Challenger: Generates executable Python image-editing programs with diverse parameter sets, creating verifiable visual transformation tasks.
• Solver: Learns to answer VQA questions derived from code execution results, receiving deterministic reward signals from the environment.

Unlike pseudo-label or template-based methods, EVE learns from an external execution environment — code execution provides deterministic, verifiable feedback independent of model confidence.

Installation

git clone https://github.com/0001Henry/EVE.git
cd eve

# Create conda environment
conda env create -f env.yml
conda activate eve

# Install package
pip install -e .

Requirements: Python 3.10+, PyTorch 2.8+, vLLM 0.11+, transformers 4.57+, CUDA-capable GPU(s)

Quick Start

1. Set environment variables

export BASEDIR=/path/to/eve
export OUTPUT_DIR=${BASEDIR}/output
export SAVE_CHECKPOINT_DIR=${OUTPUT_DIR}/checkpoints
export HF_HOME=/path/to/huggingface/cache

export BASE_MODEL=/path/to/Qwen3-VL-8B-Instruct

2. Run the full training loop

cd ${BASEDIR}
bash recipe/EVE/main.sh

This alternates between Challenger training and Solver training for ITER_NUM iterations.

Training Details

Challenger Training (C_train.sh)

Trains the Challenger policy to generate executable image-editing programs. The Challenger is rewarded for:

• Format validity and code executability
• Difficulty calibration (targeting ~50% Solver accuracy)
• Diversity of generated programs

# Called internally by main.sh; can also run standalone:
challenger_model=/path/to/model
solver_model=/path/to/model
experiment_name=my_experiment
bash recipe/EVE/C_train.sh

Solver Training (S_train.sh)

Trains the Solver policy on VQA tasks derived from Challenger-generated code executions. Two question types:

• Type-0: Given parameters, identify which image was produced
• Type-1: Given an edited image, identify which parameters produced it

bash recipe/EVE/S_train.sh

Key Configuration (recipe/EVE/configs/grpo.yaml)

Parameter	Default	Description
PQ_K	50	Priority queue size for code examples
TOTAL_TRAINING_STEPS	10	Training steps per iteration
ITER_NUM	5	Number of evolution iterations

Code Structure

EVE/
├── recipe/EVE/
│   ├── CS.py                  # Challenger-Solver dataset & reward logic
│   ├── main.sh                # Main training entry point
│   ├── C_train.sh             # Challenger training script
│   ├── S_train.sh             # Solver training script
│   ├── sandbox.py             # Code execution sandbox
│   ├── make_code_pq.py        # Priority queue initialization script
│   ├── gen_S_data_from_C.py   # Generate Solver data from Challenger output
│   ├── prepare_rl_data.py     # Data preprocessing pipeline
│   └── configs/grpo.yaml      # Training configuration
├── verl/
│   ├── trainer/main_ppo.py    # PPO/GRPO training entry point
│   └── utils/
│       ├── cs_sandbox.py      # Sandbox utilities
│       ├── cs_prompts.py      # System/user prompts
└── VLMEvalKit/                # Evaluation suite

Evaluation

We use VLMEvalKit for benchmark evaluation.

cd VLMEvalKit

# Configure model path, then run:
bash qwen3vl_8b_eval.sh

Citation

If you find this work useful, please cite:

@misc{heng2026eveverifiableselfevolutionmllms,
      title={EVE: Verifiable Self-Evolution of MLLMs via Executable Visual Transformations}, 
      author={Yongrui Heng and Chaoya Jiang and Han Yang and Shikun Zhang and Wei Ye},
      year={2026},
      eprint={2604.18320},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.18320}, 
}

Acknowledgments

This code builds upon:

• VERL — RL training infrastructure
• VLMEvalKit — Multimodal evaluation suite
• Vision-SR1-47K - Image dataset

License

This project is licensed under the Apache License 2.0. See LICENSE for details.