MM-ReCoder

CVPR 2026  |  Project Page  |  arXiv  |  Code  |  SFT Cold-Start

MM-ReCoder is the 7B vision-language model from the CVPR 2026 paper MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction. It converts a chart image into the matplotlib code that reproduces it. At inference time the model renders its own code with a sandboxed matplotlib tool, inspects the result, and self-corrects across multiple turns.

This is the final RL-trained checkpoint. It is fine-tuned from Qwen/Qwen2.5-VL-7B-Instruct via:

1. SFT cold-start — released separately as cwbc/MM-ReCoder-SFT-Cold-Start.
2. Multi-turn RL (GRPO), stage 1 — shared-first-turn optimization.
3. Multi-turn RL (GRPO), stage 2 — full-trajectory optimization, resumed from stage 1.

Usage

The recommended way to use MM-ReCoder is through the inference scripts in the official repository, which wrap the model with the self-correction agent loop (render → critique → revise):

git clone https://github.com/ZitianTang/MM-ReCoder.git
cd MM-ReCoder
# Follow the Installation section in the repo README.

# Downalod the MM-ReCoder checkpoint from Hugging Face
hf download cwbc/MM-ReCoder

# Two-turn self-correction on ChartMimic.
bash examples/mmrecoder/inference/chartmimic_2turns.sh

Direct single-turn use (no self-correction)

You can also load the model in a single-pass setting via transformers:

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from PIL import Image
import torch

model_id = "cwbc/MM-ReCoder"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

image = Image.open("path/to/chart.png").convert("RGB")
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Generate the matplotlib code that reproduces this chart."},
    ],
}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
print(processor.batch_decode(out[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0])

This emits code in one shot. The full self-correction behavior requires the agent loop in the repository.

Training

• Base model: Qwen2.5-VL-7B-Instruct.
• RL algorithm: GRPO with chart-specific rule-based rewards (format, color, text, layout, type) plus an LLM-as-a-judge model reward.
• RL data: Chart2Code-160k prompts.
• Evaluation: ChartMimic (direct-600), Plot2Code, and ChartX.

See the repository for full training scripts and configs.

Citation

@inproceedings{tang2026mmrecoder,
    title={MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction},
    author={Zitian Tang and Xu Zhang and Jianbo Yuan and Yang Zou and Varad Gunjal and Songyao Jiang and Davide Modolo},
    booktitle={CVPR},
    year={2026}
}

License

Released under the Apache 2.0 License, inheriting from the base Qwen2.5-VL-7B-Instruct license.