README.md

---
base_model: zai-org/Glyph
library_name: transformers
license: other
pipeline_tag: image-text-to-text
tags:
- llama-factory
- full
- generated_from_trainer
- vision-language-model
- reasoning
model-index:
- name: vtc-r1-glyph
  results: []
---

# VTC-R1-Glyph

VTC-R1 (Vision-Text Compression for Efficient Long-Context Reasoning) is an efficient reasoning paradigm that integrates vision-text compression into the reasoning process. This repository contains the fine-tuned version of [zai-org/Glyph](https://huggingface.co/zai-org/Glyph) (based on GLM-4V) using this paradigm.

- **Paper:** [VTC-R1: Vision-Text Compression for Efficient Long-Context Reasoning](https://huggingface.co/papers/2601.22069)
- **Repository:** [https://github.com/w-yibo/VTC-R1](https://github.com/w-yibo/VTC-R1)

## Model Description

VTC-R1 addresses efficiency bottlenecks in long-context reasoning for Vision-Language Models (VLMs). Instead of processing lengthy textual traces, VTC-R1 renders intermediate reasoning segments into compact images, which are iteratively fed back into the model as "optical memory." 

Key features:
- **Efficiency:** Achieves 3.4x token compression and 2.7x speedup in end-to-end latency.
- **Performance:** Outperforms standard long-context reasoning on benchmarks like MATH500, AIME25, AMC23, and GPQA-D.
- **Scalability:** Integrates vision-text compression directly into the reasoning process without needing external compression models.

## Setup & Inference

### Installation
To use this model, install the required dependencies:
```bash
apt-get install poppler-utils # or conda install -c conda-forge poppler
pip install torch==2.6.0
pip install transformers==4.57.1 
pip install reportlab
pip install pdf2image 
```

### Inference
You can run the inference code provided in the [official repository](https://github.com/w-yibo/VTC-R1) to generate VTC-R1 style reasoning:
```bash
python inference.py # replace your model path in the script
```

## Training Procedure

The model was fine-tuned using [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) on a dataset derived from OpenR1-Math-220K.

### Training Hyperparameters

The following hyperparameters were used during training:
- **learning_rate:** 1e-05
- **train_batch_size:** 1
- **eval_batch_size:** 8
- **seed:** 42
- **distributed_type:** multi-GPU
- **num_devices:** 8
- **gradient_accumulation_steps:** 8
- **total_train_batch_size:** 64
- **total_eval_batch_size:** 64
- **optimizer:** AdamW with betas=(0.9,0.999) and epsilon=1e-08
- **lr_scheduler_type:** cosine
- **lr_scheduler_warmup_ratio:** 0.1
- **num_epochs:** 1

## Citation

If you find this work useful, please cite:

```bibtex
@misc{wang2026vtcr1visiontextcompressionefficient,
      title={VTC-R1: Vision-Text Compression for Efficient Long-Context Reasoning}, 
      author={Yibo Wang and Yongcheng Jing and Shunyu Liu and Hao Guan and Rong-cheng Tu and Chengyu Wang and Jun Huang and Dacheng Tao},
      year={2026},
      eprint={2601.22069},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.22069}, 
}
```

## Framework Versions

- Transformers 4.57.1
- Pytorch 2.6.0+cu124
- Datasets 4.0.0
- Tokenizers 0.22.1