[2026.04.09] We officially open-source the CRUX dataset and ViRC models as scheduled. 🚀
[2026.03.01] The CURX dataset and ViRC models are ready and are currently under internal review at Ant Group. Full open-sourcing will be completed no later than 2026.04.19.⚡️
[2026.02.21] ViRC has been accepted by CVPR 2026. 🎉
[2025.12.16] We release the arxiv paper and the code. 🔥
👀 About VɪRC
Existing MLLMs typically perform textual reasoning solely from a single static mathematical image, overlooking dynamic visual acquisition during reasoning. In contrast, humans repeatedly examine visual image and employ step-by-step reasoning to prove intermediate propositions. We propose a ViRC framework for multimodal mathematical tasks, introducing a Reason Chunking mechanism that structures multimodal mathematical CoT into consecutive Critical Reasoning Units (CRUs) to simulate human expert problem-solving patterns. CRUs ensure intra-unit textual coherence for intermediate proposition verification while integrating visual information across units to generate subsequent propositions and support structured reasoning.
To this end, we present CRUX dataset by using three visual tools and four reasoning patterns to provide explicitly annotated CRUs across multiple reasoning paths for each mathematical problem.
Leveraging the CRUX dataset, we propose a progressive training strategy inspired by human cognitive learning, which includes Instructional SFT, Practice SFT, and Strategic RL, aimed at further strengthening the Reason Chunking ability of the model.
The resulting ViRC-7B model achieves a 18.8% average improvement over baselines across multiple mathematical benchmarks and cross-domain high-resolution image benchmarks.
1pip installdatasets==4.0.0
2# Start Qwen2.5-VL-72B-Instruct with vLLM, and update Lines 32, 36, and 37 accordingly (the model-related settings).3\cp -f src/train/reward_for_scale_dynamic.py ./verl/utils/reward_score/init.py
4bash src/train/stage_3_StratRL
Notes on Training Config
We provide src/train/merge_rl_result.sh to merge Strategic RL outputs and export the final model weights in safetensors format.
Our training data and scripts support two image-resolution settings:
Dynamic resolution (for models such as Qwen2.5VL).
Fixed resolution (images resized to 1000×1000, for models such as Qwen2VL/Qwen3VL).
These are distinguished by the suffixes _scale_dynamic and _scale_fixed under data/ and src/.
For reproducibility, we release a 50K subset of the training data (suffix _50K) and the full 100K set (suffix _100K).
💫 Inference
We provide inference scripts for two image-resolution settings:
Dynamic resolution (src/evaluation/ViRC_scale_dynamic.py): for ViRC-7B and ViRC-3B, based on the Qwen2.5-VL-Instruct series.
Fixed resolution (1000×1000) (src/evaluation/ViRC_scale_fixed.py): for ViRC-Qwen2VL-7B and ViRC-Qwen2VL-2B, based on the Qwen2-VL-Instruct series.
src/evaluation/ also includes a sample input image (image.png) and an expected output example in src/evaluation/response/ for a quick sanity check.
Start the ViRC model with vLLM, then run the evaluation script:
We would like to thank LLaMA-Factory and verl, upon which our repo is built.
✅ Citation
@article{wang2025virc,
title={ViRC: Enhancing Visual Interleaved Mathematical CoT with Reason Chunking},
author={Lihong, Wang and Liangqi, Li and Weiwei, Feng and Jiamin, Wu and Changtao, Miao and Tieru, Wu and Rui, Ma and Bo, Zhang and Zhe, Li},
journal={arXiv preprint arXiv:2512.14654},
year={2025}
}
