MetaStoneTec/XBai-o4

News

• 2025.08.01: We release XBai o4, where o=open, and o4 represents our fourth-generation open-source large model technology. XBai o4 excels in complex reasoning capabilities and has now completely surpassed OpenAI-o3-mini in Medium mode. View Github to get more information!

Introduction

XBai o4 is trained based on our proposed reflective generative form, which combines “Long-CoT Reinforcement Learning” and “Process Reward Learning” into a unified training form. This form enables a single model to simultaneously achieve deep reasoning and high-quality reasoning trajectory selection. By sharing the backbone network between the PRMs and policy models, XBai o4 significantly reduces the inference cost of PRMs by 99%, resulting in faster and higher-quality responses.

For full details please refer to our paper and our official website.

Performance

Model	AIME24	AIME25	LiveCodeBench v5	C-EVAL
s1-32B	56.7	50.0	-	-
QwQ-32B	79.5	69.5	62.7	88.4
R1-Distill-Qwen-32B	72.6	49.6	54.5	82.2
GLM-Z1-32B-0414	80.8	63.6	-	-
DeepSeek-R1-671B-0120	79.8	70.0	64.3	91.8
Claude-3.5-Sonnet1022	16.0	7.4	40.2	76.7
GPT-4o-0513	9.3	11.6	32.3	-
OpenAI-o1-mini	63.6	50.7	49.4	68.9
OpenAI-o1-1217	79.2	-	63.9	-
OpenAI-o3-mini-medium	79.6	74.8	66.3	75.9
Claude Opus 4	75.7	75.5	61.3	-
Qwen3-32B	81.4	72.9	65.7	87.3
XBai o4-low	82.4	74.8	66.6	89.4
XBai o4-medium	85.4	77.6	67.0	89.5
XBai o4-high	86.5	77.9	67.2	89.7

Model

We save the parameters of the policy model and the SPRM head into two files:

• "model.safetensors" is the checkpoint of the policy model.

• "score_module.pt" is the checkpoint of the SPRM head.

You can find other sizes of MetaStone‑S1 below:

Model	Transformers(HF)	ModelScope
XBai o4	XBai o4	XBai o4

Training & Evaluation

Since Huggingface models do not directly support inference on SPRM. Please refer to our github repository for the detailed training and testing pipeline.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{wang2025testtimescalingreflectivegenerative,
 title={Test-Time Scaling with Reflective Generative Model}, 
 author={Zixiao Wang and Yuxin Wang and Xiaorui Wang and Mengting Xing and Jie Gao and Jianjun Xu and Guangcan Liu and Chenhui Jin and Zhuo Wang and Shengzhuo Zhang and Hongtao Xie},
 year={2025},
 eprint={2507.01951},
 archivePrefix={arXiv},
 primaryClass={cs.LG},
 url={https://arxiv.org/abs/2507.01951}, 
}