Imagen de Portada

Step-DeepResearch Technical Report

Fuente: arXiv

Guardado en:

Detalles Bibliográficos
Autores principales:	Hu, Chen, Du, Haikuo, Wang, Heng, Lin, Lin, Chen, Mingrui, Liu, Peng, Miao, Ruihang, Yue, Tianchi, You, Wang, Ji, Wei, Yuan, Wei, Deng, Wenjin, Yuan, Xiaojian, Zhang, Xiaoyun, Liu, Xiangyu, Liu, Xikai, Xu, Yanming, Cao, Yicheng, Zhang, Yifei, Wang, Yongyao, Shu, Yubo, Zhang, Yurong, Zhang, Yuxiang, Gong, Zheng, Chang, Zhichao, Li, Binyan, Ma, Dan, Jia, Furong, Wang, Hongyuan, Liu, Jiayu, Bai, Jing, Liu, Junlan, Liu, Manjiao, Wang, Na, Wu, Qiuping, Du, Qinxin, Li, Shiwei, Sun, Wen, Gong, Yifeng, Chen, Yonglin, Zhao, Yuling, Lin, Yuxuan, Ren, Ziqi, Wang, Zixuan, Zhang, Aihu, Li, Brian, Ma, Buyun, An, Kang, Xie, Li, Li, Mingliang, Li, Pan, Yang, Shidong, Chen, Xi, Liu, Xiaojia, Luo, Yuchu, Song, Yuan, Ding, YuanHao, Liang, Yuanwei, Li, Zexi, Zhang, Zhaoning, Zhang, Zixin, Jiao, Binxing, Jiang, Daxin, Chen, Jiansheng, Li, Jing, Zhang, Xiangyu, Zhu, Yibo
Formato:	Preprint
Publicado:	2025
Materias:	Computation and Language
Acceso en línea:	Acceder al recurso
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

_version_	1866918265038569472
author	Hu, Chen Du, Haikuo Wang, Heng Lin, Lin Chen, Mingrui Liu, Peng Miao, Ruihang Yue, Tianchi You, Wang Ji, Wei Yuan, Wei Deng, Wenjin Yuan, Xiaojian Zhang, Xiaoyun Liu, Xiangyu Liu, Xikai Xu, Yanming Cao, Yicheng Zhang, Yifei Wang, Yongyao Shu, Yubo Zhang, Yurong Zhang, Yuxiang Gong, Zheng Chang, Zhichao Li, Binyan Ma, Dan Jia, Furong Wang, Hongyuan Liu, Jiayu Bai, Jing Liu, Junlan Liu, Manjiao Wang, Na Wu, Qiuping Du, Qinxin Li, Shiwei Sun, Wen Gong, Yifeng Chen, Yonglin Zhao, Yuling Lin, Yuxuan Ren, Ziqi Wang, Zixuan Zhang, Aihu Li, Brian Ma, Buyun An, Kang Xie, Li Li, Mingliang Li, Pan Yang, Shidong Chen, Xi Liu, Xiaojia Luo, Yuchu Song, Yuan Ding, YuanHao Liang, Yuanwei Li, Zexi Zhang, Zhaoning Zhang, Zixin Jiao, Binxing Jiang, Daxin Chen, Jiansheng Li, Jing Zhang, Xiangyu Zhu, Yibo
author_facet	Hu, Chen Du, Haikuo Wang, Heng Lin, Lin Chen, Mingrui Liu, Peng Miao, Ruihang Yue, Tianchi You, Wang Ji, Wei Yuan, Wei Deng, Wenjin Yuan, Xiaojian Zhang, Xiaoyun Liu, Xiangyu Liu, Xikai Xu, Yanming Cao, Yicheng Zhang, Yifei Wang, Yongyao Shu, Yubo Zhang, Yurong Zhang, Yuxiang Gong, Zheng Chang, Zhichao Li, Binyan Ma, Dan Jia, Furong Wang, Hongyuan Liu, Jiayu Bai, Jing Liu, Junlan Liu, Manjiao Wang, Na Wu, Qiuping Du, Qinxin Li, Shiwei Sun, Wen Gong, Yifeng Chen, Yonglin Zhao, Yuling Lin, Yuxuan Ren, Ziqi Wang, Zixuan Zhang, Aihu Li, Brian Ma, Buyun An, Kang Xie, Li Li, Mingliang Li, Pan Yang, Shidong Chen, Xi Liu, Xiaojia Luo, Yuchu Song, Yuan Ding, YuanHao Liang, Yuanwei Li, Zexi Zhang, Zhaoning Zhang, Zixin Jiao, Binxing Jiang, Daxin Chen, Jiansheng Li, Jing Zhang, Xiangyu Zhu, Yibo
contents	As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.
format	Preprint
id	arxiv_https___arxiv_org_abs_2512_20491
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	Step-DeepResearch Technical Report Hu, Chen Du, Haikuo Wang, Heng Lin, Lin Chen, Mingrui Liu, Peng Miao, Ruihang Yue, Tianchi You, Wang Ji, Wei Yuan, Wei Deng, Wenjin Yuan, Xiaojian Zhang, Xiaoyun Liu, Xiangyu Liu, Xikai Xu, Yanming Cao, Yicheng Zhang, Yifei Wang, Yongyao Shu, Yubo Zhang, Yurong Zhang, Yuxiang Gong, Zheng Chang, Zhichao Li, Binyan Ma, Dan Jia, Furong Wang, Hongyuan Liu, Jiayu Bai, Jing Liu, Junlan Liu, Manjiao Wang, Na Wu, Qiuping Du, Qinxin Li, Shiwei Sun, Wen Gong, Yifeng Chen, Yonglin Zhao, Yuling Lin, Yuxuan Ren, Ziqi Wang, Zixuan Zhang, Aihu Li, Brian Ma, Buyun An, Kang Xie, Li Li, Mingliang Li, Pan Yang, Shidong Chen, Xi Liu, Xiaojia Luo, Yuchu Song, Yuan Ding, YuanHao Liang, Yuanwei Li, Zexi Zhang, Zhaoning Zhang, Zixin Jiao, Binxing Jiang, Daxin Chen, Jiansheng Li, Jing Zhang, Xiangyu Zhu, Yibo Computation and Language As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.
title	Step-DeepResearch Technical Report
topic	Computation and Language
url	https://arxiv.org/abs/2512.20491