The Year in AI Papers
The year/Independent research

Paper 2512.20491

Step-DeepResearch Technical Report

Chen Hu, Haikuo Du, Heng Wang, Lin Lin, Mingrui Chen, Peng Liu, Ruihang Miao, Tianchi Yue, Wang You, Wei Ji, Wei Yuan, Wenjin Deng, et al.

Published
Dec 2025
Research lab
Independent
Citations
11
GitHub
569 stars

01 In brief

Summary

Step-DeepResearch is a 32B-parameter, end-to-end Deep Research agent model developed by StepFun.

It uses a data synthesis strategy based on atomic capabilities (planning, information seeking, reflection, and report writing) and a progressive training pipeline (agentic mid-training, SFT, RL) with a Checklist-style Judger reward.

The model achieves a score of 61.42 on the Scale AI ResearchRubrics benchmark, outperforming larger open-source models and rivaling proprietary systems like OpenAI DeepResearch and Gemini DeepResearch, while costing less than 0.50 RMB per report.

The authors also introduce ADR-Bench, a Chinese benchmark for realistic Deep Research scenarios, on which Step-DeepResearch outperforms comparable models in expert Elo ratings.

The system uses a single-agent ReAct architecture with a specialized toolset, including authority-aware search, token-efficient file editing, and multimodal perception.

Key findings show that mid-training improves human preference, and the model excels in implicit criteria and citation quality, though it lags in STEM and philosophy domains.

The work demonstrates that medium-sized models can achieve expert-level Deep Research capabilities cost-effectively.

02 From the paper

Abstract

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.