Step-DeepResearch - A deep research AI model launched by Step-Leap Star.

What is Step-DeepResearch?

Step-DeepResearch, a 3.2 billion parameter deep research AI model developed by Step-Leap Star, can complete complex research tasks and generate professional reports in a single inference iteration. The model employs a monolithic architecture and achieves a highly efficient and low-cost research loop through four core capabilities: planning, deep search, reflective verification, and report writing. It scored 61.42 in the ResearchRubrics benchmark, demonstrating performance close to top-tier closed-source models, but with a single report cost as low as 0.5 RMB. Its unique feature lies in strengthening the model's decision-making capabilities through mid-training, making it particularly adept at professional fields such as finance and healthcare. It supports features like focused search and user-selected information sources, making it suitable for industry research, academic analysis, and other scenarios.

The main functions of Step-DeepResearch

• Atomic capability integrationComplex research tasks are broken down into trainable atomic capabilities such as planning, information retrieval, reflection and cross-validation, and professional report generation, and deeply internalized at the model level to ensure closed-loop reflection and dynamic correction in a single inference.
• Progressive training processEstablish a complete optimization path from mid-term training of the agent to supervised fine-tuning (SFT) and reinforcement learning (RL), reshape the training objective to "determine the next atomic action", and improve the model's adaptability and generalization performance.
• Outstanding performance across model scalesWith only 32 billion parameters, it achieved a score of 61.4% in the Scale AI Research Rubrics test, on par with OpenAI Deep Research and Gemini Deep Research. In the ADR-Bench expert evaluation, its Elo score significantly outperformed larger-scale models.
• Excellent cost-effectivenessIt maintains expert-level research capabilities while having extremely low deployment and inference costs, making it the most cost-effective deep research proxy solution in the industry.
• High-quality data synthesisBy employing an atomic capability-based data synthesis strategy, "doctoral-level" training data with detailed inference trajectories is generated, overcoming the challenge of scarce research data.
• Authoritative Information Acquisition and ProcessingWe employ a carefully selected authoritative indexing strategy, isolating 600+ authoritative domains to ensure factual basis. Knowledge-intensive retrieval maximizes the information density of a single token at the paragraph level, and prioritizes highly trustworthy sources when semantic relevance is comparable.
• Dynamic Dual-Circulation Cognitive ArchitectureThe system constructs a dual-loop workflow of "dynamic programming-hierarchical synthesis". After the planner agent initially generates the research outline, it continuously optimizes the research path based on new discoveries through reinforcement learning algorithms.
• Professional report generationBased on an "industry template library + dynamic knowledge base", it generates vertical domain documents that meet the format requirements. Important viewpoints in the reports have clear information source citations and have the credibility of professional research.

The technical principles of Step-DeepResearch

• Single agent architecture and dynamic loopBased on a single-agent architecture and following the ReAct paradigm, deep research tasks are restructured into a dynamic reasoning-action-observation cycle. Through core stages such as planning and reflection, tool execution, feedback and cross-validation, a specialized toolset is used to generate comprehensive research reports.
• The construction and training of atomic capabilitiesThis approach breaks down complex research tasks into trainable atomic capabilities such as planning, information retrieval, reflection and cross-validation, and professional report generation. Data is generated through specific closed-loop processes, such as designing error reflection loops in long-term inference, to improve the model's robustness and cross-validation performance.
• Progressive training strategyThe training process employs a phased approach, including mid-term training with medium-length contexts, mid-term training with long contexts, and a reinforcement learning phase. By progressively expanding the context length and optimizing the model's atomic capabilities, the model's performance on complex tasks is improved.
• Data synthesis and reinforcement learningA data synthesis strategy based on atomic capabilities is employed to generate "doctoral-level" training data with detailed inference trajectories. Through reinforcement learning algorithms, expert-aligned scale judgments are converted into binary reward signals, accelerating the model's convergence towards expert-aligned behavior.
• Authoritative Information Acquisition and ProcessingWe employ a carefully selected authoritative indexing strategy, isolating 600+ authoritative domains to ensure factual basis. Knowledge-intensive retrieval maximizes the information density of a single token at the paragraph level, and prioritizes highly trustworthy sources when semantic relevance is comparable.
• Dynamic Programming and OptimizationThe system constructs a dual-loop workflow of "dynamic programming-hierarchical synthesis". After the planner agent initially generates the research outline, it continuously optimizes the research path based on new discoveries through reinforcement learning algorithms.

Step-DeepResearch project address

• Github repositoryhttps://github.com/stepfun-ai/StepDeepResearch
• arXiv technical paperhttps://arxiv.org/pdf/2512.20491

Application scenarios of Step-DeepResearch

• academic researchIt helps researchers quickly generate literature reviews, research plans, and preliminary research reports, accelerating the academic research process.
• Business AnalysisIt provides business analysts with market trend analysis, competitor research, and industry report generation to support business decisions.
• Policy making: Assist policymakers in conducting policy background research, impact assessments, and writing policy recommendation reports to support the scientific and rational nature of policy formulation.
• Technology DevelopmentIn the technology field, it is used for new technology research, technology trend analysis, and feasibility study report generation to promote technological innovation.
• EducationIt assists teachers and students in curriculum research, project design, and academic paper writing, thereby improving the quality of education and research capabilities.
• HealthcareSupport medical researchers in conducting disease research, evaluating treatment methods, and reviewing medical literature, thereby promoting the development of medical technology.