StepFun introduced Step-DeepResearch – a 32B parameter deep research agent with long-horizon reasoning and tool usage, as well as structured reporting. Model is built using Qwen2.5-32B Base and trained as an agent to plan, investigate sources, verify evidence, and write reports that include citations while maintaining low inference costs.
Deep Research: From Searching to Deep Research
The vast majority of existing web agents use benchmarks to answer questions in multiple-hop answers. The agents try to find answers that are based on the truth for simple questions. The process is similar to retrieval targeted to specific questions, but not real research. The deep research tasks differ. Deep research tasks are different. They require latent intention recognition, long-horizon decision-making, the use of multi-turn tools, structured reasoning, and cross-source validation under uncertainty.
The Step-DeepResearch team reframes it as sequential decision making based on a small set of capabilities. The research team has defined 4 capabilities: task and planning decomposition, information seeking in depth, reflection and validation, and the generation of professional reports. Instead of orchestrating multiple external agents, this system internallyizes a loop that determines each action.
Data Synthesis Around Atomic Capabilities
The research team creates separate pipelines of data for each capability to teach these atomic abilities. They start with high-quality technical reports, surveys papers, and financial analysis documents. The reverse engineering of realistic research and task tree plans is done from title, abstract and structure. They then produce trajectories following these plans. The model is exposed to the long-horizon project structure, and not just short questions.
They construct queries based on graphs over knowledge graphs like Wikidata5m or CN-DBpedia for deep information searching. Sample subgraphs are expanded using search and questions synthesized that involve multihop reasoning based on entities and documents. Another pipeline makes use of a Wiki hyperlink index for cross-document retrieval. The training is focused on the hard problems, and easy questions are eliminated. A strong model will already be able to solve them with ReAct’s simple strategy.
Self-correction cycles and multiple-agent teachers traces generate data for reflection and verification. Teacher agents will extract claims, create checks and facts to verify, replan when inconsistencies are found, then only write reports. They are then cleaned up and assigned to a single agent as supervision. The report generation process is divided into two phases: mid-training for style and depth of domain using query reports pairs and then fine tuning with strict format and plan consistency restrictions.
Qwen2.5 32B Base – Progressive Training
Three stages are included in the training pipeline: agentic midtraining, fine-tuning supervised and reinforcement learning. The team introduces contexts up to 32k tokens in mid-training stage-1. Data on active reading, synthesizing reasoning traces and summarizing and reflecting is included. As training increases to 150B tokens or more, the research team shows steady gains in SimpleQA, TriviaQA, and FRAMES. FRAMES is showing the most gains, as it emphasizes structured reasoning.
Stage-2 extends the context to 128k symbols and introduces explicit tool calls. Model learns to perform tasks like URL-based questions, deep web searches, document summarization, and dialogue reasoning. In this stage, the model aligns with actual research scenarios in which search, browsing, and analysis are all mixed together.
The 4 atomic capabilities of the system are combined to form full depth search and deep investigation traces during supervised tuning. Cleansing data ensures a short, correct path in terms both of tools and steps. Pipelines inject controlled tool mistakes followed by corrective actions to increase robustness and enforce citation formats, so reports are grounded in the sources retrieved.
The agent is then optimized in the real-tool environment using reinforcement learning. The team creates tasks and check-lists using reverse synthesis and then trains Rubrics Judge, a style of checklist judge to score the reports in fine-grained dimension. The design of the reward converts binary ternary labels to asymmetric rewards, which capture both positive goals and violations. This policy will be trained using PPO, a learner critic and generalized advantage estimation (near zero discount) to ensure that the trajectory is not cut short.
The Single Agent ReAct Architecture Stack and the Search Stack
Step-DeepResearch’s inference agent is run as one ReAct style agent. This alternates tool calls, thinking and observations, until the report output decision. A todo list, shell commands, file operations, and batch web searching are all included in the tool set. The execution runs within a sandbox, with terminal persistence via tmux. By using the perceptual distance, a browser that is oriented towards perception reduces duplicate page captures. The multimodal input is supported with tools such as audio transcription, image analysis and document parsing.
Two resources are used to acquire information. StepFun states that the Search API of its Search Engine is based on over 20M premium papers, and 600 premium indexes. The team describes an authority indexing strategy which isolates over 600 domains including academic, government and institutional websites. Retrieval is performed at the paragraph level, and authority-aware ranking ensures that domains with high levels of trust are given preference when similar relevance exists.
File tools allow for patch-based editing. The agent is able to update just the modified parts of a document. Summary aware storage writes tool outputs into local files, and only injects compact summaries in the context. The external memory acts to prevent context overload for large projects.
Assessment, Access and Cost
ADR-Bench is a Chinese Benchmark with 110 open ended questions across nine domains. The 70 tasks are divided into general areas such as science, education, social life and engineering. They were evaluated side-by-side by experts. The rubrics for 40 finance and law tasks are based on atomicity, verifiability and fidelity constraints.
On Scale AI The Research RubricsStep-DeepResearch achieves 61.42 per cent rubric compliance. This is similar to OpenAI DeepResearch and Gemini DeepResearch. It also surpasses multiple proprietary and open baselines. Elo expert ratings on ADR-Bench show the 32B model to be superior to larger open models like MiniMaxM2, DeepSeekV3.2 and GLM4.6. It is also competitive with KimiResearcher, MiniMaxAgentPro and other systems.
The Key Takeaways
- Design for single agent, Atomic CapabilityIt internalizes four atomic abilities, including planning, information gathering, verification and reflection, as well as professional report generation.
- Data synthesis targeted for each skillThe team creates separate data pipelines to support planning, information searching, reflection, and writing reports. They use reverse engineered plans from actual reports, graph queries on Wikidata5m, CN-DBpedia and multi-agent teachers’ traces, as well as strict data formatting for report formats.
- Training in three stages with long-term context and RLSFT creates deep research trajectories and PPO based RL optimizes reports against fine grained check-lists.
- The ReAct architecture includes curated search, external memory and a reacted architecture.The model uses an inference loop at inference time that runs ReAct, which calls for tools to perform batch web searches, todos, shell operations and file operations. It also relies on a Search API based on more than 20M documents and 600 premium indexes as well 600+ trusted domains.
- Low cost, high quality productsIn ADR Bench, the model achieved a 67.1 percent success rate or tie against strong baselines.
Take a look at the Paper The following are some examples of how to get started: Repo. Also, feel free to follow us on Twitter Join our Facebook group! 100k+ ML SubReddit Subscribe now our Newsletter. Wait! What? now you can join us on telegram as well.

