StepFun has launched Step-DeepResearch, a 32B parameter finish to finish deep analysis agent that goals to show internet search into precise analysis workflows with lengthy horizon reasoning, instrument use and structured reporting. The mannequin is constructed on Qwen2.5 32B-Base and is skilled to behave as a single agent that plans, explores sources, verifies proof and writes reviews with citations, whereas maintaining inference price low.
From Search to Deep Analysis
Most current internet brokers are tuned for multi-hop question-answering benchmarks. They attempt to match floor reality solutions for brief questions. That is nearer to focused retrieval than to actual analysis. Deep analysis duties are totally different. They contain latent intent recognition, lengthy horizon choice making, multi-turn instrument use, structured-reasoning and cross-source verification beneath uncertainty.
Step-DeepResearch reframes this as sequential choice making over a compact set of atomic capabilities. The analysis crew defines 4 atomic capabilities, planning and activity decomposition, deep-information searching for, reflection and verification, {and professional} report era. As an alternative of orchestrating many exterior brokers, the system internalizes this loop right into a single mannequin that decides the subsequent motion at every step.
Information Synthesis round Atomic Capabilities
To show these atomic capabilities, the analysis crew builds separate information pipelines for every talent. For planning, they begin from prime quality technical reviews, survey papers and monetary evaluation paperwork. They reverse-engineer real looking analysis plans and activity bushes from titles, abstracts and construction, then generate trajectories that comply with these plans. This exposes the mannequin to lengthy horizon challenge constructions, not solely brief query templates.
For deep data searching for, they assemble graph primarily based queries over information graphs equivalent to Wikidata5m and CN-DBpedia. They pattern subgraphs, develop them utilizing search, and synthesize questions that require multi hop reasoning throughout entities and paperwork. A separate pipeline makes use of a Wiki model hyperlink index to pressure cross doc retrieval and mixture of proof. Straightforward questions {that a} robust mannequin can already clear up with a easy ReAct model technique are filtered out, so coaching focuses on arduous search issues.
Reflection and verification information is generated by means of self-correction loops and multi-agent instructor traces. Trainer brokers extract claims, plan checks, confirm information, replan if inconsistencies seem and solely then write reviews. The ensuing trajectories are cleaned and used as supervision for a single scholar agent. Report era is skilled in 2 phases, mid coaching for area model and depth utilizing question report pairs, then supervised fine-tuning with strict formatting and plan consistency constraints.
Progressive Coaching on Qwen2.5-32B-Base
The coaching pipeline has 3 levels, agentic mid-training, supervised fine-tuning and reinforcement studying. In mid coaching stage-1, the crew injects atomic capabilities with out instruments, utilizing context size as much as 32k tokens. The information covers energetic studying, artificial reasoning traces, summarization and reflection. The analysis crew present regular good points on SimpleQA, TriviaQA and FRAMES as coaching scales as much as about 150B tokens, with the biggest good points on FRAMES, which stresses structured reasoning.
In stage-2, the context extends to 128k tokens and express instrument calls are launched. The mannequin learns duties equivalent to URL primarily based question-answering, deep internet search, lengthy doc summarization and lengthy dialogue reasoning. This stage aligns the mannequin with actual analysis eventualities the place search, searching and evaluation should be blended in a single trajectory.
Throughout supervised fine-tuning, the 4 atomic capabilities are composed into full deep search and deep analysis traces. Information cleansing retains trajectories which can be right and brief when it comes to steps and power calls. The pipeline injects managed instrument errors adopted by correction to enhance robustness, and enforces quotation codecs in order that reviews keep grounded within the retrieved sources.
Reinforcement studying then optimizes the agent in an actual instrument surroundings. The analysis crew builds duties and checklists by means of reverse synthesis, and trains a guidelines model Rubrics Choose to attain reviews alongside high quality grained dimensions. The reward design converts ternary rubric labels into uneven binary rewards that seize each optimistic targets and violations. The coverage is skilled with PPO and a discovered critic, utilizing generalized benefit estimation with close to zero low cost in order that lengthy trajectories usually are not truncated.
Single Agent ReAct Structure and Search Stack
At inference time, Step-DeepResearch runs as a single ReAct model agent that alternates pondering, instrument calls and observations till it decides to output a report. The instrument set contains batch internet search, a todo supervisor, shell instructions and file operations. Execution runs in a sandbox with terminal persistence by means of tmux. A notion oriented browser reduces redundant web page captures through the use of perceptual hash distance. Instruments for doc parsing, audio transcription and picture evaluation assist multimodal inputs.
Data acquisition makes use of 2 associated assets. StepFun crew states that its Search API is grounded in additional than 20M prime quality papers and 600 premium indices. The analysis crew then describes a curated authority indexing technique that isolates greater than 600 trusted domains, together with authorities, educational and institutional websites. Retrieval operates at paragraph degree and makes use of authority conscious rating so that prime belief domains are most well-liked when relevance is analogous.
The file instruments assist patch primarily based modifying, so the agent can replace solely modified sections of a report. A abstract conscious storage scheme writes full instrument outputs to native information and injects solely compact summaries into the context. This acts as exterior reminiscence and avoids context overflow for lengthy initiatives.
Analysis, Value and Entry
To measure deep analysis conduct, the crew introduce ADR-Bench, a Chinese language benchmark with 110 open ended duties throughout 9 domains. 70 duties cowl basic domains equivalent to schooling, science and engineering and social life, evaluated by professional facet by facet comparability. 40 duties in finance and regulation are scored with express rubrics that comply with atomicity and verifiability constraints.
On Scale AI Analysis Rubrics, Step-DeepResearch reaches 61.42 p.c rubric compliance, which is corresponding to OpenAI-DeepResearch and Gemini-DeepResearch, and clearly forward of a number of open and proprietary baselines. On ADR-Bench, expert-based Elo rankings present that the 32B mannequin outperforms bigger open-models equivalent to MiniMax-M2, GLM-4.6 and DeepSeek-V3.2, and is aggressive with techniques like Kimi-Researcher and MiniMax-Agent-Professional.
Key Takeaways
- Single agent, atomic functionality design: Step-DeepResearch is a 32B parameter single agent constructed on Qwen2.-32B-Base, it internalizes 4 atomic capabilities, planning, deep data searching for, reflection and verification, {and professional} report era, as an alternative of counting on many exterior brokers.
- Focused information synthesis for every talent: The analysis crew builds separate information pipelines for planning, deep data searching for, reflection and report writing, utilizing reverse-engineered plans from actual reviews, graph-based queries over Wikidata5m and CN-DBpedia, multi-agent instructor traces and strict report formatting information.
- Three stage coaching with lengthy context and RL: Coaching makes use of mid coaching, supervised fine-tuning and reinforcement studying, with mid coaching as much as 150B tokens at 32k after which 128k context, SFT composes full deep analysis trajectories, and PPO primarily based RL with a Rubrics Choose optimizes reviews towards high quality grained checklists.
- ReAct structure with curated search and exterior reminiscence: At inference time the mannequin runs a ReAct loop that calls instruments for batch internet search, todo, shell and file operations, makes use of a Search API grounded in additional than 20M papers and 600 premium indices together with 600+trusted domains, and depends on patch modifying and abstract conscious storage to behave as exterior reminiscence.
- Aggressive high quality with decrease price: On Scale AI Analysis Rubrics the mannequin reaches 61.42 p.c rubric compliance and is aggressive with OpenAI-DeepResearch and Gemini-DeepResearch, on ADR Bench it achieves 67.1 p.c win or tie fee towards robust baselines.
Take a look at the Paper and Repo. Additionally, be at liberty to comply with us on Twitter and don’t overlook to affix our 100k+ ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you’ll be able to be part of us on telegram as nicely.
