Skip to content
Bud's Wiki

Step-DeepResearch Technical Report

TL;DR

Section titled “TL;DR”

Step-DeepResearch is StepFun’s open 32B end-to-end deep-research agent. It scores 61.4% on Scale AI’s ResearchRubrics, matching OpenAI Deep Research and Gemini Deep Research while running at a small fraction of frontier-model inference cost. The recipe is a Data Synthesis Strategy Based on Atomic Capabilities — decompose research into planning, information seeking, reflection / cross-validation, and report writing, synthesize trajectories per capability — combined with a progressive Agentic Mid-Training → SFT → RL pipeline, gated by a Checklist-style Judger as reward signal. The paper also ships ADR-Bench, a Chinese-domain deep-research benchmark with expert Elo evaluation.

Key claims

Section titled “Key claims”
  • A 32B base trained with the atomic-capability data synthesis + Mid-Training → SFT → RL pipeline reaches 61.4% on Scale AI ResearchRubrics, rivaling OpenAI Deep Research and Gemini Deep Research at industry-leading cost-efficiency [Abstract, §1].
  • The training objective is reframed from “predict the next token” to “decide the next atomic action” — the four atomic capabilities (planning, information seeking, reflection / cross-validation, report generation) are each made trainable units of supervision [§1, README Model Summary].
  • The pipeline is staged: Agentic Mid-Training first internalizes tool-use and ReAct-style closed-loop reflection; SFT then fits curated trajectories per atomic capability; RL with the Checklist-style Judger as reward closes the loop [§1, §5].
  • The agent is single-agent ReAct over a streamlined toolset — batch_web_surfer (batch web search/browse), file (read/write/edit), todo (task state), shell (interactive command) — released both as a framework with local implementations and behind a StepFun API [README System Architecture].
  • ADR-Bench is introduced as a realistic Chinese-domain deep-research benchmark with expert Elo evaluation; Step-DeepResearch’s Elo significantly outperforms DeepSeek-v3.2 and GLM-4.6 across all dimensions and rivals top-tier closed models [Abstract, README Cost-Efficiency & ADR-Bench comparison].
  • A follow-on model, Step 3.5 Flash (released Feb 2026), reaches 65.27 on ResearchRubrics from the same lineage at higher inference efficiency [README News, README Performance].

Method

Section titled “Method”

Step-DeepResearch is a single-agent ReAct-paradigm system: reasoning → action → reflection in a dynamic loop, executing against a small fixed toolset (batch web surfer, file, todo, shell). The training side is what’s novel. The team decomposes “deep research” into four atomic capabilities — planning, information seeking, reflection and cross-validation, and professional report generation — and synthesizes per-capability trajectory data using internal pipelines. Training then proceeds in three stages:

  1. Agentic Mid-Training. A new pre/post-training boundary where the base LLM is exposed to tool-use traces and ReAct-style closed-loop reflection at scale, internalizing the “decide next atomic action” objective before any task-specific supervision.
  2. SFT. Curated atomic-capability trajectories teach the format and structure of each phase (planning checklists, search-result triangulation, report skeletons).
  3. RL. A Checklist-style Judger (a structured rubric of correctness/format/coverage criteria, not a single scalar) provides the reward signal; the model is optimized end-to-end on multi-step research trajectories.

ADR-Bench is constructed in parallel as a Chinese-domain evaluation: open-ended research scenarios with expert reviewers producing Elo ratings across multiple research-quality dimensions.

Results

Section titled “Results”
  • ResearchRubrics (Scale AI): Step-DeepResearch 32B → 61.4%, matching OpenAI Deep Research and Gemini Deep Research; Step 3.5 Flash → 65.27 [Abstract, README Performance].
  • ADR-Bench Elo: Step-DeepResearch’s Elo significantly outperforms DeepSeek-v3.2 and GLM-4.6 across all evaluated dimensions and is competitive with top-tier closed-source models [README ADR-Bench comparison].
  • Cost positioning: the team explicitly frames the 32B model at the “high-efficiency frontier” — near-top quality at a substantially lower per-query cost in RMB than frontier closed-source DR systems [README Cost-Efficiency].

Why it’s interesting

Section titled “Why it’s interesting”

This is the first filed 32B-scale deep-research agent that matches the closed frontier (OpenAI / Gemini Deep Research) on the canonical rubric benchmark, and it does so with an open Apache-2.0 release including the agent harness. It directly extends the Tool-Use Agents design space staked out by MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling (interaction-depth scaling, SFT→DPO→GRPO) and Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing (closed multi-agent data loop, GRPO + gated rewards) — but its differentiating bet is the Atomic Capabilities decomposition: instead of synthesizing one undifferentiated trajectory pool, each research sub-skill becomes its own training signal, and the Checklist-style Judger provides a structured (not scalar) RL reward. This is the same gated/shaped-reward direction the Reasoning RL cluster is converging on, applied with sharper task decomposition than prior open recipes. It also adds another datapoint to Synthetic Training Data — the entire pipeline is synthesized via internal data generation per atomic capability, with the Judger acting as the verification gate.

See also

Section titled “See also”