ZTE-DeepResearch-Bench

This is a codebase demonstrating the test results of DeepResearch Agent, a tool developed by ZTE's NLM team, on the DeepResearch Bench leaderboard.

ZTE-Nebula-DeepResearch on DeepResearch Bench (RACE)

Evaluation on the DeepResearch Bench RACE benchmark: 100 PhD-level research queries (50 Chinese, 50 English) scored by Gemini-2.5-Pro against expert-written reference articles.

Results

Overall RACE Score: 57.27 — #2 on the leaderboard.

Model	overall	comp.	insight	inst.	read.
DuMate/Qianfan-DeepResearch	58.03	59.48	61.48	53.87	54.34
ZTE-Nebula-DeepResearch	57.27	58.37	59.76	54.06	54.66
iFlow-Researcher	57.08	58.24	59.74	53.24	55.05
Zhipu Deep Research	57.06	58.15	60.14	53.47	53.88
Xiaoyi DeepResearch	57.00	58.58	59.38	53.58	53.99
Cellcog Max	56.67	57.4	60.01	53.25	53.21
1688AILab-DeepResearch	56.53	57.32	59.27	53.51	53.36
Octen Deep Research	56.31	56.89	59	53.39	53.83
Grep DeepResearch	56.23	56.82	58.92	53.38	53.44
Nvidia-aiq(Nemotron3, GPT 5.2)	55.95	56.9	58.49	52.89	53.43

Performance by Dimension

ZTE-Nebula-DeepResearch leads in Instruction Following (54.06)

ZTE-Nebula-DeepResearch Advantage Over Every Competitor

Score Distribution

Every single query (100/100) scored above the expert-written reference — ZTE-Nebula-DeepResearch doesn't just win on average, it wins consistently.

Scoring Variance

Three identical runs (same articles, score via official pipeline) to measure Gemini-2.5-pro judge non-determinism

How ZTE-Nebula-DeepResearch Works

ZTE‑Nebula‑DeepResearch is a self‑developed deep research agent system by ZTE. It mainly consists of the following five core modules:

1. Planning: A hierarchical mechanism and a rubric are introduced to guide the planning process, ensuring in‑depth analysis and decomposition of the research topic, accurately understanding user intent, and reasonably outputting detailed descriptions of subtasks.

2. Research: Based on the detailed subtask descriptions, end‑to‑end trained sub‑agents execute the research process dynamically in a Directed Acyclic Graph (DAG) manner, supporting both sequential and parallel execution. These sub‑agents can interact with online environments (e.g., web search, professional sources) through multiple rounds to obtain necessary information.

3. Research Draft: The reports written by sub‑agents are deeply integrated, supplemented with necessary summaries, research analysis, and conclusions, to form a complete research draft document.

4. Draft Optimization: Each chapter is examined for comprehensiveness and factual correctness. Further research is triggered when necessary, and the draft is iteratively improved.

5. Final Report: After verifying that the document structure and citations comply with the required standards, the final research report is published.

Acknowledgements

We extend our thanks to the DeepResearch Bench team for developing this comprehensive benchmark to evaluate deep research agents. For complete rankings and detailed evaluation criteria, please visit the DeepResearch Bench Leaderboard.

---

Information updated on 2026.05.25