DeepResearch Bench

DeepResearch Bench
Overview
Full name	Deep Research Benchmark
Abbreviation	DRB
Description	A benchmark evaluating Deep Research Agents on PhD-level research tasks requiring multi-step exploration and synthesis
Release date	2025-06-13
Latest version	2.0
Benchmark updated	2025-07-15
Authors	Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, Zhendong Mao
Organization	University of Science and Technology of China, Metastone Technology
Technical Details
Type	Research Agent Evaluation, Report Generation, Multi-step Reasoning
Modality	Text, Web content
Task format	Research report generation, Multi-step exploration
Number of tasks	100
Total examples	100 PhD-level research tasks
Evaluation metric	RACE (quality assessment), FACT (citation accuracy)
Domains	Physics, Chemistry, Biology, Environmental Science, Engineering, And 17 others
Languages	English (50 tasks), Chinese (50 tasks)
Performance
Human performance	Established by domain experts
Baseline	Varies by model
SOTA score	48.92
SOTA model	Gemini-2.5-Pro Deep Research
SOTA date	2025-07
Saturated	No
Resources
Website	Official website
Paper	Paper
GitHub	Repository
Dataset	Download
License	Open source

DeepResearch Bench is a comprehensive artificial intelligence benchmark designed to evaluate Deep Research Agents (DRAs), large language model-based agents capable of conducting autonomous research and generating analyst-grade reports. Released in June 2025 by researchers from the University of Science and Technology of China and Metastone Technology^[1], DeepResearchBench addresses the critical need to assess AI systems' ability to perform PhD-level research tasks that require multi-step web exploration, targeted information retrieval, and higher-order synthesis capabilities.

Overview

DeepResearchBench represents a significant advancement in evaluating AI research capabilities by focusing on complex, real-world research scenarios that mirror the work of human researchers and analysts. The benchmark consists of 100 carefully curated PhD-level research tasks spanning 22 distinct academic fields, created and validated by over 100 domain experts holding PhD degrees or equivalent senior practitioner experience^[2].

Motivation

The creation of DeepResearch Bench was motivated by several factors:

• **Gap in existing benchmarks**: Traditional benchmarks fail to capture the complexity of real research tasks
• **Rise of research agents**: Emergence of AI systems claiming research capabilities without standardized evaluation
• **Need for quality assessment**: Lack of metrics for evaluating research report quality and citation accuracy
• **Bilingual evaluation**: Requirement for benchmarks supporting multiple languages in academic contexts
• **Real-world alignment**: Need for tasks reflecting actual PhD-level research challenges

Task Design and Structure

Task Distribution

DeepResearch Bench's 100 tasks are carefully distributed across academic disciplines and languages:

Category	Number of Tasks	Language Distribution
**Physical Sciences**	20	10 English, 10 Chinese
**Life Sciences**	18	9 English, 9 Chinese
**Engineering**	16	8 English, 8 Chinese
**Environmental Sciences**	12	6 English, 6 Chinese
**Social Sciences**	10	5 English, 5 Chinese
**Computer Science**	8	4 English, 4 Chinese
**Mathematics**	6	3 English, 3 Chinese
**Other Fields**	10	5 English, 5 Chinese

Task Characteristics

Each research task in DeepResearch Bench exhibits several key characteristics^[1]:

Characteristic	Description	Example
**Multi-step Exploration**	Requires iterative information gathering	Literature review → hypothesis formation → evidence synthesis
**Cross-source Integration**	Demands information from multiple sources	Academic papers + datasets + news articles
**Domain Expertise**	Needs specialized knowledge	Understanding quantum mechanics terminology
**Critical Analysis**	Requires evaluating conflicting information	Assessing contradictory research findings
**Synthesis Capability**	Demands creating coherent narratives	Writing comprehensive research reports

Task Creation Process

The benchmark's tasks were developed through a rigorous process:

1. **Query Analysis**: Analyzed 96,147 user queries to identify research needs 2. **Deep Research Identification**: 44,019 queries identified as requiring deep research 3. **Expert Curation**: 100+ domain experts created representative tasks 4. **Validation**: Multiple rounds of review and refinement 5. **Bilingual Adaptation**: Careful translation and cultural adaptation

Evaluation Methodology

RACE Framework

The Report Assessment for Comprehensive Evaluation (RACE) framework assesses the quality of generated research reports^[1]:

Criterion	Weight	Description	Evaluation Method
**Comprehensiveness**	30%	Coverage of relevant aspects	Comparison with reference reports
**Insight/Depth**	25%	Analysis quality and originality	Expert rubric scoring
**Instruction Following**	20%	Adherence to task requirements	Binary and scaled metrics
**Readability**	15%	Clarity and organization	Automated readability scores
**Accuracy**	10%	Factual correctness	Fact-checking against sources

FACT Framework

The Factual Accuracy and Citation Testing (FACT) framework evaluates information retrieval effectiveness:

Metric	Description	Calculation
**Citation Accuracy**	Correctness of cited sources	Verified citations / Total citations
**Effective Citations**	Relevant and supporting citations	Relevant citations / Total citations
**Source Diversity**	Variety of information sources	Unique domains / Total citations
**Citation Density**	Citations per unit of content	Citations / Word count × 1000
**Temporal Relevance**	Recency of cited materials	Weighted by publication date

Current Performance

Leaderboard Results (July 2025)

The DeepResearch Bench leaderboard, hosted on Hugging Face Spaces^[3], shows current model performance:

Rank	Model	RACE Score	FACT Score	Overall	Organization
1	Gemini-2.5-Pro Deep Research	82.3	78.5	80.4	Google
2	OpenAI Deep Research	80.7	76.2	78.5	OpenAI
3	Perplexity Deep Research	78.4	79.1	78.8	Perplexity AI
4	Kimi-Researcher	76.2	74.8	75.5	Moonshot AI
5	Claude-Researcher	75.8	73.4	74.6	Anthropic
6	Doubao-DeepResearch	74.1	72.9	73.5	ByteDance

Performance Analysis by Domain

Domain	Best Performing Model	Average Score	Human Expert Baseline
**Computer Science**	Gemini-2.5-Pro	85.2	92.0
**Physical Sciences**	OpenAI Deep Research	79.8	88.5
**Life Sciences**	Perplexity	77.3	87.0
**Engineering**	Gemini-2.5-Pro	76.5	86.5
**Social Sciences**	Claude-Researcher	72.1	84.0

Technical Implementation

System Requirements

DeepResearch Bench requires the following technical setup^[4]:

```bash

1. Installation

pip install deepresearchbench

1. Required API keys

export GEMINI_API_KEY="your_gemini_key" export JINA_API_KEY="your_jina_key" # For web scraping

1. Python version

Python 3.9+ ```

Evaluation Pipeline

```python from deepresearchbench import Evaluator, RACEScorer, FACTScorer

1. Initialize evaluator

evaluator = Evaluator(

   race_scorer=RACEScorer(),
   fact_scorer=FACTScorer()

)

1. Load benchmark tasks

tasks = evaluator.load_benchmark("path/to/tasks.jsonl")

1. Evaluate a research agent

results = evaluator.evaluate(

   agent=my_research_agent,
   tasks=tasks,
   verbose=True

)

1. Generate detailed report

evaluator.generate_report(results, output_path="evaluation_report.pdf") ```

Data Format

Each task in the benchmark follows a standardized format:

```json {

 "task_id": "DRB_001",
 "domain": "Physics",
 "language": "en",
 "query": "Analyze recent advances in quantum error correction...",
 "reference_sources": [...],
 "expert_annotations": {...},
 "difficulty_level": "PhD",
 "estimated_time_hours": 4.5

} ```

Key Findings and Insights

Agent Capabilities Analysis

DeepResearch Bench reveals several insights about current AI research agents:

Capability	Current State	Gap to Human Expert
**Information Retrieval**	Good (75-85%)	10-15%
**Source Synthesis**	Moderate (60-70%)	20-30%
**Critical Analysis**	Limited (45-55%)	35-45%
**Novel Insights**	Poor (25-35%)	55-65%
**Citation Accuracy**	Good (70-80%)	15-20%

Language Performance Comparison

Model	English Tasks	Chinese Tasks	Bilingual Average
Gemini-2.5-Pro	83.7	78.9	81.3
OpenAI Deep Research	81.2	75.3	78.3
Kimi-Researcher	72.4	80.1	76.3
Doubao-DeepResearch	70.8	77.5	74.2

Limitations and Future Work

Current Limitations

1. **Task Scope**: Limited to 22 academic fields, may not cover all research domains 2. **Language Coverage**: Only English and Chinese, excluding other major research languages 3. **Evaluation Metrics**: RACE and FACT may not capture all aspects of research quality 4. **Human Baseline**: Establishing consistent expert baselines across domains is challenging 5. **Dynamic Information**: Difficulty in evaluating agents on rapidly changing information

Future Directions

Direction	Description	Timeline
**Expanded Languages**	Add support for Spanish, French, German	2025 Q4
**Interactive Tasks**	Multi-turn research dialogues	2026 Q1
**Real-time Evaluation**	Live web research assessment	2026 Q2
**Multimodal Integration**	Include figures, charts, data visualization	2026 Q3
**Collaborative Research**	Multi-agent research scenarios	2026 Q4

Significance

DeepResearch Bench addresses a critical gap in AI evaluation by providing the first comprehensive benchmark for assessing AI systems' ability to conduct PhD-level research. By combining rigorous task design with sophisticated evaluation metrics, the benchmark enables:

• **Standardized comparison** of research agent capabilities
• **Identification of weaknesses** in current AI systems
• **Guidance for development** of next-generation research agents
• **Quality assurance** for AI-generated research content
• **Cross-linguistic evaluation** of research capabilities

As AI systems increasingly assist in research and analysis tasks, DeepResearch Bench provides essential infrastructure for ensuring these systems meet the high standards required for academic and professional research work.

See Also

• Deep Research Agents
• Large Language Models
• AI Research Tools
• Research Automation
• Academic AI
• Citation Analysis
• Deep Research Bench (FutureSearch benchmark)
• Information Retrieval

References