ReportBench: Evaluating Deep Research Agents via Academic Survey Tasks

If you like our project, please give us a star ⭐ on GitHub for the latest update.

Overview

ReportBench is a comprehensive benchmark for evaluating the factual quality and citation behavior of Deep Research agents. Leveraging expert-authored survey papers as ground truth, ReportBench reverse-engineers domain-specific prompts and provides automated tools to assess both cited and non-cited content.

ReportBench addresses this need by:

• Leveraging expert surveys: Uses high-quality, peer-reviewed survey papers from arXiv as gold-standard references.
• Reverse-engineering prompts: Generates task-specific prompts matching each survey’s scope, methods, and temporal constraints.
• Automated validation: Employs a dual-path evaluation to verify citation consistency and factual correctness of non-cited statements.

---

Benchmark Construction

The dataset construction pipeline consists of four phases:

1. Survey Paper Identification

   • Start from an arXiv metadata snapshot (post-2020).
   • Filter titles/abstracts for “survey” or “review” and confirm publication status via metadata and LLMs classification.
   • Retain 600 high-quality, peer-reviewed survey papers.

2. Fine-Grained Reference Extraction

   • Download and parse LaTeX sources to extract all in-text citation commands.
   • Build a gold-standard set of references mirroring the true citation pattern of each survey.

3. Prompt Generation

   • Reverse-engineer three levels of prompts (sentence, paragraph, detail-rich) via LLMs.
   • Enforce temporal constraints matching each paper’s publication cutoff.
   • Add explicit instructions to avoid citing the original survey itself.

4. Application Domain Distribution

   • Classify surveys into ten domains using LLMs.
   • Perform downsampling while adjusting the distribution to achieve balance, and sample one of the three prompt types to form a 100-task benchmark.

---

Evaluation Framework

ReportBench’s evaluation workflow consists of two complementary validation procedures:

1. Content Quality

• URL Extraction: Extract all URL citations from the report, including from the base model and Deep Research outputs.
• Normalization and Retrieval: Normalize and deduplicate URLs, then retrieve the content of each web page.
• Document Type Classification: Use an LLM to determine whether each URL corresponds to a scholarly article and extract its title if applicable.
• Title Matching: Compare the extracted titles against ground-truth references from the expert-authored report and compute an overlap ratio.

2.1. Cited Statements

• Statement Extraction: Identify all sentences in a generated report containing explicit citations.
• Source Retrieval: Scrape the full text of each cited source.
• Semantic Matching: Use an LLM to locate supporting passages and verify consistency.
• Scoring: Compute a citation alignment score for each report.

2.2. Non-Cited Statements

• Statement Extraction: Extract factual claims without citations, filtering out common-sense content.
• Web-Connected Fact Checking: Query multiple web-connected LLMs (Gemini Pro and Flash) to independently verify each claim.
• Voting Mechanism: Aggregate judgments via majority vote to compute factual accuracy.

---

Evaluation Results

We evaluated two Deep Research products alongside their corresponding base LLMs using the ReportBench benchmark. Table 1 summarizes precision, recall, average references per report, citation match rate, cited statement count, non-cited factual accuracy, and non-cited statement count.

Test Model	Precision	Recall	Avg Refs	Cit. Match Rate	Cit. Stmt Count	Non-Cit Acc	Non-Cit Stmt Count
OpenAI Deep Research	0.385	0.033	9.89	78.87%	88.2	95.83%	38.9
Gemini Deep Research	0.145	0.036	32.42	72.94%	96.2	92.21%	49.6
gemini-2.5-flash	0.237	0.012	5.47	44.88%	12.1	98.52%	11.5
gemini-2.5-pro	0.269	0.010	4.27	59.24%	6.58	96.08%	9.35
o3	0.299	0.031	12.26	31.43%	16.16	82.22%	11.51
claude4-sonnet	0.337	0.021	6.74	73.67%	14.93	92.64%	17.07

Table 1. Performance metrics of Deep Research products and their base models.

Product-Level

• OpenAI Deep Research: Highest precision (0.385) and citation match rate (78.87%), indicating focused and accurate retrieval with fewer references.
• Gemini Deep Research: Generates many more citations (32.42 vs. 9.89) but yields only marginal recall gain, suggesting over-generation without proportional coverage benefits.

Model-Level

• OpenAI vs. o3: Comparable retrieval metrics, but Deep Research produces far more cited statements (88.2 vs. 16.16) and non-cited statements (38.9 vs. 11.51), and achieves much higher alignment (78.87% vs. 31.43%) and accuracy (95.83% vs. 82.22%).
• Gemini Deep Research vs. gemini-2.5-pro: Trades off precision (0.145 vs. 0.269) for higher recall and citation volume, while maintaining strong alignment (72.94% vs. 59.24%) but slightly lower non-cited statement accuracy.
• claude4-sonnet: Most balanced baseline—moderate precision (0.337), recall (0.021), citation consistency (73.67%), and non-cited statement accuracy (92.64%).

---

🛠️ Installation and Usage

Prerequisites

Environment Requirements:

• Python 3.8+
• Required Python packages (install via pip)

Install Dependencies:

pip install pandas pyyaml langchain-openai tenacity tqdm requests pathlib beautifulsoup4 firecrawl-py python-dotenv

API Keys Setup: Create a .env file in the project root with the following configurations:

# OpenAI Configuration
OPENAI_PROVIDER=openai  # or "azure" for Azure OpenAI
OPENAI_API_KEY=your_openai_api_key_here
OPENAI_MODEL_NAME=gpt-4o-mini
TEMPERATURE=0.0
MAX_TOKENS=8192

# Azure OpenAI (if using Azure)
AZURE_OPENAI_ENDPOINT=your_azure_endpoint
AZURE_OPENAI_API_VERSION=2023-12-01-preview
AZURE_OPENAI_DEPLOYMENT_NAME=your_deployment_name
AZURE_OPENAI_API_KEY=your_azure_api_key

# Web Scraping (Required for citation verification)
FIRECRAWL_API_KEY=your_firecrawl_api_key

# Search API (Optional, for enhanced fact-checking)
SERPAPI_API_KEY=your_serpapi_key

Project Structure

ReportBench/
|
├── 📁 ReportBench Release v1.1
│   ├── ReportBench_v1.1.jsonl      # Main dataset file in JSON Lines format
│   └── ReportBench_v1.1_GT         # Ground truth reference data
|
├── 📁 Core Processing Scripts
│   ├── openai_processor.py          # Process OpenAI Deep Research outputs
│   ├── gemini_processor.py          # Process Gemini Deep Research outputs  
│   ├── statement_evaluator.py       # Extract and evaluate factual statements
│   ├── related_work_evaluator.py    # Evaluate citation accuracy and recall
│   └── metrics_calculator.py        # Calculate final performance metrics
│
├── 📁 Configuration & Utilities
│   ├── config.py                    # API keys and model configurations
│   ├── utils.py                     # Common utilities (LLM clients, CSV ops)
│   ├── cache_utils.py               # URL caching and normalization
│   └── .env                         # Environment variables (create this)
│
├── 📁 Evaluation Modules
│   ├── statement/                   # Statement extraction and verification
│   │   ├── extract_citations.py    # Extract cited statements
│   │   ├── extract_no_citations.py # Extract non-cited statements
│   │   ├── scrape_content.py       # Web scraping for citation sources
│   │   ├── match_text.py           # Semantic matching of statements
│   │   ├── verify_alignment.py     # Verify citation-statement alignment
│   │   └── verify_no_citations_web.py # Web-based fact-checking
│   └── process/                     # Data processing utilities
│       ├── extract_activity_structured.py
│       ├── extract_reference_structured.py
│       └── html2markdown.py
│
├── 📁 Scripts & Templates
│   ├── run_test.sh                  # Main evaluation pipeline
│   ├── process_prompt.py            # Prompt processing utilities
│   └── prompt_template/             # Evaluation prompt templates
│
└── 📄 Configuration Files
    ├── README.md                    # Project documentation
    ├── .gitignore                   # Git ignore rules
    └── .env.example                 # Environment variables template

Key Components:

• Processing Pipeline: openai_processor.py → statement_evaluator.py + related_work_evaluator.py → metrics_calculator.py
• Input Format: Model outputs as JSON files with response/content field containing the generated survey text
• Output: Comprehensive evaluation metrics including citation alignment scores and factual accuracy rates

Quick Start

1. Prepare Your Model Data

Important Note: The data preparation process differs based on your model type:

For OpenAI Deep Research & Gemini Deep Research (Web-based)

These are web-based products that require special data collection:

1. Use Chrome Extension: Since these are web interfaces, you need to use a Chrome extension to capture the conversation records
2. Process with Dedicated Scripts: Use the corresponding processor to parse the captured data:

   # For OpenAI Deep Research
   python openai_processor.py --input=captured_data_dir --output=parsed_output_dir --markdown
   
   # For Gemini Deep Research  
   python gemini_processor.py --input=captured_data_dir --output=parsed_output_dir

For Other Models (API/Local Models)

Input Format: Your model outputs should be saved as JSON files with the following structure:

{
  "response": "Your model's generated survey text here...",
  "arxiv_id": "2024.12345",  // Optional: will be extracted from filename if not present
  "query": "Original query prompt",  // Optional
  // ... other metadata fields
}

Alternative accepted field names for the main content:

• response, content, text, message, output, result

File Organization:

# Create your model's output directory
mkdir -p your-model-name

# Save each evaluation result as: {arxiv_id}.json
# Example filenames:
your-model-name/2003.00653.json
your-model-name/2004.05937.json
# ... (100 files total for the full benchmark)

2. Run Evaluation Pipeline

# 1. Process raw outputs (ONLY for web-based Deep Research products)
# For OpenAI Deep Research:
python openai_processor.py --input=data/test_data/raw_data/your-model-name --output=your-model-name-parsed --markdown
# For Gemini Deep Research:
python gemini_processor.py --input=data/test_data/raw_data/your-model-name --output=your-model-name-parsed --markdown
# For other models: Skip this step, use your JSON files directly

# 2. Extract and evaluate statements
python statement_evaluator.py your-model-name-parsed --output-dir your-model-name-stat-results

# 3. Evaluate citation accuracy  
python related_work_evaluator.py --survey-dir your-model-name-parsed --ground-truth-dir ReportBench_v1.1_GT --result-dir your-model-name-related-work-results

# 4. Calculate final metrics
python metrics_calculator.py your-model-name-stat-results

3. View Results

Results will be saved in structured directories:

• Statement Evaluation: your-model-name-stat-results/ - Contains citation alignment and factual accuracy scores
• Related Work Evaluation: your-model-name-related-work-results/ - Contains precision/recall for citation discovery
• Final Metrics: Summary CSV files with aggregated performance metrics

Key Output Files:

• Individual paper results in {arxiv_id}/ subdirectories

4. HuggingFace Dataset (Optional)

You can easily access ReportBench with the following code:

from datasets import load_dataset
dataset = load_dataset("ByteDance-BandAI/ReportBench")

Citation

@misc{li2025reportbenchevaluatingdeepresearch,
      title={ReportBench: Evaluating Deep Research Agents via Academic Survey Tasks}, 
      author={Minghao Li and Ying Zeng and Zhihao Cheng and Cong Ma and Kai Jia},
      year={2025},
      eprint={2508.15804},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.15804}, 
}