🔥 FIRE: Fact-checking with Iterative Retrieval and Verification

FIRE is a novel agent-based framework for fact-checking atomic claims, designed to integrate evidence retrieval and claim verification in an iterative and cost-effective manner. Unlike traditional systems that fix the number of web queries before verifying, FIRE dynamically decides whether to stop or continue querying based on confidence.

🔍 Why FIRE?

Compared to previous systems like FACTCHECKGPT, FACTOOL, and SAFE, FIRE:

• Integrates reasoning and retrieval instead of separating them
• Dynamically controls the retrieval depth
• Reduces LLM cost by 7.6× and search cost by 16.5×
• Performs comparably or better on public datasets like FacTool-QA, FELM-WK, BingCheck

📌 Features

• Iterative agent-based reasoning
• Unified decision function for retrieval or finalization
• Optimized for low-cost verification
• Supports both proprietary and open-source LLMs
• Extensive evaluations and ablations available

🧠 How It Works

Input Claim
   │
   ▼
[FIRE Decision Module]
   ├── confident → Output Label (True / False)
   └── uncertain → Generate Search Query
                      │
                      ▼
          Web Search (via SerperAPI)
                      │
                      ▼
            Update Evidence Set
                      │
                      └── Loop until confident or max steps

📊 Performance Snapshot

🔍 FIRE vs. Baseline Systems

FIRE is compared against state-of-the-art frameworks including FactCheckGPT, FACTOOL, and SAFE.

🔧 Performance Across Datasets

---

💰 Cost and Time Efficiency

🚀 Quickstart

git clone https://github.com/mbzuai-nlp/fire.git
cd fire
pip install -r requirements.txt

# Run FIRE with GPT-4o-mini
python run_fire.py --model gpt-4o-mini --dataset factcheck_bench

📄 Citation

@inproceedings{xie-etal-2025-fire,
 address = {Albuquerque, New Mexico},
 author = {Xie, Zhuohan  and
Xing, Rui  and
Wang, Yuxia  and
Geng, Jiahui  and
Iqbal, Hasan  and
Sahnan, Dhruv  and
Gurevych, Iryna  and
Nakov, Preslav},
 booktitle = {Findings of the Association for Computational Linguistics: NAACL 2025},
 isbn = {979-8-89176-195-7},
 pages = {2901--2914},
 publisher = {Association for Computational Linguistics},
 title = {{FIRE}: Fact-checking with Iterative Retrieval and Verification},
 url = {https://aclanthology.org/2025.findings-naacl.158/},
 year = {2025}
}

👥 Authors

Developed by Zhuohan Xie, Rui Xing, Yuxia Wang, Jiahui Geng, Hasan Iqbal, Dhruv Sahnan, Iryna Gurevych, and Preslav Nakov
Affiliations: MBZUAI, The University of Melbourne

For questions or collaborations, contact:
📬 zhuohan.xie@mbzuai.ac.ae

---

“Fact-checking, now with FIREpower.”