FBD boosts LLM fact verification via debates

TL;DR

FBD Method: Researchers introduce Fact-Based Debating (FBD), a multi-agent debate using large language models for fact verification.[[1]](https://ieeexplore.ieee.org/document/11229172)[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)
Superior Performance: Experiments on standard datasets show FBD outperforms prior methods while using fewer resources.[[1]](https://ieeexplore.ieee.org/document/11229172)[[3]](https://ieeexplore.ieee.org/iel8/11227166/11227148/11229172.pdf)
Misinformation Tackle: FBD addresses online misinformation challenges through guided, fact-grounded LLM debates.[[1]](https://ieeexplore.ieee.org/document/11229172)

The story at a glance

A team led by Mufan Yu from Tianjin University's College of Intelligence and Computing presents FBD, a novel approach for fact verification that employs multi-agent debating with large language models (LLMs). The paper argues that grounding debates in provided evidence improves accuracy over existing LLM-based methods. It appears now as part of the 2025 International Joint Conference on Neural Networks (IJCNN 2025) proceedings.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)

Key points

Proposes FBD to combat misinformation proliferation on internet and social media via LLM-driven debates.
Uses multi-agent framework where agents debate claims strictly based on given evidence (index terms: Fact Verification, Large Language Models, Multi-agent Debate).[[3]](https://ieeexplore.ieee.org/iel8/11227166/11227148/11229172.pdf)
Experiments on standard fact verification datasets confirm FBD's superior performance compared to prior techniques.[[1]](https://ieeexplore.ieee.org/document/11229172)
Achieves high accuracy with lower computational cost than competing methods.[[3]](https://ieeexplore.ieee.org/iel8/11227166/11227148/11229172.pdf)
8-page conference paper from IJCNN 2025, DOI 10.1109/IJCNN64981.2025.11229172.[[4]](https://www.researchgate.net/publication/397633849_FBD_Fact-Based_Debating_for_Fact_Verification_through_Large_Language_Models)

Details and context

The paper originates from researchers primarily at Tianjin University's College of Intelligence and Computing: Mufan Yu, Guozheng Rao, Xin Wang, Kaijia Tian, Jiayin Zhang; with Li Zhang from Tianjin University of Science and Technology's School of Economics and Management.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)

Fact verification typically involves checking claims against evidence, a task where LLMs struggle due to hallucinations or bias; FBD enforces fact-based arguments in a debate format to refine verdicts.[[1]](https://ieeexplore.ieee.org/document/11229172)

Presented at IJCNN 2025 (June 30-July 5, Rome, Italy), it fits into broader neural networks research on reliable AI amid rising LLM applications.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)

Full text is paywalled on IEEE Xplore; details drawn from abstract, metadata, and conference program.[[5]](https://ieeexplore.ieee.org/document/11229172/)

Key quotes

Omitted; no direct quotes visible in accessible sources.

Why it matters

Fact verification tools like FBD are crucial as misinformation undermines public discourse, especially with AI-generated content surging. For AI developers and researchers, it offers a practical way to boost LLM reliability in evidence-based tasks without heavy compute. Watch for peer reviews, code releases, or extensions to real-time social media verification, though full benchmarks await open access.[[1]](https://ieeexplore.ieee.org/document/11229172)

FAQ

Q: What is the FBD method in this paper?

A: FBD is Fact-Based Debating, a multi-agent system where LLMs debate claims using only provided evidence to reach accurate fact verification. It structures arguments around facts to reduce errors common in standalone LLMs. The approach is detailed in an 8-page IJCNN paper.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)

Q: Who authored the FBD paper?

A: Lead author Mufan Yu and co-authors Guozheng Rao, Xin Wang, Li Zhang, Kaijia Tian, Jiayin Zhang, mostly from Tianjin University's College of Intelligence and Computing. One from Tianjin University of Science and Technology. Affiliations confirm academic focus on AI.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)

Q: How does FBD perform compared to others?

A: Experiments on standard datasets show FBD outperforms existing methods in accuracy. It also maintains strong results with lower resource use than alternatives. Specific metrics are in the paywalled full text.[[3]](https://ieeexplore.ieee.org/iel8/11227166/11227148/11229172.pdf)

Q: When and where was this research presented?

A: Published in IJCNN 2025 proceedings, held June 30 to July 5 in Rome, Italy. The conference program lists it among technical papers on neural networks and AI.[[2]](https://confcats-siteplex.s3.us-east-1.amazonaws.com/ijcnn25/IJCNN_2025_Program_77b2d8aef4.pdf)