Hybrid system combining knowledge graphs, AI, and search agents sets new benchmark for fact-checking accuracy

Fact-checking at scale remains a difficult problem: while large language models demonstrate impressive text fluency, they often lack reliable grounding in verified information. Meanwhile, systems built on structured knowledge bases deliver transparent, auditable evidence but struggle with incomplete coverage and slow response times.

A new research paper from arXiv describes a solution that bridges these limitations by combining the strengths of multiple verification approaches into a unified pipeline.

Three-step verification process#

The system operates through a layered architecture designed to handle different scenarios efficiently:

• Knowledge Graph Lookup: The pipeline first attempts rapid searches in DBpedia, retrieving direct factual connections to test claims without external delays.
• Language Model Classification: When initial lookups succeed, a specially designed prompt guides the model's reasoning, producing outputs backed by interpretable rule-based logic.
• Web Search Agent: Only when structured knowledge proves insufficient does the system invoke real-time web search, ensuring coverage while maintaining efficiency.

This modular design allows each component to handle what it does best while fallback mechanisms ensure comprehensive coverage.

Benchmark results and real-world insights#

The hybrid approach achieves an F1 score of 0.93 on the FEVER benchmark's Supported/Refuted classification task—a competitive result achieved without task-specific fine-tuning of any component.

Beyond headline metrics, the researchers conducted a deeper investigation into claims originally flagged as "Not Enough Information." Their reannotation study found that the system frequently uncovers valid supporting or refuting evidence for these cases. Both expert human reviewers and LLM-based evaluators confirmed that evidence the system discovered would justify reclassifying many NEI claims, suggesting that earlier labeling standards may have been overly restrictive.

Implications for deployable fact-checking#

The work demonstrates that interpretability and accuracy need not be at odds. By integrating knowledge graphs' explainability with LLMs' reasoning capacity and search agents' comprehensiveness, the researchers create a system that produces verifiable results. The team released their pipeline as open-source software, signaling potential for adoption across fact-checking platforms and research applications.