WKGFC: A New AI System Combines Knowledge Graphs and Multi-Agent Reaso

When a false claim spreads across social media, the damage often happens faster than fact-checkers can respond. By the time a human analyst verifies the information, millions may have already seen and shared it. A team of researchers from Australia and Qatar believes artificial intelligence can change this equation—and their new system suggests they might be right.

The researchers have developed WKGFC, a framework that combines knowledge graphs with multi-agent AI reasoning to automatically verify claims against structured evidence. Unlike existing systems that rely on simple text matching, WKGFC uses a reasoning agent that navigates knowledge graphs like a detective following leads, retrieving connected facts and assessing their relevance to the claim at hand.

“Misinformation spreading over the Internet poses a significant threat to both societies and individuals, necessitating robust and scalable fact-checking that relies on retrieving accurate and trustworthy evidence.” — Research Team

The Knowledge Graph Approach

Traditional fact-checking systems rely on semantic similarity—matching words in a claim to words in potential evidence documents. This approach works for straightforward cases but struggles with complex claims that require connecting multiple pieces of information.

WKGFC takes a different path. It treats authorized knowledge graphs as core evidence sources, using LLM-enabled retrieval to assess claims and extract the most relevant knowledge subgraphs. These structured evidence networks capture relationships between entities that text-based methods often miss.

The multi-hop reasoning capability is what sets WKGFC apart. When evaluating a claim like “Company X acquired Company Y in 2023,” the system doesn’t just search for documents containing those words. It navigates the knowledge graph to find acquisition relationships, cross-references dates, and validates the entities involved—mimicking how a human fact-checker would approach the problem.

Multi-Agent Evidence Retrieval

The system implements an automatic Markov Decision Process (MDP) where a reasoning LLM agent decides what actions to take based on current evidence and the claim being evaluated. This agentic approach allows the system to adapt its retrieval strategy dynamically.

Prompt optimization fine-tunes the agent’s behavior specifically for fact-checking tasks. Rather than using a generic LLM prompt, the researchers optimized the agent to recognize when it has sufficient evidence versus when it needs to search further—a critical capability for balancing accuracy against computational cost.

Web content augmentation complements the knowledge graph evidence. When the structured data is incomplete, the system retrieves additional web content to fill gaps, creating a hybrid approach that leverages both curated knowledge bases and the broader internet.

“Previous methods rely on semantic and social-contextual patterns learned from training data, which limits their generalization to new data distributions.” — Research Authors

Addressing the Generalization Problem

One of the persistent challenges in automated fact-checking is generalization. Systems trained on one domain or time period often fail when faced with claims from different contexts. The researchers designed WKGFC to address this limitation by grounding verification in structured knowledge rather than pattern matching.

Knowledge graphs provide a more stable foundation than training data because they represent verified facts about the world rather than statistical patterns from text. When political circumstances change or new scientific discoveries emerge, updating a knowledge graph is more straightforward than retraining a machine learning model.

The multi-sourced approach also improves robustness. By combining evidence from knowledge graphs, web content, and the reasoning process itself, WKGFC reduces dependence on any single information source—a design choice that makes the system more resilient to source-specific biases or gaps.

Implications for the Misinformation Ecosystem

The release of WKGFC comes as platforms struggle to contain misinformation at scale. Human fact-checkers cannot possibly evaluate the volume of claims generated daily across social media, news sites, and messaging apps. Automated systems that can operate in real-time represent a potential solution—if they can achieve sufficient accuracy.

The researchers’ approach of combining structured knowledge with agentic reasoning reflects broader trends in AI development. Rather than relying solely on ever-larger language models, the field is increasingly exploring how to augment AI with external knowledge sources and structured reasoning processes.

For fact-checking specifically, the knowledge graph approach offers advantages in explainability. When WKGFC verifies or refutes a claim, the reasoning path through the knowledge graph provides a transparent trail that human reviewers can audit—addressing one of the key concerns about automated content moderation.

The coming months will reveal whether systems like WKGFC can achieve the accuracy and speed necessary for real-world deployment. The technical approach is promising, but the ultimate test will be performance against the adversarial creativity of misinformation producers.

This article was reported by the ArtificialDaily editorial team. For more information, visit arXiv.