Researchers blend real and synthetic data to improve fake news detection

When news stories spread online, the patterns of who shares and engages with them reveal clues about whether the content is genuine or fabricated. But this propagation data is often incomplete—many user interactions go unrecorded or invisible. Researchers have begun using large language models to fill these gaps by simulating missing interactions, yet those synthetic signals carry their own reliability problems.

The core challenge#

Fake news detection systems typically rely on how information propagates through social networks. When actual propagation data is incomplete, researchers can employ LLMs to role-play missing users and generate synthetic engagement. However, mixing unreliable simulated data directly with real observations can distort the model's learned representations and undermine detection accuracy.

A new framework for fusion#

Computer scientists have developed InfoPDF (Information-theoretic Propagation Denoising and Fusion), an approach that treats synthetic propagation graphs as probabilistic distributions and intelligently combines them with real data. The framework uses principles from information theory—specifically mutual information—to filter out noise while preserving the most useful signals.

The method operates in two key phases:

• Generation: LLMs create attribute-specific synthetic propagation patterns tailored to the news story's characteristics
• Fusion: A reliability-aware mechanism learns which signals matter most, suppressing noise across synthetic propagation while maintaining alignment between real and synthetic representations

Results and insights#

Testing on three real-world datasets showed consistent improvements across multiple fake news detection tasks. Beyond accuracy gains, the framework also reveals attribute-level reliability scores, showing researchers which synthetic signals are trustworthy and producing more interpretable propagation representations for distinguishing false from authentic content.

This approach suggests that carefully structured fusion of synthetic and real data—guided by information-theoretic principles—can compensate for incomplete online traces while maintaining detection robustness.