LLMs Are Not Enough for Multimodal Fake News Detection: Why Global Label Propagation Helps

2026-04-26 · Source: NLP on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

The ACL 2025 paper "Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection" introduces GLPN-LLM, a framework designed to enhance multimodal fake news detection by integrating Large Language Models (LLMs) more effectively. The authors argue that LLMs, while powerful, are insufficient as direct classifiers due to noise, lack of task-specific optimization, and the multimodal nature of fake news. Instead, GLPN-LLM treats LLMs as "noisy but useful weak annotators," generating initial pseudo-labels. These labels are then refined through global label propagation, leveraging relationships among multimodal samples (textual, visual, semantic) to smooth and correct individual LLM predictions. This approach aims to make LLM-generated supervision more reliable by considering the dataset's global structure, moving beyond isolated LLM outputs.

Key takeaway

For research scientists developing misinformation detection systems, directly using LLMs as classifiers is suboptimal due to their inherent noise and lack of task-specific alignment. You should instead integrate LLMs as weak supervision sources, refining their initial pseudo-labels through global label propagation over multimodal sample relationships. This approach, exemplified by GLPN-LLM, improves reliability by leveraging dataset structure, leading to more robust detection models.

Key insights

LLMs are better as weak annotators than direct classifiers for multimodal fake news detection.

Principles

• Fake news detection requires task-specific supervision.
• LLM predictions can be noisy and unreliable.
• Global data structure improves label reliability.

Method

GLPN-LLM combines LLM-generated pseudo-labels, multimodal representation learning, and global label propagation to refine noisy LLM signals using dataset structure.

In practice

• Integrate LLMs as auxiliary knowledge sources.
• Explicitly model noisy LLM outputs.
• Consider global structure in social media tasks.

Topics

• Multimodal Fake News Detection
• Global Label Propagation
• LLM Weak Supervision
• Pseudo-label Learning
• GLPN-LLM Framework

Code references

• TSCenter/GLPN-LLM

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• A Distribution-Free Framework for Rewrite-Based Human-text Detection via Knockoff Filtering — A distribution-free framework provides finite-sample FDR guarantees for rewrite-based human-text detectors without retraining.
• Beyond Surface Statistics: Robust Conformal Prediction for LLMs via Internal Representations — Internal LLM representations offer more robust uncertainty signals for conformal prediction, especially under distribution shifts.
• A Few-Shot LLM Framework for Extreme Day Classification in Electricity Markets — LLMs can effectively classify electricity price spikes with limited data by interpreting natural language prompts of system conditions.
• KITE: A Tri-Modal Transformer Integrating Text, Images, and Knowledge Graphs for Fake News Detection — KITE integrates text, images, and knowledge graphs via a tri-modal transformer for enhanced fake news detection and interpretability.
• Explain the Flag: Contextualizing Hate Speech Beyond Censorship — A hybrid LLM and vocabulary approach improves hate speech detection and explanation across multiple languages.
• CommunityFact: A Dynamic, Multilingual, Multi-domain Benchmark for Misinformation Detection in the Wild — CommunityFact offers a dynamic benchmark revealing LLM misinformation detection challenges and web-search benefits.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NLP on Medium.