EARD: An Imitation Agent Framework for LLM-based Early Rumor Detection

Core Overview of the EARD Framework

EARD (LLM-based Early Rumor Detection with Imitation Agent) is a lightweight LLM-driven early rumor detection method that achieves training-free, data-efficient detection through an imitation agent mechanism. This framework addresses the challenges of traditional rumor detection in scenarios with scarce data and high timeliness requirements, demonstrating excellent performance in early social media rumor identification tasks. By combining the semantic understanding capabilities of LLMs with the advantages of lightweight reasoning, it provides key technical support for information governance.

Research Background: Timeliness Challenges in Rumor Detection

In the era of social media, rumors spread rapidly and have a wide impact. Early identification is key to information governance. Traditional methods have limitations:

• Deep learning methods rely on large amounts of labeled data, but early rumor data is scarce
• Feature engineering methods rely on manual design and struggle to adapt to the rapid evolution of rumor forms The emergence of LLMs offers new ideas, but how to efficiently utilize LLMs in the early stage of scarce data while controlling computational costs remains an open problem. The EARD research proposes a solution to this challenge.

Core Innovation: Imitation Agent Mechanism

Design Principles

EARD introduces an imitation agent mechanism that mimics the decision-making process of human experts: extracting key information, comparing with existing knowledge, evaluating source credibility, and making comprehensive judgments. It uses phased lightweight reasoning (information extraction → evidence evaluation → comprehensive judgment) to reduce computational costs.

Technical Architecture

It includes four main components:

• Sequence Encoder: Encodes post time sequence, interaction features, and semantic information
• Imitation Agent: Performs observation, reasoning, decision-making, and learning functions
• Evidence Aggregator: Integrates multi-source evidence and handles conflicts
• Time Series Processor: Models propagation dynamics and extracts early signals

Method Details

Few-shot Context Learning

Guiding the model through carefully designed examples:

• Example selection: Similar cases, covering different types, including reasoning processes
• Prompt engineering: Clear task objectives, structured output, guiding reasoning
• Dynamic update: Optimizing the example library based on feedback

Training-free Adaptability

• Zero-shot capability: Cross-language, cross-platform, and new-type rumor identification
• Rapid deployment: No labeled data or fine-tuning required
• Continuous adaptation: Adapting to new scenarios by updating examples

Early Detection Optimization

• Sparse data reasoning: Probabilistic judgment and confidence evaluation
• Incremental information fusion: Dynamic judgment update and information gain evaluation
• Early signal extraction: Identifying propagation patterns and key node behaviors

Experimental Evaluation and Results

Datasets

Experiments were conducted on four public datasets: Twitter15/Twitter16, Weibo, and PHEME.

Evaluation Metrics

• Accuracy: Accuracy, Precision, Recall, F1-Score
• Early detection: Earliness, accuracy at different time points
• Efficiency: Inference latency, resource consumption

Key Results

• Leading performance: Outperforms existing state-of-the-art methods, with more obvious advantages in early detection
• Cross-domain adaptation: Adapts to new platforms and languages without training
• Computational efficiency: Inference speed is better than large end-to-end models

Ablation Experiments

Verifies the key contributions of the imitation agent, few-shot examples, and time series modeling.

Application Value and Scenarios

Social Media Platforms

Real-time content moderation, trend monitoring, user prompts

News Organizations

Clue discovery, preliminary evaluation, accelerating fact-checking

Government and Public Sectors

Public opinion monitoring, public communication, emergency response support

Enterprise Brand Protection

Brand monitoring, competitor analysis, crisis PR initiation

Limitations and Future Directions

Current Limitations

• Limited multi-modal content processing capability
• Insufficient deep reasoning for complex rumors
• Robustness against adversarial attacks needs to be enhanced
• Real-time processing of ultra-large-scale data streams needs optimization

Future Directions

• Technical improvements: Introducing external knowledge via RAG, multi-modal detection, cross-platform tracking
• Application expansion: Specific domains (medical/finance), multi-language, source tracing
• System integration: Integration with fact-checking databases, human-machine collaborative moderation

Summary

EARD combines LLM semantic understanding with lightweight reasoning through an imitation agent mechanism to achieve training-free, data-efficient early rumor detection. Its significance lies in:

• Methodological innovation: Proposes a new paradigm of imitation agents, demonstrating the application potential of LLMs in data-scarce scenarios
• Practical value: Can be deployed immediately, lowers technical barriers, and supports real-time large-scale applications
• Research implications: Re-examines the deep learning paradigm dependent on training data, promoting research on efficient AI applications This framework provides a powerful tool for building a healthy information ecosystem.