Early Rumor Detection for LLMs Using Imitation Agents: Keeping Large Models in Zero-Training State

The research team from Singapore Management University proposed the EARD framework, which combines autonomous agents with LLM detection models to solve the problem of early rumor detection. The agent handles temporal decision-making, while the LLM remains in a zero-training state and is only used for inference. The framework outperforms existing methods on four real datasets, balancing accuracy and timeliness.

Challenges of Early Rumor Detection

In the era of social media, rumors spread rapidly. Early detection requires identifying the earliest reliable classification point when data is scarce. Traditional methods rely on large amounts of annotations or delayed judgments, making timely intervention difficult.

Dilemmas of LLMs

LLMs excel at static text but are not adapted to temporal decision-making. Additionally, their training and inference costs are high, making it hard to process massive social data in real time.

EARD Framework Design

The agent (temporal decision-maker) determines whether information is sufficient to trigger detection, while the LLM focuses on semantic analysis and classification. This division of labor achieves complementary advantages, and the LLM remains in a zero-training state.

Imitation Learning Training

The agent learns when to make decisions by imitating expert strategies, quickly adapting from a small number of demonstrations to balance accuracy and timeliness.

Validated on four real datasets:

1. Versatility: Multiple LLM backbone models achieved performance improvements;
2. Superiority: Outperformed existing methods in both accuracy and earliness;
3. Low cost: Zero-training for LLMs reduces deployment and maintenance costs.

1. Platforms can integrate EARD into real-time content streams to monitor suspicious topics in real time and trigger manual review or automatic labeling promptly;
2. Small and medium-sized platforms can also deploy it to enhance content governance capabilities.

Technical Limitations

Evaluated based on historical datasets; performance in dynamic social environments remains to be verified; 2. Mainly targeted at English content; cross-language/cultural effects need further research.

Future Directions

Multimodal detection, cross-language transfer, and adversarial rumor recognition.

The EARD framework demonstrates the value of collaboration between different components. By dividing labor between agents and LLMs to solve complex problems, it provides a reference idea of "complementary expertise" for AI applications.