Where Detectors Fail: Closing the Tail-Domain Gap with Expert-Guided Mutual Distillation
Quick Answer
This paper shows that The Expert-Guided Mutual Distillation (EGMD) method enhances multimodal fake news detection by addressing domain-specific biases, achieving state-of-the-art accuracy and reducing domain bias by up to 57.3% across four datasets in two languages.
Quick Take
A new benchmark, Weibo_Balanced, isolates the effect of data imbalance on generalization.
Key Points
- EGMD learns to trust reliable evidence across the prediction pipeline.
- Input-level calibration improves coherence before data fusion.
- Domain-specific patterns are concentrated in specialized experts.
- Prototype-anchored students inherit features while reducing local biases.
- Weibo_Balanced benchmark isolates imbalance effects on generalization.
DeepSignal Analysis
What happened
The Expert-Guided Mutual Distillation (EGMD) method was developed to improve multimodal fake news detection. It addresses issues of domain-specific biases and unreliable evidence, achieving significant accuracy improvements and reducing domain bias across multiple datasets.
Key evidence
- EGMD reduces domain bias by up to 57.3% across four datasets in two languages, indicating a substantial improvement in generalization capabilities.
- The Weibo_Balanced benchmark was created to isolate the effects of data imbalance on generalization, providing a clearer understanding of model performance.
- The method incorporates input-level calibration, representation-level alignment, and decision-level learning to enhance the reliability of evidence used in predictions.
Why it matters
Improving the accuracy of fake news detection systems is critical as misinformation continues to proliferate across various platforms. By addressing domain-specific biases, EGMD could lead to more reliable detection methods that generalize better across different contexts, potentially reducing the spread of false information.
What to watch
Paper Resources
Source Excerpt
Multimodal fake news detectors often generalize poorly across domains because they learn to trust unreliable evidence: domain-specific shortcuts amplified by imbalanced data and semantically inconsistent text-image pairs that make cross-modal evidence unreliable. We propose Expert-Guided Mutual Distillation (EGMD), which learns what evidence to trust across the prediction pipeline. At the input level, input-level calibration encodes pair-level coherence as a shared gain before fusion. At the rep
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.