Research map / Large language models / Language, safety and interpretability

Collaborative Evolution: Multi-Round Learning Between Large and Small Language Models for Emergent Fake News Detection

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Data and hallucination detection · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MRCD, a multi-round collaboration framework where an LLM (Llama3-8B etc.) with two-stage retrieval demonstrations and Wikipedia knowledge and a fine-tuned SLM (RoBERTa or FTT) iteratively pseudo-label and select clean data for emergent fake news detection.

Problem

Detecting fake news about emergent events when no annotated data exists and the distribution shifts over time; SLMs need supervised data and frozen LLMs perform poorly zero/few-shot.

Approach

Combine LLM in-context learning (with pseudo-labeled demonstrations retrieved from search engines, a news corpus, and Wikipedia knowledge) and an SLM fine-tuned on confidently agreed pseudo-labels, iterating over rounds so each model's clean pool improves the other.

Key results

• MRCD (Llama3+RoBERTa) improves accuracy over RoBERTa by 7.4% on Pheme and 12.8% on Twitter16.
• Llama3+FTT achieves the best results: 0.814 accuracy on Pheme and 0.794 on Twitter16.
• Zero-shot LLMs (Llama2-7B, Llama3-8B, GPT-3.5) perform near chance on Pheme (accuracy 0.503-0.535).

Measured effects reported in the paper

Surprising findings

• The SLM's performance declines from round 4 onward as noisy samples contaminate the clean pool, while the LLM is less affected because in-context learning relies more on demonstration content than label accuracy.
• Pseudo-label synonyms (e.g., realistic/unrealistic) for demonstrations outperform the true task labels (real/fake) for LLM in-context learning, attributed to avoiding copy effects.

Limitations

• Two-stage retrieval (search and Wikipedia) adds time and resource cost.
• Hyperparameters (round threshold N, confidence threshold omega) are selected manually.
• Performance of the SLM declines in later rounds as noisy pseudo-labels accumulate in the clean pool.

Experimental setup

Datasets: Pheme Twitter16 Wikipedia Bing News search news corpus

Benchmarks: Pheme Twitter16

Metrics: accuracy precision recall F1

Baselines: RoBERTa EANN M3FEND FTT Llama2-7B zero-shot/few-shot Llama3-8B zero-shot/few-shot GPT-3.5 zero-shot/few-shot ARG

Builds on: ARG BM25 FTT RoBERTa

Closest papers by meaning

• Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.89)
• DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.85)
• Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.85)
• External Reliable Information-enhanced Multimodal Contrastive Learning for Fake News Detection (AAAI 2025 · similarity 0.85)
• Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.84)
• Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.84)