Research map / Large language models / Language, safety and interpretability
Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection
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.
Shows GPT-3.5 underperforms a fine-tuned BERT at fake news detection but gives useful multi-perspective rationales, and proposes ARG (and distilled ARG-D) where BERT selectively absorbs these rationales.
Problem
Small language models used for fake news detection lack knowledge and reasoning, and it is unclear whether LLMs can help; direct LLM prompting underperforms fine-tuned SLMs.
Approach
Treat the LLM as an advisor rather than a detector: generate perspective-specific rationales (textual description, commonsense) from GPT-3.5, and train a BERT-based network with news-rationale cross-attention, an LLM-judgment predictor, and a rationale usefulness evaluator to adaptively select rationales; distill the result into a rationale-free model.
Key results
- ARG reaches 0.784 macro F1 on Chinese (Weibo21) and 0.790 on English (GossipCop), versus 0.753 and 0.765 for fine-tuned BERT.
- GPT-3.5-turbo few-shot reaches 0.725 (Chinese) and few-shot CoT 0.702 (English), below fine-tuned BERT in both datasets.
- ARG-D, the rationale-free distilled model, reaches 0.771 (Chinese) and 0.778 (English) macro F1, still above all SLM-only baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ARG instead of the BERT baseline | BERT-based fake news detection · Weibo21 test | macro F1: 0.753 -> 0.784 | Helped |
| Use ARG instead of the BERT baseline | BERT-based fake news detection · GossipCop test | macro F1: 0.765 -> 0.790 | Helped |
| Distill ARG into a rationale-free model | BERT-based fake news detection · Weibo21 test | macro F1: 0.753 -> 0.771 | Helped |
| Distill ARG into a rationale-free model | BERT-based fake news detection · GossipCop test | macro F1: 0.765 -> 0.778 | Helped |
| Add the LLM Judgment Predictor (ablation) | ARG · Weibo21 test | macro F1: 0.773 -> 0.784 | Helped |
| Add the LLM Judgment Predictor (ablation) | ARG · GossipCop test | macro F1: 0.786 -> 0.790 | Helped |
| Add the Rationale Usefulness Evaluator (ablation) | ARG · Weibo21 test | macro F1: 0.781 -> 0.784 | Helped |
| Add the Rationale Usefulness Evaluator (ablation) | ARG · GossipCop test | macro F1: 0.782 -> 0.790 | Helped |
| Add both the LLM Judgment Predictor and Rationale Usefulness Evaluator (ablation) | ARG · Weibo21 test | macro F1: 0.769 -> 0.784 | Helped |
| Add both the LLM Judgment Predictor and Rationale Usefulness Evaluator (ablation) | ARG · GossipCop test | macro F1: 0.780 -> 0.790 | Helped |
| Fine-tune BERT for the task instead of using GPT-3.5-turbo few-shot prompting | Fake news detection · Weibo21 test | macro F1: 0.725 -> 0.753 | Helped |
| Fine-tune BERT for the task instead of using GPT-3.5-turbo few-shot CoT prompting | Fake news detection · GossipCop test | macro F1: 0.702 -> 0.765 | Helped |
Surprising findings
- GPT-3.5 underperforms fine-tuned BERT under all four prompting strategies, even though its rationales are informative.
- Chain-of-thought prompting sometimes decreased performance (e.g., few-shot CoT on Chinese: 0.725 to 0.681 for zero/few-shot comparison).
Limitations
- Only GPT-3.5-turbo was evaluated; Claude and Ernie Bot were unavailable.
- Only textual description and commonsense perspectives were used in the final model; factuality was excluded due to hallucination.
- Best results still fall short of the oracle voting upper bound, leaving headroom.
Experimental setup
Datasets: Weibo21 GossipCop
Benchmarks: Weibo21 GossipCop
Metrics: macro F1 accuracy F1 real F1 fake
Baselines: GPT-3.5-turbo zero-shot/few-shot/CoT prompting BERT fine-tuning EANN_T Publisher-Emo ENDEF Baseline+Rationale SuperICL
Builds on: BERT chain-of-thought prompting EANN SuperICL knowledge distillation
Closest papers by meaning
- Collaborative Evolution: Multi-Round Learning Between Large and Small Language Models for Emergent Fake News Detection (AAAI 2025 · similarity 0.89)
- DAMMFND: Domain-Aware Multimodal Multi-view Fake News Detection (AAAI 2025 · similarity 0.88)
- Contradicted in Reliable, Replicated in Unreliable: Dual-Source Reference for Fake News Early Detection (AAAI 2025 · similarity 0.88)
- Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection (AAAI 2025 · similarity 0.87)
- Reinforced Adaptive Knowledge Learning for Multimodal Fake News Detection (AAAI 2024 · similarity 0.87)
- Is LLMs Hallucination Usable? LLM-based Negative Reasoning for Fake News Detection (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.