FactGuard: Event-Centric and Commonsense-Guided Fake News Detection

AI summary

Overview

• Research area: Automated fake news detection using large language models (LLMs) combined with event-centric content extraction and commonsense reasoning.
• Technical level: Advanced (assumes familiarity with transformer encoders, cross-attention, knowledge distillation, and multi-objective loss design).
• Scope: The paper introduces FactGuard, an LLM-and-small-language-model framework for style-robust fake news detection, plus a distilled variant (FactGuard-D) for cold-start and resource-constrained deployment, evaluated on the Weibo21 (Chinese) and GossipCop (English) datasets.

What This Paper Is About

Most existing fake news detectors rely on writing style, and this becomes unreliable once adversaries imitate the style of authentic news. The authors propose FactGuard, which uses LLMs to strip stylistic noise by extracting the core event topic and content, then combines that with LLM-generated commonsense reasoning and a module that decides how much to trust the LLM's advice. A distilled version, FactGuard-D, is designed to work without calling an LLM at inference time.

Key Contributions

1. FactGuard framework: An event-centric detection framework that reduces the influence of textual style and integrates LLM-based reasoning through cross-attention and dynamic reliability weighting, targeted at resource-rich settings.
2. LLM-based news extraction: A prompting method that extracts key topics and principal event content from news articles, with a two-stage constraint mechanism (a text similarity metric during extraction and an information density metric after extraction) to maintain fidelity to the original article.
3. Rationale usability module: A dual-branch structure treating the LLM as an advisor, where one branch adaptively controls the influence of LLM judgments and the other emphasizes contradictions or ambiguity surfaced by commonsense reasoning.
4. FactGuard-D via knowledge distillation: A lightweight, LLM-free student model for cold-start and resource-constrained scenarios, initialized from the trained FactGuard and supervised by a feature distillation loss.

Main Findings

• FactGuard leads on both benchmarks: On Weibo21 it reaches macF1 0.801, Accuracy 0.804, F1_real 0.824, and F1_fake 0.777. On GossipCop it reaches macF1 0.805, Accuracy 0.892, F1_real 0.935, and F1_fake 0.675.
• Margin over the strongest baseline: On Weibo21, FactGuard outperforms the strongest baseline (TED) by 0.8% in accuracy and 0.9% in F1_real, with additional macF1 gains. TED scored macF1 0.795, Accuracy 0.798, F1_real 0.815, F1_fake 0.774 on Weibo21.
• Fewer prompts than multi-agent debate: The paper states FactGuard achieves higher accuracy than TED's multi-agent debate framework using only two simple prompts, which it attributes to lower computational and inference cost.
• Distilled variant stays competitive: FactGuard-D achieves macF1 0.788, Accuracy 0.790, F1_real 0.807, F1_fake 0.769 on Weibo21, and macF1 0.790, Accuracy 0.888, F1_real 0.933, F1_fake 0.647 on GossipCop. It surpasses ARG (Weibo21 macF1 0.784) and ARG-D (Weibo21 macF1 0.771).
• Original news is the most important input: In the ablation study, removing the original news representation caused the largest drop (Weibo21 macF1 falls from 0.801 to 0.718; GossipCop from 0.805 to 0.773).
• Both LLM-derived signals matter and must be combined: Using LLM-extracted topic-content alone yields Weibo21 macF1 0.690, and commonsense rationale alone yields 0.678, both below the original-news-only setting (0.768). The paper states the two modules must be used in conjunction for maximum improvement.
• Usability module helps: Removing the LLM rationale usability module lowers Weibo21 macF1 to 0.778; substituting ARG's usefulness module gives 0.782.
• Loss weights are tuned per dataset: Optimal α and β were α=0.40, β=0.16 for Weibo21 and α=0.50, β=0.58 for GossipCop. The distillation coefficient λ was set to 8 for both the Chinese and English FactGuard-D models.
• Performance gap explained by class balance: The paper attributes the smaller GossipCop gains to class imbalance (more fake news in Weibo21, more real news in GossipCop).
• Efficiency measurements: The paper asserts efficiency advantages for FactGuard-D over LLM-invoking pipelines, but no specific latency, throughput, or parameter-count figures appear in the provided content.

Methodology in Plain English

The authors start from the observation that news describes real-world events, while stylistic rewriting is a common deception strategy. So instead of classifying raw text directly, they first ask an LLM to read each article and produce two things: a topic-and-content summary of the underlying event, and a commonsense rationale explaining whether anything in the article contradicts common sense. The extraction is checked two ways: a text similarity metric keeps the summary faithful to the original article, and an information density metric checks that the summary is actually informative.

The original news text, the extracted topic-content, and the commonsense rationale are each encoded with a small pretrained language model (BERT for Chinese, RoBERTa for English). A cross-attention module lets the topic-content and rationale features exchange information in both directions. One of these cross-attention outputs becomes a weight factor inside a "Rationale Usability Evaluator," which is a dual-branch MLP with a sigmoid output. The idea is that one branch is trained to shrink the LLM's influence (since the LLM's own detection ability is limited), while the other branch increases influence when commonsense reasoning flags contradictions or uncertainty. The news features and the weighted LLM features are concatenated and passed to an MLP classifier.

Training uses three losses: a binary cross-entropy classification loss, a usability loss that supervises the branch weights (one branch supervised toward 0, the other toward the LLM's own judgment label), and an auxiliary text loss that aligns the extracted content and rationale representations with ground-truth labels and LLM judgments. These are summed with weights α and β.

For FactGuard-D, the student's news encoder and classifier are initialized from the trained FactGuard, and a four-layer Transformer encoder plus linear attention acts as a feature simulator to internalize the teacher's reasoning. The student is trained with the classification loss plus a distillation loss that minimizes the mean squared error between its feature representation and the teacher's. At inference, FactGuard-D uses only the original news text, so no LLM calls are needed. Experiments used a single NVIDIA A100 (40GB) GPU, AdamW with a learning rate of 2e-4 and weight decay of 5e-5, early stopping with patience of 5 epochs, a fixed seed of 3759, and PyTorch 1.13.0; topic extraction used locally deployed DeepSeek-R1-Distill-Llama-8B for Weibo21 and SOLAR-10.7B-Instruct-v1.0-uncensored for GossipCop, with commonsense rationales adopted from prior work (Hu et al. 2024).

Why This Matters

• Shifts detection away from style: By modeling events rather than prose, the approach directly targets a known failure mode of stylometric detectors that adversaries can evade by mimicking authentic writing.
• Turns LLM advice into a measured input: Rather than trusting or ignoring LLM judgments wholesale, the framework learns when to weight them, addressing documented problems with hallucination and low few-shot/chain-of-thought accuracy in LLM detectors.
• Addresses deployment cost: The distilled variant provides an LLM-free path for settings where LLMs cannot be called, which the paper frames as the resource-constrained and cold-start case.

Real-world applications implied by the work:

• Early-stage content moderation on social platforms such as Sina-Weibo and Facebook, where only the article text is available at publication time.
• Multilingual or cross-lingual misinformation screening (the paper evaluates both Chinese and English).
• Rapid triage during major public events, where the paper notes misinformation can trigger social panic and disrupt governance.
• Edge or on-device deployment where LLM inference is unavailable, via FactGuard-D.

Industry relevance: The framework's design target is the practical trade-off that platform operators face: LLM-quality reasoning is too expensive to run per item at scale, so distilling that reasoning into a lightweight model is directly relevant to production moderation pipelines. The code is released at https://github.com/ryliu68/FACTGUARD.

Future Directions

1. Develop customized methods tailored separately to Chinese and English fake news, rather than one cross-lingual design.
2. Optimize the model for edge deployment.
3. Improve the interpretability of the usability evaluation module to increase transparency and credibility.
4. Study the role of text style at different stages of news dissemination detection.
5. Address benchmark data contamination in the LLMs used, and extend cross-domain adaptation to emerging platforms and multimodal signals.

Target Audience

Researchers and practitioners working on misinformation detection, LLM-augmented NLP pipelines, and knowledge distillation for efficient inference. The paper is most useful to readers already comfortable with transformer architectures, cross-attention fusion, and multi-objective training, and to engineers evaluating whether an LLM-assisted detector can be compressed into a deployable model.

Authors’ abstract

Fake news detection methods based on writing style have achieved remarkable progress. However, as adversaries increasingly imitate the style of authentic news, the effectiveness of such approaches is gradually diminishing. Recent research has explored incorporating large language models (LLMs) to enhance fake news detection. Yet, despite their transformative potential, LLMs remain an untapped goldmine for fake news detection, with their real-world adoption hampered by shallow functionality exploration, ambiguous usability, and prohibitive inference costs. In this paper, we propose a novel fake news detection framework, dubbed FactGuard, that leverages LLMs to extract event-centric content, thereby reducing the impact of writing style on detection performance. Furthermore, our approach introduces a dynamic usability mechanism that identifies contradictions and ambiguous cases in factual reasoning, adaptively incorporating LLM advice to improve decision reliability. To ensure efficiency and practical deployment, we employ knowledge distillation to derive FactGuard-D, enabling the framework to operate effectively in cold-start and resource-constrained scenarios. Comprehensive experiments on two benchmark datasets demonstrate that our approach consistently outperforms existing methods in both robustness and accuracy, effectively addressing the challenges of style sensitivity and LLM usability in fake news detection.