Organizations: Key Laboratory of Dependable Service Computing in Cyber Physical Society (Chongqing University), Ministry of Education, Chongqing 401331, China · School of Big Data and Software Engineering, Chongqing University, Chongqing 400044, China · Department of Plastic Surgery, Chongqing University Central Hospital, Chongqing Emergency Medical Center, Chongqing 400000, China · Informatics Research Centre, University of Reading, Reading, United Kingdom
Fake news detection under low resource conditions is challenged by the scarcity of labeled data and the difficulty of capturing subtle deceptive patterns. In this work, we focus on low resource fake news detection, where only a small labeled set is available for training, while evaluation is performed under a fixed and unchanged test protocol. This setting is motivated by costly, ambiguous fact-checking and emerging topics and domains that are poorly covered by existing benchmarks. Existing data augmentation methods, such as synonym substitution, generative adversarial networks, and large language model (LLM)-based generation, often produce fluent but overly generic content, lacking the task-specific relevance needed for accurate detection. To address this, we propose SharP, a boundary-aware text generation framework guided by soft and hard prompts, which leverages LLMs as tools adapted to produce task-specific synthetic samples. SharP employs a dual prompt mechanism: hard prompts provide structured guidance and factual constraints, while soft prompts introduce learned semantic patterns from the data. These two types of prompts are combined to guide the generation process, enabling the model to produce samples that closely align with the characteristics of news. To further enhance detection performance, we incorporate a boundary-aware strategy that steers generation toward areas where the classifier is less confident, helping to clarify subtle distinctions between real and fake content. In addition, we adopt a dual objective optimization that balances semantic alignment with the source data and classifier feedback. This encourages the generation of samples that are both domain-consistent and helpful for refining decision boundaries, ultimately improving model robustness and generalization in low resource fake news detection.
Figures & tables
Dataset
Source
Lang.
Scale
Labels / Structure
Social rumor verification
PHEME ( Zubiaga et al., 2016 )
Twitter events
EN
330/4.8k
Conversation threads; rumor verification / rumor vs. non-rumor.
Twitter15/16 ( Ma et al., 2017a ; Ma et al., 2017b )
Twitter events
EN
1490/818
Event-level rumor detection with reply structure.
Weibo (rumor) ( Ma et al., 2016 )
Sina Weibo
ZH
2,560 (1280 fake + 1280 true)
Post-level rumor vs. non-rumor (common benchmark variants).
News/claim fake news detection
LIAR ( Wang, 2017 )
PolitiFact claims
EN
12.8k
6-way truthfulness; often binarized in detection.
Table 1: Representative benchmark datasets for fake news and rumor detection
Work
LLM role
Datasets / key results (as reported in the original papers)
Chen et al. (2025)
Prompted LLM as detector
On FND23, GPT-3.5-turbo with a task-specific prompt reports Acc=0.875 and Macro- F1 =0.528; the paper additionally benchmarks multiple datasets/prompting variants.
Raza et al. (2025)
LLMs vs. BERT-like (zero-/few-shot, fine-tune)
The authors report F1 : on their GPT-labeled dataset vs. NELA-GT, RoBERTa achieves 0.897 (vs. 0.836), while Mistral-7B-v0.2 fine-tuned achieves 0.802 (vs. 0.737).
Choi et al. (2025)
LLM-based data augmentation
Finds LLM-generated augmentation yields limited or inconsistent gains for COVID-19 misinformation stance detection across settings, cautioning that fluency does not guarantee task-useful diversity.
PHEME (5 events, 1%–70% labeled): Acc improves from 0.760 (1%) to 0.938 (70%); consistently outperforms strong baselines under low resource splits (see Table 6 ).
Table 2: Representative LLM-related studies for fake news detection and their reported results
Figure 1: Model Architecture of SharP
Symbol
Description
DLt=(XL,YL)
Labeled pool at round t , where XL and YL denote labeled texts and their labels.
DUt=XU
Unlabeled pool at round t , where XU denotes unlabeled texts.
St
Queried unlabeled batch selected from DUt at round t .
St′
Augmented labeled set generated from St (inherits labels).
b
Query batch size in each active learning round.
T
Total number of active learning rounds.
Table 3: Notation used in the Method section
Figure 2: Method Flow
Figure 3: Architecture of Data Augmentation Module
Figure 4: Vocabulary Mapping
LD=LossD(D(xL),yL)
Algorithm 1 Model Pre-training Process
St={(xu(i),yu(i))}i=1b.
Algorithm 2 Model Fine-tuning Process
Dataset
Cha.
Fer.
Ger.
Ott.
Syd.
Total
Labeled Fake News
318
144
95
330
382
1269
Labeled Real News
1481
719
94
280
559
3133
Unlabeled Fake News
140
140
143
140
140
703
Unlabeled Real News
140
140
137
140
140
697
Table 4: Dataset statistics for Group A (PHEME) used in our labeled/unlabeled pool setting
Dataset
LIAR
Twitter15/16
Labeled Fake News
5,514
445
Labeled Real News
6,997
429
Unlabeled Fake News
143
134
Unlabeled Real News
137
146
Total
12,791
1,154
Table 5: Dataset statistics for Group B (LIAR and Twitter15/16) used in our labeled/unlabeled pool setting
Dataset
Stra.
1%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
70%
Germanwings-crash
HAN
0.500
0.734
0.713
0.691
0.691
0.745
0.734
0.755
0.723
0.766
0.745
0.713
dEFEND
0.670
0.745
0.723
0.702
0.745
0.755
0.745
0.787
0.681
0.745
0.755
0.713
HMCAN
0.250
0.272
0.652
0.757
0.766
0.712
0.610
0.727
0.651
0.750
0.768
0.813
DDT
0.740
0.729
0.760
0.781
0.792
0.823
0.833
0.854
0.865
0.854
0.885
0.927
Deepseek-7b
0.617
0.617
0.617
0.617
0.628
0.628
0.628
0.617
0.606
0.617
0.617
0.606
SharP
0.760
0.802
0.792
0.833
0.833
0.813
0.854
0.885
0.875
0.875
0.917
0.938
Table 6: Performance comparison on Group A datasets (Acc.)
Figure 5: Sample Space Visualization
Dataset
Stra.
1%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
70%
PHEME (Germanwings-crash)
HAN
0.500
0.734
0.713
0.691
0.691
0.745
0.734
0.755
0.723
0.766
0.745
0.713
dEFEND
0.670
0.745
0.723
0.702
0.745
0.755
0.745
0.787
0.681
0.745
0.755
0.713
HMCAN
0.250
0.272
0.652
0.757
0.766
0.712
0.610
0.727
0.651
0.750
0.768
0.813
DDT
0.740
0.729
0.760
0.781
0.792
0.823
0.833
0.854
0.865
0.854
0.885
0.927
Deepseek-7b
0.617
0.617
0.617
0.617
0.628
0.628
0.628
0.617
0.606
0.617
0.617
0.606
SharP
0.760
0.802
0.792
0.833
0.833
0.813
0.854
0.885
0.875
0.875
0.917
0.938
Table 7: Performance comparison on Group B datasets (Acc.)
Figure 6: Ablation Study
Figure 7: Hyperparameter Sensitivity Analysis
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Item
Content
Template Th(x,st)
Rewrite this ONE sentence using <s_t> word choices (minimal substitutions). Keep meaning/facts; keep names/numbers unchanged. Return ONE sentence only, no extra text. <x>
Style tokens
st∈{common,professional,intense,neutral}
Example instantiation (common style)
Prompt ( st=common )
Rewrite this ONE sentence using common word choices (minimal substitutions). Keep meaning/facts; keep names/numbers unchanged. Return ONE sentence only, no extra text. The government announced a $5 million relief plan on Tuesday.
Appendix
Table 8: Hard prompt template used for style-controlled rewriting, and an example instantiation with the common style (the best-performing style token in our experiments)
The growing spread of misinformation in digital media highlights the need for reliable fake news detection systems, yet progress in under-resourced languages such as Bangla is limited by small and imbalanced datasets. This study investigates whether Large Language Model (LLM) based augmentation can effectively address this limitation and improve Bangla fake news classification. Existing datasets remain valuable but highly imbalanced, limiting model performance, and LLM based augmentation for Bangla has been scarcely explored. To fill this gap, we propose a systematic augmentation framework that generates synthetic Bangla news articles using the instruction tuned Gemma 3 27B IT model, supported by semantic filtering and controlled subsampling to preserve label consistency and diversity. We compare zero shot and few shot prompting, evaluate multiple augmentation rates, and examine random versus similarity-based selection strategies. Our experiments show that augmenting only the minority class with a high augmentation rate and random subsampling yields the strongest gains, raising the Fake News F1 score from 0.85 to 0.88. To support reproducibility and further research in this low-resource domain, we publicly release 4,545 synthetically generated Bangla fake news samples along with our full implementation. These findings demonstrate that well-designed LLM-driven augmentation can significantly improve fake news detection in low resource settings and provide a practical foundation for advancing multilingual misinformation research.
Ahmed Alfey Sani, Kazi Akib Zaoad, Shefayat E Shams Adib +2
Islamic University of Technology Gazipur, Bangladesh
Recent advances in large language models (LLMs) have enabled the large-scale generation of highly fluent and deceptive news-like content. While prior work has often treated fake news detection as a binary classification problem, modern fake news increasingly arises through human-AI collaboration, where strategic inaccuracies are embedded within otherwise accurate and credible narratives. These mixed-truth cases represent a realistic and consequential threat, yet they remain underrepresented in existing benchmarks. To address this gap, we introduce MANYFAKE, a synthetic benchmark containing 6,798 fake news articles generated through multiple strategy-driven prompting pipelines that capture many ways fake news can be constructed and refined. Using this benchmark, we evaluate a range of state-of-the-art fake news detectors. Our results show that even advanced reasoning-enabled models approach saturation on fully fabricated stories, but remain brittle when falsehoods are subtle, optimized, and interwoven with accurate information.
Xinyu Wang, Sai Koneru, Wenbo Zhang +3
Pennsylvania State University University Park, PA 16802, USA
To efficiently combat the spread of LLM-generated misinformation in the news domain, we present RADAR, a Retrieval-Augmented Detector with Adversarial Refinement for adaptive LLM-generated fake news detection. Our approach employs a generator that rewrites real articles with factual perturbations, paired with a lightweight detector that verifies claims using dense passage retrieval. To enable effective co-evolution, we introduce Verbal Adversarial Feedback (VAF). Rather than relying on scalar rewards, VAF issues structured natural-language critiques; these guide the generator toward more sophisticated evasion attempts, compelling the detector to adapt and improve. Experiments on an LLM-generated fake news benchmark show that RADAR outperforms retrieval-augmented trainable baselines and general-purpose LLMs with retrieval. Further analysis shows that retrieval on both the generator and detector sides improves performance, while VAF and few-shot demonstrations offer complementary benefits. RADAR also transfers better to fake news generated by an unseen external attacker, suggesting improved generalization beyond the specific co-evolved generator used during training.