Lifelong Evolution: Collaborative Learning between Large and Small Language Models for Continuous Emergent Fake News Detection

Chaozhuo Li; Litian Zhang; Philip S. Yu; Xiaoming Zhang; Yibo Zhang; Zhenyu Guan; Ziyi Zhou

REVIEW 3 major objections 4 minor 52 references

Lifelong Evolution: Collaborative Learning between Large and Small Language Models for Continuous Emergent Fake News Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims a two-model collaboration detects emergent fake news accurately with 10-15% labeled samples.

desk verdict A well-built continual-learning framework whose experimental protocol trains on the test split, so the headline SOTA claim is not supported as stated. read the letter →

arxiv 2506.04739 v1 pith:G55IVXZ5 submitted 2025-06-05 cs.CL cs.AI

Ziyi Zhou , Xiaoming Zhang , Litian Zhang , Yibo Zhang , Zhenyu Guan , Chaozhuo Li , Philip S. Yu This is my paper · ORCID

classification cs.CLcs.AI

keywords fakenewsdetectioncontinuallearninglargelanguagemodelssmallactivesemi-supervisedknowledgeeditingmixtureofexperts

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Fake news on social media arrives in a stream of emergent events, each with few labels and a shifted distribution, and models trained once on past events decay in accuracy. The paper proposes C2EFND, a framework in which a large language model (LLM) and a small language model (SLM) collaborate over multiple rounds to detect continuous emergent fake news. It claims that the LLM's generalization and the SLM's classification expertise can be combined so that only 10-15 percent of samples need manual annotation, while both models keep updating as events arrive and retain the ability to detect past events. On the Pheme and Twitter16 datasets, C2EFND reports accuracy and F1 improvements over single-event, LLM-assisted, and multi-domain baselines. This matters because the condition it targets, new misinformation arriving with scarce labels and without forgetting past events, is the one real-world detection systems face.

What carries the argument

The load-bearing mechanism is the multi-round collaboration between the LLM and the SLM, formalized in Algorithm 1 and Equations 13-14. In each round, the LLM and SLM each classify the unlabeled event data; samples where the two models agree and the SLM's confidence exceeds a threshold $\omega$ enter a clean pool used to fine-tune the SLM, while conflicting or low-confidence samples enter a noisy pool from which a second active-learning stage selects a few for human labeling. The LLM is updated through an MoE-based lifelong knowledge editing module: each new event adds a dedicated feedforward expert $FFN_i$ that is trained while the original and earlier experts stay frozen, with a router network selecting the top-$k_1$ experts at inference. The SLM is kept from forgetting through a rehearsal replay loss over a bounded memory bank and a KL-divergence distillation loss from its previous version. These machinery pieces together convert scarce labeled samples and abundant unlabeled data into incremental updates of both models.

What would settle it

Run C2EFND with a strict held-out evaluation subset per event that is never touched by the two-stage active learning, pseudo-label filtering, or multi-round updates, and compare accuracy and F1 on that untouched subset against the reported numbers; if the gains over baselines shrink or disappear, the reported state-of-the-art result comes from training on the evaluation set rather than from generalization.

Watch

Extended reading notes

Core claim

The paper's central claim is that continuous emergent fake news detection can be solved by a multi-round collaborative learning loop rather than by retraining a classifier on each new event. In its design, a two-stage active learning module first selects diverse samples via distance to a memory bank, then selects uncertain samples where the LLM and SLM disagree or the SLM has low confidence; the LLM is updated with a per-event Mixture-of-Experts feedforward network, so new event knowledge is stored in a newly added expert while old experts stay frozen, and the SLM is updated with replay from a bounded memory bank plus knowledge distillation from its previous version. The two models then annotate the remaining unlabeled data through in-context learning and rationale generation, keeping samples on which they agree as clean pseudo-labels and sending disagreements to a second round of human annotation. The paper reports that this loop outperforms existing methods on target unseen events and on all accumulated events across Pheme and Twitter16, and that the edited LLM outperforms frozen-parameter prompting configurations, while ablations show each component contributes to the result.

Load-bearing premise

The framework assumes that the unlabeled test split of each event may be used for active learning, pseudo-label generation, and multi-round model updating, and that reporting accuracy on that same split is a valid measure of generalization to emergent news.

Editorial extensions

If this is right

On the reported experimental setup, C2EFND claims higher accuracy and F1 than single-event, LLM-assisted, and multi-domain baselines on most target events of Pheme and Twitter16, with gains holding on all accumulated events.
The framework claims to sustain detection on past events without full retraining, using a memory bank capped at 400 samples, so rumor resurgence can be handled at a fraction of the training cost of replay-based baselines.
With only 5-20 percent labeled data, C2EFND reports accuracy and F1 improvements over ARG and M3FEND, suggesting the two-stage active learning and multi-round pseudo-labeling compensate for annotation scarcity.
Knowledge-editing the LLM improves its zero-shot, few-shot, and in-context-learning veracity judgments, and the multi-round analysis shows SLM accuracy rises from about 80.2 percent to 88.1 percent by round 3 on Twitter16 E1, after which extra rounds add little or hurt.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editor's inference: because the unlabeled split doubles as the test set and the model updates on samples drawn from it, the reported numbers measure the framework's ability to fit the event's own test distribution rather than generalization to a truly unseen future event; a held-out event or a never-updated test split would be the decisive check.
Editor's inference: the collaboration mechanism is not specific to news text and could be transferred to other streaming classification tasks with drifting distributions, such as rumor detection in other languages or fraud alerts, as long as the LLM can generate rationales and the two models' agreement is a usable confidence signal.
Editor's inference: the per-event MoE expert design implies the number of trainable experts grows with the number of events; over very long horizons, expert routing and memory of early events would need empirical study, and the paper's five-event benchmarks may understate the challenge.
Editor's inference: the conflict-based uncertainty sampling in Equation 14 could be tested as a general annotation-efficiency device, since it uses model disagreement rather than a single model's confidence as the signal for human review.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The framework is a thoughtful engineering integration: two-stage active learning, MoE-based lifelong knowledge editing for the LLM, replay-plus-distillation for the SLM, and a multi-round collaborative loop. The MoE editing and the replay/distillation combination are genuinely new relative to your own AAAI 2025 paper, and the ablation study is reasonably thorough. It addresses a real practical problem, and the paper is well-written enough that a reader can follow exactly what was done.

That last point is what makes the central problem so clear. The paper defines the unlabeled samples of each event as the test set (Section III), then Algorithm 1 uses that test set for active-learning annotation (10% first stage, 5% second), pseudo-label selection, and multi-round fine-tuning. Section V-C3 confirms 80% of each event is reserved for testing. So the accuracy reported in Tables III, IV, VI, and VII is measured on data whose inputs and labels were directly used to train the model. Baselines are trained only on the 20% labeled split and evaluated on the untouched 80%. That asymmetry means the SOTA gains are likely an artifact of transductive label leakage, not evidence of generalization to emergent news. The knowledge-edit analysis in Table VIII has the same circularity: the LLM is edited with labeled samples from an event and then evaluated on that same event's test split.

This is not a minor flaw; it is load-bearing. The abstract claims significant improvement in continuous emergent fake news scenarios, but the experiments do not support that claim as an inductive result. That said, the paper could be fixed. The natural framing is a transductive or self-training evaluation, with baselines given the same access to unlabeled data. Or the authors could use a clean split: train on the labeled 20%, use a separate validation set for active learning and pseudo-label selection, and test on a truly held-out set. Either way, the experimental protocol needs to be rebuilt before the central claim can be assessed.

Who is this for? Researchers working on continual fake-news detection with LLM+SLM collaboration. The framework ideas may be worth borrowing, but I would not cite the empirical results until the evaluation is redone. A serious referee should see it, because the underlying system is plausible and the flaw is fixable, but the current form would be a clear reject. I recommend sending it to peer review with major revisions required.

Referee Report

3 major / 4 minor

Summary. The paper proposes C2EFND, a continual-learning framework for fake news detection on streaming events. The method combines two-stage active learning, a Mixture-of-Experts lifelong knowledge-editing module for an LLM, a replay- and distillation-based continual learning module for an SLM, and multi-round collaborative pseudo-labeling between the LLM and SLM. Experiments on Pheme and Twitter16 report state-of-the-art accuracy and F1 on target (new) events and on all previously seen events. The central claim is that C2EFND significantly outperforms existing methods in continuous emergent fake news detection while using only 10–15% annotated samples plus unlabeled data.

Significance. The problem formulation — continuous detection of emergent fake news with scarce labels and a dynamic knowledge-update requirement — is timely and practically motivated. The proposed architecture is novel in combining LLM knowledge editing, SLM continual learning, and collaborative pseudo-labeling, and the paper provides a detailed algorithm and ablation analysis. However, the empirical evidence for the headline claim is invalidated by a circular evaluation protocol: the data used for active learning, pseudo-labeling, and additional training is the same 80% split on which accuracy is subsequently reported, while baselines are trained on the remaining 20% and tested on an untouched held-out split. This asymmetry means the reported gains likely reflect training on the evaluation set rather than generalization to emergent news. Because this issue is load-bearing for all the main results, the paper's empirical contribution is not currently supported.

major comments (3)

[Section III, Section IV-D, Section V-C3, Algorithm 1] The evaluation protocol is circular and gives C2EFND an unfair advantage over all baselines. Section III explicitly defines the unlabeled samples of each event as 'the testset', and Section V-C3 states that 80% of each event is reserved for testing. Yet the method uses this test set throughout the learning loop: Equation (14) forms Dclean and Dnoisy from test-set predictions, Equation (15) draws a second batch of manually annotated samples Dlabel from Dnoisy, and Algorithm 1 lines 21–25 fine-tune the SLM on Dclean and edit the LLM on Dlabel. The accuracy and F1 reported in Tables III, IV, VI, and VII are therefore measured on data whose inputs and labels were directly used to train the model. Baselines such as RoBERTa, ARG, and M3FEND are trained only on the 20% labeled split and evaluated on the untouched 80%, so the comparison is not apples-to-apples. This invalidates the central claim that C2EFND generalizes better to unseen emergent events.
[Section V-F, Table VIII] The evidence for the value of the lifelong knowledge-editing module is also confounded by test-set leakage. In Table VIII, the knowledge-edited LLM is edited on labeled samples associated with an event (which, per the protocol, are drawn from that event's test split) and then evaluated on 'the overall performance of all events', i.e., the same event test splits. The reported improvements of ICL+KE over ICL (e.g., from 0.672 to 0.706 accuracy on Twitter) could be due to the LLM memorizing the label information of test samples rather than to a generalizable ability to incorporate event knowledge. The paper should evaluate knowledge editing on a held-out portion of each event that is never used for editing, or on future events, to support the claim that editing improves reasoning about emergent news.
[Section V-D1, 'Results on Target Events'] The phrase 'target (unseen emergent) events' in Section V-D1 is misleading given the described protocol. Although each event is processed sequentially and the model is updated per event, the 80% test split of each event is not unseen: it provides the unlabeled data for active learning, pseudo-labeling, and multi-round training. Thus the results in Tables III and IV do not demonstrate generalization to unseen events; they demonstrate performance on a transductive or semi-supervised setting where the model is allowed to consume the test inputs and, through human annotation, some test labels. The paper should either adopt a genuinely held-out test split (e.g., a separate 20% partition that is never touched by any module) and rerun all experiments, or explicitly reframe the evaluation as a transductive/self-training comparison and include baselines that also use the unlabeled test data under the same protocol.

minor comments (4)

[Figure 3 caption] The caption reads 'The architecture of C 2FEND', which is inconsistent with the paper's notation 'C2EFND'; please unify the abbreviation throughout.
[Throughout] The paper uses 'continue learning' in several places (e.g., the Index Terms and Section II-B) where 'continual learning' is the standard and more accurate term; consider consistent terminology.
[Section V-C3] The description of the training/validation split says '20% of each event-specific dataset for training and validation', but it is unclear how the validation subset is used and whether the reported numbers are selected based on validation or test performance; please clarify the model selection procedure.
[Section V-F, Table VIII] The table reports accuracy and F1 for the LLM alone, but it does not specify the number of samples used for knowledge editing or the exact edit data source; this information is needed to assess the contribution of knowledge editing independently of the main framework.

Circularity Check

2 steps flagged · score 7.0 of 10

The headline SOTA result is obtained by training on the test split via pseudo-labeling and active annotation, then reporting accuracy on that same split; the claimed generalization to unseen emergent events is not measured inductively.

fitted input called prediction [Section III (Problem Formulation); Section IV-D2 Eqs. (14)-(15); Algorithm 1 lines 17-29; Section V-C3]
"The remaining unlabeled samples are defined as the testset: Xi = {xi} Nei i= ˆNei. ... all unlabeled news {(x, ˆy1, ˆy2), x∈ Xtest} are divided into clean data pool Dclean and noisy data pool Dnoisy ... Edit L with Dlabel and fine-tune S by Dclean ∪ M ... we allocate 20% of each event-specific dataset for training and validation, while the remaining 80% is reserved for testing."

The test set is defined as the unlabeled remainder of each event (Section III). Eq. (14) partitions that test set into Dclean and Dnoisy; Eq. (15) selects Dlabel from Dnoisy for human annotation; Algorithm 1 fine-tunes the SLM on Dclean and knowledge-edits the LLM on Dlabel. Reported accuracies in Tables III, IV, VI, and VII use the same events under the V-C3 80% test split. Because Dlabel contributes true labels from the test set and Dclean supplies inputs and pseudo-labels for training, the evaluation includes data the model was fitted on, while baselines are trained only on the 20% split. The 'target unseen events' claim is therefore not an inductive comparison.
fitted input called prediction [Section V-F (Table VIII), with Algorithm 1 line 21 and Eq. (15)]
"our knowledge-edited LLM integrates the labeled news samples and knowledge directly into its model weights before performing the classification task. As illustrated in Table VIII, the knowledge-edited LLM consistently outperforms its unedited counterparts in all three scenarios. ... Edit L with Dlabel and fine-tune S by Dclean ∪ M"

Table VIII is presented as evidence that lifelong knowledge editing improves the LLM. In the pipeline, the LLM is edited on Dlabel (Algorithm 1 line 21), and Dlabel is human-annotated from the test set via Eq. (15). The table reports accuracy on the overall performance of all events, so the edited LLM is scored on event data whose true labels were written into its weights during editing. The gain over the frozen LLM is consequently at least partly a memorization effect rather than evidence of generalizing to emergent events.

full rationale

The central circularity is evaluative: the manuscript's own split (V-C3), problem formulation (Section III), data-selection equations (14)-(15), and Algorithm 1 jointly render the 'test' split training material. The self-citation [28] (Zhou et al., AAAI 2025) appears only in a related-work claim that LLMs struggle in zero-shot veracity settings and is not load-bearing. No uniqueness theorem, ansatz-smuggling citation, or renaming-of-known-result pattern is present. If an untouched held-out split (or an explicit exclusion of Dlabel/Dclean from evaluation) were used, the architecture comparisons would be meaningful; as written, the headline SOTA claim is not independent of its evaluation data.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central evaluation depends on several tunable hyperparameters (P1, P2, omega, N, k1, k2, Mmax, lambda1, lambda2) and on domain assumptions about Wikipedia knowledge quality, pseudo-label reliability, and expert interference; none are independently validated.

free parameters (9)

P1 (first-stage active learning ratio) = 0.10
Proportion of unlabeled event data selected for annotation in the first stage; set in Section V-C3.
P2 (second-stage active learning ratio) = 0.05
Proportion of noisy pool selected for manual annotation in the second stage; set in Section V-C3.
omega (confidence threshold) = 0.85
SLM confidence threshold for accepting pseudo-labels; used in Equation 14.
N (number of rounds) = 3
Number of collaborative multi-round iterations; set in Section V-C3.
k1 (number of selected experts) = 2
Top-k experts selected by the MoE router in Equation 5.
k2 (number of demonstrations) = 8
Number of BM25-retrieved demonstrations for in-context learning in Equation 12.
Mmax (memory bank capacity) = 400
Maximum number of samples retained in the memory bank; used in Equation 16.
lambda1 (replay loss weight) = 1.0
Weight of the replay loss in Equation 11.
lambda2 (distillation loss weight) = 0.4
Weight of the distillation loss in Equation 11.

assumptions (4)

domain assumption The Wikipedia API returns accurate and timely entity knowledge that improves fake-news detection.
Used in Section IV-A for knowledge extraction and throughout for conditioning both LLM and SLM; no verification of Wikipedia coverage for the events is provided.
domain assumption Agreement between the LLM and SLM predictions, combined with SLM confidence omega, identifies clean labels from unlabeled data.
Section IV-D2, Equation 14; the correctness of the pseudo-labels is assumed rather than measured.
ad hoc to paper The frozen original FFN plus added per-event experts retains past knowledge without harming new-event performance.
Section IV-B, Equation 5; no analysis of expert interference or router collapse is provided.
domain assumption RoBERTa embeddings are a valid distance metric for diverse-sample selection.
Section IV-A, Equation 1; the euclidean distance in embedding space is assumed to reflect diversity of news content.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Lifelong Evolution: Collaborative Learning between Large and Small Language Models for Continuous Emergent Fake News Detection." pith.science (2026). https://pith.science/paper/G55IVXZ5

@misc{pith2026250604739,
  author       = {Pith},
  title        = {Pith review of: Lifelong Evolution: Collaborative Learning between Large and Small Language Models for Continuous Emergent Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G55IVXZ5}},
  note         = {Machine review of arXiv:2506.04739}
}

abstract

The widespread dissemination of fake news on social media has significantly impacted society, resulting in serious consequences. Conventional deep learning methodologies employing small language models (SLMs) suffer from extensive supervised training requirements and difficulties adapting to evolving news environments due to data scarcity and distribution shifts. Large language models (LLMs), despite robust zero-shot capabilities, fall short in accurately detecting fake news owing to outdated knowledge and the absence of suitable demonstrations. In this paper, we propose a novel Continuous Collaborative Emergent Fake News Detection (C$^2$EFND) framework to address these challenges. The C$^2$EFND framework strategically leverages both LLMs' generalization power and SLMs' classification expertise via a multi-round collaborative learning framework. We further introduce a lifelong knowledge editing module based on a Mixture-of-Experts architecture to incrementally update LLMs and a replay-based continue learning method to ensure SLMs retain prior knowledge without retraining entirely. Extensive experiments on Pheme and Twitter16 datasets demonstrate that C$^2$EFND significantly outperforms existed methods, effectively improving detection accuracy and adaptability in continuous emergent fake news scenarios.

Figures

Figures reproduced from arXiv: 2506.04739 by the authors.

**Figure 2.** Challenges in continuous emergent fake news detection: distribution shift, sparse labels, and static models. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The architecture of C2FEND. C2FEND is composed of four main modules: a data initialization module, a lifelong knowledge editing module for the LLM, a continual learning module for the SLM, and a multi-round collaborative learning module between the LLM and SLM. layer can lead the model to forget previous edits, focusing solely on current events and resulting in catastrophe forgetting [33]. Inspired by the MoE archi… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Prompt for the LLM to generate rationales. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Hyper-parameter sensitivity analysis on Twitter16 and Pheme. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

[1]

Fake news on social media: the impact on society,

F. Olan, U. Jayawickrama, et al. , “Fake news on social media: the impact on society,” Information Systems Fron- tiers, vol. 26, no. 2, pp. 443–458, 2024

work page 2024
[2]

Inoculating against fake news about covid-19,

S. van Der Linden, J. Roozenbeek, et al. , “Inoculating against fake news about covid-19,” Frontiers in psychol- ogy, vol. 11, p. 566790, 2020

work page 2020
[3]

The diffusion of misinformation on social media: Temporal pattern, message, and source,

J. Shin, L. Jian, et al., “The diffusion of misinformation on social media: Temporal pattern, message, and source,” Computers in Human Behavior , vol. 83, pp. 278–287, 2018

work page 2018
[4]

Bert: Pre-training of deep bidirectional trans- formers for language understanding,

J. Devlin, “Bert: Pre-training of deep bidirectional trans- formers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

arXiv 2018
[5]

New explainability method for bert-based model in fake news detection,

M. Szczepa ´nski, M. Pawlicki, et al., “New explainability method for bert-based model in fake news detection,” Scientific reports, vol. 11, no. 1, p. 23705, 2021

work page 2021
[6]

Reinforced adaptive knowl- edge learning for multimodal fake news detection,

L. Zhang, X. Zhang, et al., “Reinforced adaptive knowl- edge learning for multimodal fake news detection,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, vol. 38, pp. 16777–16785, 2024

work page 2024
[7]

Dpsg: Dynamic propagation social graphs for multi-modal fake news detection,

C. Jing, H. Gao, et al. , “Dpsg: Dynamic propagation social graphs for multi-modal fake news detection,” In- formation Fusion, vol. 113, p. 102595, 2025

work page 2025
[8]

Embracing domain differences in fake news: Cross-domain fake news detection using multi-modal data,

A. Silva, L. Luo, et al., “Embracing domain differences in fake news: Cross-domain fake news detection using multi-modal data,” in Proceedings of the AAAI confer- ence on artificial intelligence, vol. 35, pp. 557–565, 2021

work page 2021

Show all 52 references

[9]

A survey on evaluation of large language models,

Y . Chang, X. Wang, et al. , “A survey on evaluation of large language models,” ACM Transactions on Intelligent Systems and Technology, vol. 15, no. 3, pp. 1–45, 2024

2024
[10]

A survey of large language models,

W. X. Zhao, K. Zhou, et al., “A survey of large language models,” arXiv preprint arXiv:2303.18223 , 2023

2023 arXiv
[11]

Bad actor, good advisor: Exploring the role of large language models in fake news detection,

B. Hu, Q. Sheng, et al. , “Bad actor, good advisor: Exploring the role of large language models in fake news detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 22105–22113, 2024

2024
[12]

Dell: Generating reactions and explanations for llm-based misinformation detection,

H. Wan, S. Feng, et al., “Dell: Generating reactions and explanations for llm-based misinformation detection,” arXiv preprint arXiv:2402.10426 , 2024

2024 arXiv
[13]

Fake news in sheep’s clothing: Robust fake news detection against llm-empowered style attacks,

J. Wu, J. Guo, et al. , “Fake news in sheep’s clothing: Robust fake news detection against llm-empowered style attacks,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pp. 3367–3378, 2024

2024
[14]

Mdfend: Multi-domain fake news detection,

Q. Nan, J. Cao, et al., “Mdfend: Multi-domain fake news detection,” in Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pp. 3343–3347, 2021

2021
[15]

Memory-guided multi-view multi-domain fake news detection,

Y . Zhu, Q. Sheng, et al. , “Memory-guided multi-view multi-domain fake news detection,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 7, pp. 7178–7191, 2022

2022
[16]

A survey of fake news: Fun- damental theories, detection methods, and opportunities,

X. Zhou and R. Zafarani, “A survey of fake news: Fun- damental theories, detection methods, and opportunities,” ACM Computing Surveys (CSUR) , vol. 53, no. 5, pp. 1– 40, 2020

2020
[17]

Mvae: Multimodal varia- tional autoencoder for fake news detection,

D. Khattar, J. S. Goud, et al., “Mvae: Multimodal varia- tional autoencoder for fake news detection,” in The world wide web conference , pp. 2915–2921, 2019

2019
[18]

Content-based fake news detection with machine and deep learning: a systematic review,

N. Capuano, G. Fenza, et al., “Content-based fake news detection with machine and deep learning: a systematic review,” Neurocomputing, vol. 530, pp. 91–103, 2023

2023
[19]

Mmdfnd: Multi-modal multi- domain fake news detection,

Y . Tong, W. Lu, et al. , “Mmdfnd: Multi-modal multi- domain fake news detection,” in ACM Multimedia 2024 , 2024

2024
[20]

Robust domain misinforma- tion detection via multi-modal feature alignment,

H. Liu, W. Wang, et al. , “Robust domain misinforma- tion detection via multi-modal feature alignment,” IEEE Transactions on Information Forensics and Security , vol. 19, pp. 793–806, 2023

2023
[21]

A comprehensive survey of continual learning: theory, method and application,

L. Wang, X. Zhang, et al. , “A comprehensive survey of continual learning: theory, method and application,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

2024
[22]

Incremental task learning with incremental rank updates,

R. Hyder, K. Shao, et al., “Incremental task learning with incremental rank updates,” in European Conference on Computer Vision, pp. 566–582, Springer, 2022

2022
[23]

A comprehensive study of class incremental learning algorithms for visual tasks,

E. Belouadah, A. Popescu, et al. , “A comprehensive study of class incremental learning algorithms for visual tasks,” Neural Networks, vol. 135, pp. 38–54, 2021

2021
[24]

Three types of incremental learning,

G. M. Van de Ven, T. Tuytelaars, et al. , “Three types of incremental learning,” Nature Machine Intelligence , vol. 4, no. 12, pp. 1185–1197, 2022

2022
[25]

Learning without forgetting,

Z. Li and D. Hoiem, “Learning without forgetting,” IEEE transactions on pattern analysis and machine intelli- gence, vol. 40, no. 12, pp. 2935–2947, 2017. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

2017
[26]

Efficient lifelong learning with a-gem,

A. Chaudhry, M. Ranzato, et al. , “Efficient lifelong learning with a-gem,” in International Conference on Learning Representations, 2018

2018
[27]

A unified approach to domain in- cremental learning with memory: Theory and algorithm,

H. Shi and H. Wang, “A unified approach to domain in- cremental learning with memory: Theory and algorithm,” Advances in Neural Information Processing Systems , vol. 36, 2024

2024
[28]

Collaborative evolution: Multi-round learning between large and small language models for emergent fake news detection,

Z. Zhou, X. Zhang, et al. , “Collaborative evolution: Multi-round learning between large and small language models for emergent fake news detection,” in Proceed- ings of the AAAI Conference on Artificial Intelligence , vol. 39, pp. 1210–1218, 2025

2025
[29]

Unveiling the generalization power of fine-tuned large language models,

H. Yang, Y . Zhang, et al., “Unveiling the generalization power of fine-tuned large language models,” in Proceed- ings of the 2024 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), ...

2024
[30]

Editing factual knowledge in language models,

N. De Cao, W. Aziz, et al., “Editing factual knowledge in language models,” in EMNLP 2021-2021 Conference on Empirical Methods in Natural Language Processing, Proceedings, pp. 6491–6506, 2021

2021
[31]

Rethinking the role of demon- strations: What makes in-context learning work?,

S. Min, X. Lyu, et al. , “Rethinking the role of demon- strations: What makes in-context learning work?,” in Proceedings of the 2022 Conference on Empirical Meth- ods in Natural Language Processing , pp. 11048–11064, 2022

2022
[32]

Knowledge injection to counter large language model (llm) hallucination,

A. Martino, M. Iannelli, et al. , “Knowledge injection to counter large language model (llm) hallucination,” in European Semantic Web Conference , pp. 182–185, Springer, 2023

2023
[33]

Aging with grace: Lifelong model editing with discrete key-value adaptors,

T. Hartvigsen, S. Sankaranarayanan, et al., “Aging with grace: Lifelong model editing with discrete key-value adaptors,” Advances in Neural Information Processing Systems, vol. 36, 2024

2024
[34]

A survey on in-context learning,

Q. Dong, L. Li, et al., “A survey on in-context learning,” arXiv preprint arXiv:2301.00234 , 2022

2022 arXiv
[35]

A survey on deep active learning: Recent advances and new frontiers,

D. Li, Z. Wang, et al., “A survey on deep active learning: Recent advances and new frontiers,” IEEE Transactions on Neural Networks and Learning Systems , 2024

2024
[36]

Context-aware query selection for active learning in event recognition,

M. Hasan, S. Paul, et al., “Context-aware query selection for active learning in event recognition,” IEEE trans- actions on pattern analysis and machine intelligence , vol. 42, no. 3, pp. 554–567, 2018

2018
[37]

Roberta: A robustly optimized bert pretraining approach,

Y . Liu, M. Ott,et al., “Roberta: A robustly optimized bert pretraining approach,” arXiv preprint arXiv:1907.11692 , 2019

1907 arXiv
[38]

Mixtral of experts,

A. Q. Jiang, A. Sablayrolles, et al., “Mixtral of experts,” arXiv preprint arXiv:2401.04088 , 2024

2024 arXiv
[39]

Loramoe: Alleviating world knowledge forgetting in large language models via moe- style plugin,

S. Dou, E. Zhou, et al. , “Loramoe: Alleviating world knowledge forgetting in large language models via moe- style plugin,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1932–1945, 2024

1932
[40]

Multi-modality cross attention network for image and sentence matching,

X. Wei, T. Zhang, et al., “Multi-modality cross attention network for image and sentence matching,” in Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10941–10950, 2020

2020
[41]

Dark experience for general continual learning: a strong, simple baseline,

P. Buzzega, M. Boschini, et al. , “Dark experience for general continual learning: a strong, simple baseline,” Advances in neural information processing systems , vol. 33, pp. 15920–15930, 2020

2020
[42]

A survey on semi- supervised learning,

J. E. Van Engelen and H. H. Hoos, “A survey on semi- supervised learning,” Machine learning, vol. 109, no. 2, pp. 373–440, 2020

2020
[43]

What makes good in-context examples for gpt-3?,

J. Liu, D. Shen, et al. , “What makes good in-context examples for gpt-3?,” in Proceedings of Deep Learn- ing Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pp. 100–114, 2022

2022
[44]

How to measure uncertainty in uncertainty sampling for active learning,

V .-L. Nguyen, M. H. Shaker, et al. , “How to measure uncertainty in uncertainty sampling for active learning,” Machine Learning, vol. 111, no. 1, pp. 89–122, 2022

2022
[45]

Glean: Generalized category discovery with diverse and quality-enhanced llm feed- back,

H. P. Zou, S. Singh, et al., “Glean: Generalized category discovery with diverse and quality-enhanced llm feed- back,” arXiv preprint arXiv:2502.18414 , 2025

2025 arXiv
[46]

Boididou, christina and papadopoulos, symeon and others,

Detection and visualization of misleading content on Twitter, “Boididou, christina and papadopoulos, symeon and others,” International Journal of Multimedia Infor- mation Retrieval, vol. 7, no. 1, pp. 71–86, 2018

2018
[47]

Exploiting context for rumour detection in social media,

A. Zubiaga, M. Liakata, et al. , “Exploiting context for rumour detection in social media,” in Social Informatics: 9th International Conference, SocInfo 2017, Oxford, UK, September 13-15, 2017, Proceedings, Part I 9 , pp. 109– 123, Springer, 2017

2017
[48]

Compare to the knowledge: Graph neural fake news detection with external knowl- edge,

L. Hu, T. Yang, et al. , “Compare to the knowledge: Graph neural fake news detection with external knowl- edge,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processin...

2021
[49]

Learn over past, evolve for future: Forecasting temporal trends for fake news de- tection,

B. Hu, Q. Sheng, et al. , “Learn over past, evolve for future: Forecasting temporal trends for fake news de- tection,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track), pp. 116–125, 2023

2023
[50]

Explainable fake news detection with large language model via defense among competing wisdom,

B. Wang, J. Ma, et al., “Explainable fake news detection with large language model via defense among competing wisdom,” in Proceedings of the ACM Web Conference 2024, pp. 2452–2463, 2024

2024
[51]

Eann: Event adversarial neural networks for multi-modal fake news detection,

Y . Wang, F. Ma, et al., “Eann: Event adversarial neural networks for multi-modal fake news detection,” in Pro- ceedings of the 24th acm sigkdd international conference on knowledge discovery & data mining , pp. 849–857, 2018

2018
[52]

Contrastive domain adaptation for early misinformation detection: A case study on covid-19,

Z. Yue, H. Zeng, et al., “Contrastive domain adaptation for early misinformation detection: A case study on covid-19,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pp. 2423–2433, 2022

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fake news on social media: the impact on society,

F. Olan, U. Jayawickrama, et al. , “Fake news on social media: the impact on society,” Information Systems Fron- tiers, vol. 26, no. 2, pp. 443–458, 2024

work page 2024

[2] [2]

Inoculating against fake news about covid-19,

S. van Der Linden, J. Roozenbeek, et al. , “Inoculating against fake news about covid-19,” Frontiers in psychol- ogy, vol. 11, p. 566790, 2020

work page 2020

[3] [3]

The diffusion of misinformation on social media: Temporal pattern, message, and source,

J. Shin, L. Jian, et al., “The diffusion of misinformation on social media: Temporal pattern, message, and source,” Computers in Human Behavior , vol. 83, pp. 278–287, 2018

work page 2018

[4] [4]

Bert: Pre-training of deep bidirectional trans- formers for language understanding,

J. Devlin, “Bert: Pre-training of deep bidirectional trans- formers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

arXiv 2018

[5] [5]

New explainability method for bert-based model in fake news detection,

M. Szczepa ´nski, M. Pawlicki, et al., “New explainability method for bert-based model in fake news detection,” Scientific reports, vol. 11, no. 1, p. 23705, 2021

work page 2021

[6] [6]

Reinforced adaptive knowl- edge learning for multimodal fake news detection,

L. Zhang, X. Zhang, et al., “Reinforced adaptive knowl- edge learning for multimodal fake news detection,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, vol. 38, pp. 16777–16785, 2024

work page 2024

[7] [7]

Dpsg: Dynamic propagation social graphs for multi-modal fake news detection,

C. Jing, H. Gao, et al. , “Dpsg: Dynamic propagation social graphs for multi-modal fake news detection,” In- formation Fusion, vol. 113, p. 102595, 2025

work page 2025

[8] [8]

Embracing domain differences in fake news: Cross-domain fake news detection using multi-modal data,

A. Silva, L. Luo, et al., “Embracing domain differences in fake news: Cross-domain fake news detection using multi-modal data,” in Proceedings of the AAAI confer- ence on artificial intelligence, vol. 35, pp. 557–565, 2021

work page 2021