MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection

Lehan Zhang; Naidong Zhao; Shangxi Li; Shiqi Hu Yiheng Zhou; Yinlei Cheng

REVIEW 5 major objections 5 minor 25 references

MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MRAFnd claims that a retrieval-augmented, multi-agent debate pipeline can detect zero-shot fake news more accurately than specialized baselines on three benchmarks, with the largest gain on Weibo-21.

desk verdict Useful pipeline paper with a plausible mechanism, but the headline SOTA numbers rest on an under-specified retrieval stage and test-set hyperparameter tuning; needs revision before the empirical claims can be trusted. read the letter →

arxiv 2608.01430 v1 pith:PJQWMOWC submitted 2026-08-02 cs.AI

Lehan Zhang , Yinlei Cheng , Shiqi Hu Yiheng Zhou , Shangxi Li , Naidong Zhao This is my paper

classification cs.AI

keywords fakenewsdetectionmultimodalzero-shotretrieval-augmentedreasoningmulti-agentdebatelargelanguagemodelsevidenceretrievaldisinformation

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

MRAFnd is built on the idea that fake news about new events often recycles tactics from older stories, so a suspicious article is best judged alongside similar articles rather than in isolation. The paper proposes a fully zero-shot framework—no labeled data, no training—that embeds a target article, retrieves the top-$K$ most similar articles from an unlabeled archive, has a language-model agent read that evidence forward and backward, and lets two analyst agents debate the resulting patterns with an arbiter to break ties. Reported results on Weibo, Weibo-21, and GossipCop show MRAFnd beating all tested baselines, including specialized fact-checking pipelines, with the largest margin on Weibo-21 (86.31% accuracy versus 83.96% for the strongest baseline). The authors identify the retrieval stage as the largest contributor to performance and show the framework tolerates substantial noise in the retrieved evidence. If these results hold, detectors could flag new-event misinformation immediately, without waiting for annotated examples.

What carries the argument

The central mechanism is the retrieval-evidence loop: a weighted combination of visual and textual embeddings from a pre-trained multimodal encoder is matched by cosine similarity against an unlabeled reference archive, and the top-$K$ matches become the evidence set. Around that loop, the paper builds two reasoning instruments: Bifurcated Evidential Reasoning, in which one LLM agent scans the evidence in original order and then in reverse order to produce affirmative and negational pattern summaries, and Multi-Agent Collaborative Debate, in which two analyst agents write independent reports from those summaries and an arbiter resolves disagreements. The retrieval stage carries the empirical

What would settle it

A decisive test would be to fix the reference archive, split it so that only articles published before the target are retrievable, and remove any article sharing the target's event or named entities; if MRAFnd's accuracy stays at the reported level, the retrieval evidence is genuinely about reusable patterns, while a collapse toward the no-retrieval baseline would show the reported margins rely on temporal leakage or near-duplicate content. A simpler cross-check: replace the top-$K$ retrieved articles with top-$K$ articles from a random event and observe whether the debate stage still produces

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes an empirical claim: MRAFnd, a gradient-free workflow that combines multimodal evidence retrieval, dual-direction reasoning, and multi-agent debate, is the best zero-shot multimodal fake news detector among the systems compared, across all three datasets and both accuracy and Macro-F1. It explains this success by arguing that disinformation campaigns reuse narrative frameworks and manipulation tactics; therefore, the top-$K$ archived articles similar to a target provide the comparative context needed to expose those recycled patterns. The decisive operation is retrieval: when the framework instead analyzes a target in isolation, Macro-F1 drops by up to

Load-bearing premise

The load-bearing premise is that embedding-similar archived articles carry reliable, usable evidence about whether the target is fake; if similarity is high but the retrieved articles are irrelevant or misleading, the entire gain from the retrieval stage disappears.

Editorial extensions

If this is right

Zero-shot detection becomes a feasible deployment mode: an unannotated news archive plus off-the-shelf vision-language agents can beat methods that rely on curated labels or fine-tuning.
Retrieval quality, not model size, is the main lever: the ablation and backbone-robustness results imply that improving similarity search and evidence selection will yield larger gains than scaling the reasoning model.
The debate protocol is efficient: two analyst reports plus an arbiter resolve most conflicts, so the added cost over a single-pass system is modest.
The framework is noise-tolerant: up to roughly a third of irrelevant retrieval results costs only a few Macro-F1 points, which is the condition real retrieval systems face.
Detection can be deployed with lightweight open-source vision-language backbones while keeping the margin over specialised baselines.

Reading between the lines

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

Beyond the paper, the same retrieve-then-debate recipe should transfer to other zero-shot verification problems—rumor stance, out-of-context images, or AI-generated content claims—where an unlabeled corpus of past examples exists; the paper does not test these.
Beyond the paper, the reported gains should be checked against temporal leakage: if the reference archive contains articles published after the target or covering the same event, the retrieval stage may be exploiting near-duplicates rather than reusable tactics; the paper specifies neither the construction of the reference set nor whether targets are excluded from it.
Beyond the paper, the noise-robustness result suggests the debate stage acts as a denoiser; a testable extension would be retrieval that scores articles by expected evidentiary usefulness instead of raw embedding similarity.
Beyond the paper, a balanced reference set could change the system's prior: if fake articles dominate the archive, retrieved evidence may push verdicts toward 'fake'; measuring accuracy with reference sets of varying fake/real ratios would expose that dependence.

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

Short version: this is a coherent retrieval-augmented multi-agent pipeline for zero-shot multimodal fake news detection, and the ablations suggest each piece contributes. But the headline SOTA numbers are not yet trustworthy: the reference set S_ref is never specified, retrieval quality is never measured, and the key hyperparameters K and R are selected on the test sets. The broken FactAgent citation does not help.

What is new: the assembly of CLIP retrieval, forward/reverse evidence distillation, and analyst/arbiter debate is new as a package, and the zero-shot framing on three benchmarks is a useful testbed. The ablation study is actually informative: removing retrieval causes the largest drop, and the noise-robustness experiment shows graceful degradation, which suggests the pipeline is built around a plausible mechanism.

Where it is soft: S_ref is load-bearing but described only as an 'unlabeled reference set.' No source, size, construction, or overlap control with test data. If the retrieval pool contains near-duplicates or same-event items, the 2.35-point gain on Weibo-21 could be leakage rather than transferable signal. No retrieval-quality statistics are reported—no label agreement of top-K neighbors, no similarity distributions, no qualitative examples. The paper claims p<0.05 but gives no test, and reports no error bars or repeated runs. K and R are tuned by scanning the test sets, so the reported numbers are partially fitted to the evaluation data. The FactAgent baseline is cited with '[?]' and reference [8] is a POI-recommendation paper, so the strongest comparison is unverifiable. There is also a figure note telling readers to mentally map module names to new names, which suggests the manuscript was not fully cleaned up.

None of these are fatal to the core idea. The framework is sensible and the ablations support it. But the empirical claim as stated is overreaching. It needs at minimum a full description of S_ref, leakage controls, variance reporting, and a fixed FactAgent citation.

This paper is for people working on zero-shot fake news detection and multi-agent LLM systems. It deserves a serious referee because it is a real attempt at a hard problem with a structured evaluation, but it should be sent back for major revision. I would not cite the current numbers.

Referee Report

5 major / 5 minor

Summary. The paper proposes MRAFnd, a training-free retrieval-augmented multi-agent framework for zero-shot multimodal fake news detection. Given a target article, it retrieves top-K articles from an unlabeled reference set using CLIP-based multimodal embeddings, performs affirmative and negational LLM passes over the retrieved evidence, and resolves analyst disagreement via an arbiter. On Weibo, Weibo-21, and GossipCop, it reports accuracy of 84.75/86.31/73.55%, surpassing the strongest baseline by up to 2.35 points; ablations attribute the largest contribution to retrieval.

Significance. If the empirical claims hold, MRAFnd is a meaningful advance: it is a fully zero-shot, gradient-free pipeline requiring no labeled training data, and its multi-agent debate design is transparent and interpretable. The efficiency comparison and backbone-robustness analysis are useful additions. However, the central SOTA claim is currently under-supported: the retrieval corpus is unspecified, hyperparameters are selected using the test sets, and the strongest baseline is mis-cited. These issues must be resolved before the performance numbers can be taken at face value.

major comments (5)

[§3.2 (Eqs. 1–3); §4.3 (Fig. 2)] The retrieval stage is the largest contributor in the ablation, but S_ref is only described as an 'unlabeled reference set.' Specify its source, size, and construction, and state explicitly whether it is disjoint from S_test and whether target articles or near-duplicates are removed from the retrieval pool. Without this, the 2.35-point gain over FactAgent may be inflated by leakage (e.g., retrieving same-event articles). Also report retrieval-quality statistics: label agreement of top-K neighbors, similarity-score distributions, and qualitative examples.
[§4.4 (Figs. 3–4)] K=5 and R=2 are selected by varying K and R and observing performance on the same benchmark test sets reported in Table 1. This is effectively test-set fitting, and the reported results are the selected optimum. Provide a validation split or report results for all parameter settings; at minimum, clearly disclose this selection procedure in Section 4.1.
[Table 1; ref [8]] The strongest baseline, FactAgent, appears as '[?]' and is cited to [8], which in the bibliography is a paper on next point-of-interest recommendation, not fake news. The comparison therefore cannot be verified. Please correct the citation and specify the exact version/configuration used. The caption's 'p<0.05' also lacks a description of the statistical test, number of runs, and variance; add these details or remove the claim.
[§3.4 vs §4.4] Equation (8) defines a single arbitration step, but the implementation uses R=2 and Fig. 4 varies 'debate rounds.' There is no recurrence or loop in the method description. Define mathematically how R enters the debate, including the refinement step; otherwise the sensitivity analysis and the efficiency numbers are not reproducible.
[§4.6 (Fig. 5)] The noise-robustness result is ambiguous: a 4–5% Macro-F1 drop with 30% random replacements could indicate either robust fusion or that the retrieved evidence contributes little. Plot the 'w/o Retrieval' ablation (Fig. 2) on the same axes and report retrieval-noise results together with retrieval-quality statistics to distinguish these interpretations.

minor comments (5)

[Fig. 1 caption] The note 'Module names in the figure should be mentally mapped to the new names used in this paper' indicates an inconsistency between the figure and the text; please make the figure consistent.
[Throughout] The model name 'LLaV A' appears in Tables and Figures; it should be 'LLaVA.'
[Fig. 6] The y-axis label reads 'RAMMF Performance'; this is a typo for 'MRAFnd Performance.'
[§4.5 (Table 2)] Report the number of samples or runs used for the average token/time measurements, and include any variance, so the efficiency comparison is reproducible.
[§4.1] The datasets are listed, but the paper never states which split of each dataset is used as S_ref. This is essential for reproducibility and should be stated explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Hyperparameters K and R are selected on the same test sets used for the headline SOTA numbers, making the reported gains partly a fit.

fitted input called prediction [Sec. 4.1 Implementation Details; Sec. 4.4 Parameter Sensitivity Analysis (RQ3); Table 1]
"For evidence retrieval, we set the number of retrieved articles K=5 and the number of debate rounds R=2, based on the analysis in Section 4.4. ... We therefore select K=5 as the optimal value. ... We choose R=2 for an optimal balance between performance and efficiency."

Section 4.4 sweeps K and R on the same three benchmark test sets (Weibo, Weibo-21, GossipCop) and reports Macro-F1 for each. The values K=5 and R=2 are chosen because they produce the best numbers on those test sets. Table 1 then reports MRAFnd's accuracy/Macro-F1 on those same test sets using K=5 and R=2. Hence the headline 'prediction' (e.g., 86.31% on Weibo-21, 2.35 points above FactAgent) is evaluated under hyperparameters selected using the test labels of the very datasets being 'predicted.' The claimed SOTA margin is therefore in part a selection artifact, not an independent out-of-sample result.

full rationale

No equation-level circularity, no load-bearing self-citation chain, and no imported uniqueness theorem: MRAFnd is an empirical pipeline whose components (CLIP embedding retrieval, dual-pass LLM reasoning, debate) are defined independently of the outcome. The one substantive circular-content signal is the hyperparameter selection protocol. Section 4.1 sets K=5 and R=2 'based on the analysis in Section 4.4,' and Section 4.4 chooses these values by sweeping exactly the same test sets whose results appear in Table 1. This is a fitted-input-called-prediction pattern: the reported SOTA numbers are partly produced by selecting K and R on the evaluation data. The underspecified reference set S_ref and the broken FactAgent citation are serious validity concerns, but they are not circular in the definitional sense: the paper does not define its result in terms of them. Baseline comparisons are independent, and the core retrieval/reasoning design is not derived from the target numbers, so the circularity is partial rather than total. Score 6 reflects that one of the central empirical claims is partially reduced to a test-set fit.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework rests on three main free choices: the embedding weighting coefficients (unstated), and the retrieval and debate hyperparameters K and R (both tuned on the test sets). The core axioms are domain assumptions about the usefulness of CLIP-based retrieval and the reliability of LLM agents in zero-shot reasoning. No new theoretical entities, forces, or particles are introduced.

free parameters (3)

lambda_v, lambda_t (modality weighting coefficients) = not specified
Eq. (1) uses fixed hyperparameters to combine visual and textual embeddings; no values are given, and they affect retrieval and hence the evidence used for all downstream reasoning.
K (number of retrieved articles) = 5
Chosen by scanning K in {1,3,5,8,10} on the benchmark test sets (Section 4.4, Figure 3), so the final result is fitted to the evaluation data.
R (number of debate rounds) = 2
Chosen by scanning R in {1,2,3,4} on the test sets (Section 4.4, Figure 4), likewise tuned on the evaluation data.

assumptions (4)

domain assumption Pre-trained CLIP encoders yield embeddings that capture the semantic and visual similarity needed for retrieving relevant evidence.
Invoked in Section 3.2 Eq. (1)-(3); no validation that CLIP similarity correlates with useful evidence for veracity, and lambda_v/lambda_t are unstated.
domain assumption Articles retrieved by embedding similarity from an unlabeled reference set provide evidence that helps determine the target's veracity.
This is the central hypothesis of the paper (Sections 3.2-3.3); it is not independently tested in the paper beyond the main results, and retrieval noise is only added synthetically.
domain assumption LLM agents can reliably extract disinformation patterns and render reasoned verdicts from the retrieved evidence in a zero-shot manner using the described prompts.
Sections 3.3-3.4 rely on the reasoner/analyst/arbiter agents making correct judgments; the actual prompts are not shown, so the reader cannot inspect what knowledge or bias is being injected.
domain assumption The unlabeled reference set S_ref does not contain the target test articles and has no label leakage.
The paper does not state how S_ref is constructed or whether the target article is excluded from its own retrieval pool (Section 3.2), which matters for a zero-shot claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection." pith.science (2026). https://pith.science/paper/PJQWMOWC

@misc{pith2026260801430,
  author       = {Pith},
  title        = {Pith review of: MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJQWMOWC}},
  note         = {Machine review of arXiv:2608.01430}
}

read the original abstract

The rapid dissemination of multimodal content has intensified the spread of fabricated news, presenting a substantial threat to social integrity. A formidable challenge for current detection systems is identifying misinformation related to novel events in zero-shot scenarios. Prevailing zero-shot methods typically assess news items in isolation via semantic matching, a strategy that fails to recognize the recycled disinformation tactics from past campaigns and lacks the sophisticated reasoning needed to identify subtle, cross-modal discrepancies. To surmount these deficiencies, we introduce \textbf{MRAFnd}, a novel \underline{\textbf{M}}ultimodal \underline{\textbf{R}}etrieval-\underline{\textbf{A}}ugmented Framework for Zero-Shot \underline{\textbf{F}}ake \underline{\textbf{N}}ews \underline{\textbf{D}}etection. MRAFnd emulates a collaborative team of analysts to verify news veracity. The framework initiates with \textbf{Multimodal Similarity-based News Retrieval} to assemble a corpus of contextually analogous articles from an unlabeled reference database. Subsequently, during the \textbf{Bifurcated Evidential Reasoning} stage, agents perform a dual-directional analysis to extract critical patterns from the retrieved evidence. Finally, a \textbf{Multi-Agent Collaborative Debate}, involving Analyst and Arbiter agents, engages in a structured discourse to arrive at a definitive and robust conclusion. Comprehensive experiments on three benchmark datasets reveal that MRAFnd markedly surpasses state-of-the-art baselines, achieving an accuracy gain of up to 2.35\% on the demanding Weibo-21 dataset.

Figures

Figures reproduced from arXiv: 2608.01430 by the authors.

**Figure 2.** Ablation study of MRAFnd on three datasets. Performance is reported as Macro [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Performance of MRAFnd with a varying number of retrieved articles ( [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Performance of MRAFnd with a varying number of deliberation rounds ( [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

**Figure 5.** Figure 5: Performance of MRAFnd under varying levels of retrieval noise. The noise ratio [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Performance (Macro-F1 %) of the MRAFnd framework with different MLLM [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

[8]

In: Proceedings of the 47th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval

Li, P., de Rijke, M., Xue, H., Ao, S., Song, Y ., Salim, F.D.: Large language mod- els for next point-of-interest recommendation. In: Proceedings of the 47th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1463–1472 (2024)

work page 2024
[1]

arXiv preprint arXiv:2310.09478 (2023)

Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V ., Xiong, Y ., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)

arXiv 2023
[2]

In: Proceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval

Cui, X., Lu, W., Tong, Y ., Li, Y ., Zhao, Z.: Diffusion-based multi-modal synergy interest network for click-through rate prediction. In: Proceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 581–591 (2025)

work page 2025
[3]

In: Proceed- ings of the 48th International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval

Cui, X., Lu, W., Tong, Y ., Li, Y ., Zhao, Z.: Multi-modal multi-behavior sequential recommendation with conditional diffusion-based feature denoising. In: Proceed- ings of the 48th International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval. pp. 1593–1602 (2025)

work page 2025
[4]

Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)

2023
[5]

ACM Computing Surveys55(4), 1–37 (2022)

Giachanou, A., Ghanem, B., Ríssola, E.A., Rosso, P., Crestani, F., Oberski, D.: The impact of psycholinguistic patterns in discriminating between fake and real news. ACM Computing Surveys55(4), 1–37 (2022)

work page 2022
[6]

In: The world wide web conference

Khattar, D., Goud, J.S., Gupta, M., Varma, V .: Mvae: Multimodal variational au- toencoder for fake news detection. In: The world wide web conference. pp. 2915– 2921 (2019)

work page 2019
[7]

Advances in Neural Information Processing Systems33, 9459–9474 (2020)

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V ., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems33, 9459–9474 (2020)

work page 2020

Show all 25 references

[9]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Liu, H., Li, C., Li, Y ., Lee, Y .J.: Improved baselines with visual instruction tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26296–26306 (2024)

2024
[10]

arXiv preprint arXiv:2402.07776 (2024)

Liu, H., Wang, W., Li, H., Li, H.: Teller: A trustworthy framework for ex- plainable, generalizable and controllable fake news detection. arXiv preprint arXiv:2402.07776 (2024)

2024 arXiv
[11]

In: Proceedings of the 31st ACM International Conference on Multimedia

Liu, M., Liang, K., Hu, D., Yu, H., Liu, Y ., Meng, L., Tu, W., Zhou, S., Liu, X.: Tmac: Temporal multi-modal graph learning for acoustic event classification. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 3365– 3374 (2023)

2023
[12]

In: Proceedings of the AAAI Conference on Artificial Intelli- gence

Lu, W., Tong, Y ., Ye, Z.: Dammfnd: Domain-aware multimodal multi-view fake news detection. In: Proceedings of the AAAI Conference on Artificial Intelli- gence. vol. 39, pp. 559–567 (2025) 14 Lehan Zhang et al

2025
[13]

In: Proceedings of the 33rd ACM Inter- national Conference on Multimedia

Lu, W., Yin, L.: Dmmd4sr: Diffusion model-based multi-level multimodal de- noising for sequential recommendation. In: Proceedings of the 33rd ACM Inter- national Conference on Multimedia. pp. 6363–6372 (2025)

2025
[14]

In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)

Ma, Z., Luo, M., Guo, H., Zeng, Z., Hao, Y ., Zhao, X.: Event-radar: Event-driven multi-view learning for multimodal fake news detection. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 5809–5821 (2024)

2024
[15]

In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management

Nan, Q., Cao, J., Zhu, Y ., Wang, Y ., Li, J.: Mdfend: Multi-domain fake news detec- tion. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management. pp. 3343–3347 (2021)

2021
[16]

In: Proceedings of the 36th annual acm symposium on user interface software and technology

Park, J.S., O’Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Gen- erative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th annual acm symposium on user interface software and technology. pp. 1–22 (2023)

2023
[17]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Qi, P., Yan, Z., Hsu, W., Lee, M.L.: Sniffer: Multimodal large language model for explainable out-of-context misinformation detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13052– 13062 (2024)

2024
[18]

In: Proceedings of the 44th international acm sigir conference on research and development in information retrieval

Qian, S., Wang, J., Hu, J., Li, J., Fang, Q.: Hierarchical multi-modal contextual at- tention network for fake news detection. In: Proceedings of the 44th international acm sigir conference on research and development in information retrieval. pp. 1957–1961 (2021)

1957
[19]

In: International conference on machine learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PMLR (2021)

2021
[20]

Big data8(3), 171–188 (2020)

Shu, K., Mahudeswaran, D., Wang, S., Lee, D., Liu, H.: Fakenewsnet: A data repository with news content, social context, and spatiotemporal information for studying fake news on social media. Big data8(3), 171–188 (2020)

2020
[21]

In: Proceedings of the 33rd ACM Interna- tional Conference on Multimedia

Tong, Y ., Lu, W., Cui, X., Mao, Y ., Zhao, Z.: Dapt: Domain-aware prompt-tuning for multimodal fake news detection. In: Proceedings of the 33rd ACM Interna- tional Conference on Multimedia. pp. 7902–7911 (2025)

2025
[22]

In: Proceedings of the 32nd ACM International Conference on Multimedia

Tong, Y ., Lu, W., Zhao, Z., Lai, S., Shi, T.: Mmdfnd: Multi-modal multi-domain fake news detection. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 1178–1186 (2024)

2024
[23]

In: Intrinsically-Motivated and Open-Ended Learning Workshop@ NeurIPS2023 (2023)

Wang, G., Xie, Y ., Jiang, Y ., Mandlekar, A., Xiao, C., Zhu, Y ., Fan, L., Anand- kumar, A.: V oyager: An open-ended embodied agent with large language models. In: Intrinsically-Motivated and Open-Ended Learning Workshop@ NeurIPS2023 (2023)

2023
[24]

In: Proceedings of the 24th acm sigkdd international conference on knowledge discovery & data mining

Wang, Y ., Ma, F., Jin, Z., Yuan, Y ., Xun, G., Jha, K., Su, L., Gao, J.: Eann: Event adversarial neural networks for multi-modal fake news detection. In: Proceedings of the 24th acm sigkdd international conference on knowledge discovery & data mining. pp. 849–857 (2018)

2018
[25]

ACM Computing Surveys (CSUR)53(5), 1–40 (2020)

Zhou, X., Zafarani, R.: A survey of fake news: Fundamental theories, detec- tion methods, and opportunities. ACM Computing Surveys (CSUR)53(5), 1–40 (2020)

2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [8]

In: Proceedings of the 47th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval

Li, P., de Rijke, M., Xue, H., Ao, S., Song, Y ., Salim, F.D.: Large language mod- els for next point-of-interest recommendation. In: Proceedings of the 47th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1463–1472 (2024)

work page 2024

[2] [1]

arXiv preprint arXiv:2310.09478 (2023)

Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V ., Xiong, Y ., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)

arXiv 2023

[3] [2]

In: Proceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval

Cui, X., Lu, W., Tong, Y ., Li, Y ., Zhao, Z.: Diffusion-based multi-modal synergy interest network for click-through rate prediction. In: Proceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 581–591 (2025)

work page 2025

[4] [3]

In: Proceed- ings of the 48th International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval

Cui, X., Lu, W., Tong, Y ., Li, Y ., Zhao, Z.: Multi-modal multi-behavior sequential recommendation with conditional diffusion-based feature denoising. In: Proceed- ings of the 48th International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval. pp. 1593–1602 (2025)

work page 2025

[5] [4]

Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)

2023

[6] [5]

ACM Computing Surveys55(4), 1–37 (2022)

Giachanou, A., Ghanem, B., Ríssola, E.A., Rosso, P., Crestani, F., Oberski, D.: The impact of psycholinguistic patterns in discriminating between fake and real news. ACM Computing Surveys55(4), 1–37 (2022)

work page 2022

[7] [6]

In: The world wide web conference

Khattar, D., Goud, J.S., Gupta, M., Varma, V .: Mvae: Multimodal variational au- toencoder for fake news detection. In: The world wide web conference. pp. 2915– 2921 (2019)

work page 2019

[8] [7]

Advances in Neural Information Processing Systems33, 9459–9474 (2020)

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V ., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems33, 9459–9474 (2020)

work page 2020