Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models

T0 review · 4 major / 5 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 Structured adversarial debate among domain-specialized LLM agents detects misinformation better than static classifiers and generic debate.

desk verdict A well-structured multi-agent debate pipeline for misinformation detection, but the compute-unmatched baseline and post-hoc data filtering undercut the headline F1 claims. read the letter →

arxiv 2505.18596 v4 pith:ZO3UJDPG submitted 2025-05-24 cs.CL cs.AI

Chen Han , Wenzhen Zheng , Xijin Tang This is my paper · ORCID
classification cs.CLcs.AI
keywords misinformationdetectionmulti-agentdebatelargelanguagemodelsfact-checkingdomain-specificagentsmulti-dimensionalevaluationzero-sumscoringinterpretability
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

The paper's central claim is that misinformation detection improves when it is set up as a structured adversarial debate among large-language-model agents rather than as a single-pass classification. Its Debate-to-Detect (D2D) framework gives four debaters on each side a fixed stance and a domain-specific profile, runs them through five debate stages, and has a panel of judges score the claim on five dimensions with complementary scores that sum to seven, making a tie impossible. On the Weibo21 and FakeNewsDataset benchmarks, D2D reaches 81.97% and 81.94% F1, respectively, above fine-tuned transformer classifiers and standard debate baselines. The result matters because it points toward LLM-based detection that is both more accurate and more explainable, with a verdict trail that resembles human fact-checking.

What carries the argument

The load-bearing mechanism is the five-stage structured debate with role separation and zero-sum multi-dimensional scoring. Each input text is first assigned a topical domain, then eight debaters—four defending 'real' and four defending 'fake'—receive profiles relevant to that domain and speak in fixed stages; a shared-memory summary is generated before each turn to keep agents on the central points of contention. Six judge agents then produce a neutral synopsis and five dimension-specific scores, with each dimension's two integer scores constrained to sum to 7, so the final classification cannot tie. This design does two jobs: it forces adversarial evidence production before a verdict, and it makes the verdict decomposable into named criteria.

What would settle it

Re-run D2D and its strongest baselines on the full, unfiltered Weibo21 dataset, including the removed long and low-quality samples, and compare F1 and recall; if D2D's advantage shrinks to noise or reverses, the claim of general superiority fails. A second check is to collect human fact-checker ratings on the five dimensions and compare them with D2D's dimension scores, since large disagreement would undercut the interpretability claim.

Watch

Extended reading notes

Core claim

Stated as the authors would state it: a fact-checking-shaped debate protocol, not a bigger model, is what carries misinformation detection. D2D dynamically infers the topical domain of each claim, generates concise domain profiles for debaters, and runs a five-stage exchange—opening, rebuttal, free debate, closing, judgment—with a compressed shared memory feeding each turn. Five judge agents then assign zero-sum integer scores on Factuality, Source Reliability, Reasoning Quality, Clarity, and Ethics, and the aggregate score yields a definitive REAL or FAKE verdict. The reported evidence is that full D2D outperforms every baseline on both datasets and every metric, and that the multi-dimensional judgment component is the largest single contributor in ablations, followed by domain profiles and staged design. The case study shows the debate progressively refining evidence, for example a COVID-19 liquor claim countered by WHO ethanol-inactivation thresholds.

Load-bearing premise

The preprocessing that removes the top 5% longest samples and manually judged 'low-quality' Weibo21 items is unbiased; if those are the hard cases where every method fails, D2D's measured lead over baselines may not survive on unfiltered real-world data.

Editorial extensions

If this is right

  • D2D's reported F1 of 81.97% on Weibo21 and 81.94% on FakeNewsDataset tops fine-tuned BERT/RoBERTa, zero-shot prompting, chain-of-thought, self-reflection, and standard multi-agent debate.
  • Removing the multi-dimensional judgment hurts more than removing domain profiles or the staged design, suggesting that evaluation design drives most of the gain.
  • The judgment stage is the most capability-sensitive stage: replacing GPT-4o with GPT-3.5-turbo there drops F1 by 6.87%, while GPT-4.1 raises it by 3.03%.
  • The framework is robust to speaker-order reversal and neutral relabeling, with more than 90% of perturbed samples staying within a 5-point score deviation.
  • On news published after GPT-4o's training cutoff, D2D still reaches 83.92% accuracy and 79.83% F1, indicating the gains are not just memorized-content retrieval.

Reading between the lines

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

  • The zero-sum scoring trick makes every debate resolvable by construction; a natural next test is whether the five dimension scores align with independent human fact-checker ratings, which the paper does not report.
  • The preprocessing that drops the longest and manually judged low-quality samples likely makes the evaluation optimistic; the appendix's full-dataset results with lower recall suggest D2D's advantage on clean inputs may not extend to uncurated social media streams.
  • The same debate-and-zero-sum scaffold should transfer to other binary verification tasks such as rumor stance detection, claim matching, or fact-check explanation generation, provided domain profiles can be generated automatically.
  • Because D2D runs 14 agents and five stages, its practical value depends on a cost-accuracy tradeoff; comparing it to cheaper single-LLM pipelines on a per-dollar or per-latency basis would clarify where deployment is justified.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes Debate-to-Detect (D2D), a multi-agent LLM framework that recasts misinformation detection as a structured five-stage debate among domain-profiled debaters, followed by a multi-dimensional zero-sum judgment by multiple judges. On preprocessed Weibo21 and FakeNewsDataset, D2D reports F1 scores of 81.97% and 81.94%, outperforming BERT, RoBERTa, zero-shot prompting, chain-of-thought, self-refinement, and standard multi-agent debate (SMAD). The paper also presents ablations of domain profiles, stage design, and multi-dimensional judgment, a stage-substitution analysis, perturbation tests for speaker-order and lexical-framing biases, a round-number analysis, and a generalization test on 2025 Chinese news.

Significance. If the reported results hold, D2D would be a useful contribution: it provides a concrete instantiation of debate-based fact-checking with external ground-truth evaluation, component-level ablations, and a post-training generalization check. The evaluation is not circular, since final judgments are scored against human labels. The main value is the demonstration that structured, role-specialized debate and multi-dimensional scoring can improve interpretable LLM-based detection. However, the significance is presently conditional on resolving two threats: a post-hoc dataset preprocessing step that may inflate measured gains, and an unmatched inference budget relative to the SMAD baseline.

major comments (4)
  1. [§4.1, Appendix A, Tables 2 and 7] The headline results are reported on preprocessed datasets from which the top 5% longest samples and manually judged 'low-quality' Weibo21 samples are removed. Appendix A shows that on the original Weibo21 dataset D2D's F1 falls from 81.97% to 76.79%, yet no baseline is reported on the original data, so it is unknown whether the preprocessing differentially favors D2D or simply removes hard cases for all methods. The claim that D2D 'significantly outperforms' baselines should be re-established on the original datasets, or at least with all baselines evaluated on the same included and excluded subsets, with an analysis of the removed samples.
  2. [§3 and §4.1, Table 2] D2D uses 8 debaters and 6 judges (14 agents) over 5 stages with shared-memory summarization, while the SMAD baseline uses 3 agents over 4 rounds. The F1 gaps over SMAD (5.45 points on Weibo21 and 6.97 on FakeNewsDataset) are therefore confounded by the number of LLM calls and generated tokens. Because the ablations keep the full 14-agent, 5-stage architecture, they do not control for compute. The central claim that the structured debate design, rather than additional inference budget, drives the gains requires budget-matched comparisons, e.g., SMAD with a comparable number of agents/rounds or D2D with a matched token budget.
  3. [§5.2, Table 3] The robustness analysis is internally inconsistent. The text says 'Only 7 cases show variations, all within 5 points' and 'the remaining 5 disagreements also fall within 5 points,' while Table 3 reports 7 inconsistent Fake samples and 5 inconsistent Real samples, with additional samples in the 5–10 band. The categories 'consistent' and 'inconsistent' are not defined relative to the delta ranges, and the numbers do not sum coherently. As presented, the table does not support the stated conclusion that over 90% of samples are robust.
  4. [§4.2 and §5.1, Table 2 and Figure 4] No error bars, variance estimates, or significance tests are reported for any of the main or ablation results, although differences are often small (e.g., 1.39 F1 points between D2D and RoBERTa on Weibo21). Additionally, the GPT-4o baseline F1 in Section 5.1 is 81.55% on FakeNewsDataset, whereas Table 2 reports 81.94% for the same D2D configuration; the discrepancy should be explained or corrected. Without repeated runs or significance testing, the word 'significantly' in the abstract and Section 4.2 is not supported.
minor comments (5)
  1. [§1 and §4.1] There are typos: 'judegments' should be 'judgments', 'aovid' should be 'avoid', and 'are are' should be 'are'.
  2. [§4.1, D2D w/o SD] The text says the variant 'eliminates the structured four-stage debate process,' but D2D is described as having five stages; clarify whether only the four debate stages before judgment are removed.
  3. [§3.3] The phrase 'the total score across all dimensions is inherently imbalanced' is imprecise; a tie is impossible because the five zero-sum pairs sum to 35 (an odd number), not because the score is 'inherently imbalanced'.
  4. [§5.3, Table 4] The round configurations mix debate stages and free-debate repetitions; consider defining 'rounds' consistently, since 'Round 1' includes only Opening while 'Round 4' includes four stages.
  5. [Figure 3] The case-study figure is dense and difficult to read; enlarging or annotating the debate transcript would improve interpretability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: D2D's accuracy claims are benchmarked against external ground-truth labels, with no parameters fitted to the test data and no load-bearing self-citations.

full rationale

The paper's central claim is empirical: D2D achieves higher F1 than baselines on Weibo21 and FakeNewsDataset, and on a newly constructed post-cutoff news benchmark. The final verdict is produced by LLM judges from the debate transcript, and scored against external binary labels; the scoring rubric and debate stages are fixed design choices, not fitted parameters. No equation or mechanism in the paper reduces a prediction to a fitted value or to a prior output. The preprocessing step removes long and low-quality samples, but this is a data-quality filter applied before any model prediction and is not derived from the model's errors or fitted to the labels; the appendix reports results on the original datasets, showing the filter's effect rather than hiding it. The framework's internal evaluations (e.g., zero-sum scoring) are structural properties that guarantee a binary outcome, but they do not determine accuracy against ground truth. All citations are to external prior work (e.g., Liang et al., Chan et al., Soprano et al.) and are used as design inspirations or comparative baselines, not as premises that force the reported results. The strongest residual concern is an unmatched inference budget between D2D and SMAD (14 agents over 5 stages vs. 3 agents over 4 rounds), which is a validity or confound issue for the comparison, not a circularity in the derivation chain. Therefore, no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the reliability of the datasets' ground-truth labels, on GPT-4o serving as both debater and impartial judge, and on the preprocessing exclusions being unbiased. The design hyperparameters (agent counts, score pool, dimensions) are chosen by hand and not systematically swept.

free parameters (4)
  • Zero-sum score pool per dimension = 7
    Arbitrary integer pool; ensures no tie but not fitted. Affects score granularity, not central accuracy claim.
  • Number of debater agents per side = 4
    Chosen by hand; not ablated. D2D uses 8 debaters plus 6 judges (14 agents), versus SMAD's 3 agents, so compute is not matched.
  • Number of free debate rounds = 1
    Fixed; Section 5.3 shows optimal round count varies with text length (4-6), yet final model uses 1. Potential mismatch.
  • Evaluation dimensions = 5
    Factuality, Source Reliability, Reasoning Quality, Clarity, Ethics; borrowed from Soprano et al. 2021, not tuned.
assumptions (6)
  • domain assumption Ground-truth labels of Weibo21 and FakeNewsDataset are reliable
    The accuracy claim is measured against these labels; if labels are noisy, the reported improvements are unreliable.
  • domain assumption GPT-4o can serve as both debater and judge without systematic self-bias
    The framework relies on the same model to argue and to evaluate; the paper tests only order and label robustness, not deeper self-preference bias.
  • domain assumption The five truthfulness dimensions from Soprano et al. transfer to LLM-based judgment
    The multi-dimensional rubric is taken from a crowdsourcing study; no validation that LLM judges interpret the dimensions consistently.
  • domain assumption The 'tit for tat' fixed-stance strategy improves reasoning
    Adopted from Liang et al. without re-validation in this setting; not ablated here.
  • domain assumption Post-2024 news samples are not in GPT-4o pretraining
    The authors manually checked absence of verbatim overlaps, but that does not rule out paraphrase-level memorization.
  • ad hoc to paper Excluding the top 5% longest and low-quality samples does not bias results
    The preprocessing criteria are not operationalized; this assumption is load-bearing for the main results.
invented entities (1)
  • Domain-specific agent profiles
    purpose: Specialize debater and judge agents by topical domain to improve argument quality
    Introduced by the framework; no external evidence of their independent existence beyond the paper's experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models." pith.science (2026). https://pith.science/paper/ZO3UJDPG

@misc{pith2026250518596,
  author       = {Pith},
  title        = {Pith review of: Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZO3UJDPG}},
  note         = {Machine review of arXiv:2505.18596}
}
read the original abstract

The proliferation of misinformation in digital platforms reveals the limitations of traditional detection methods, which mostly rely on static classification and fail to capture the intricate process of real-world fact-checking. Despite advancements in Large Language Models (LLMs) that enhance automated reasoning, their application to misinformation detection remains hindered by issues of logical inconsistency and superficial verification. In response, we introduce Debate-to-Detect (D2D), a novel Multi-Agent Debate (MAD) framework that reformulates misinformation detection as a structured adversarial debate. Inspired by fact-checking workflows, D2D assigns domain-specific profiles to each agent and orchestrates a five-stage debate process, including Opening Statement, Rebuttal, Free Debate, Closing Statement, and Judgment. To transcend traditional binary classification, D2D introduces a multi-dimensional evaluation mechanism that assesses each claim across five distinct dimensions: Factuality, Source Reliability, Reasoning Quality, Clarity, and Ethics. Experiments with GPT-4o on two datasets demonstrate significant improvements over baseline methods, and the case study highlight D2D's capability to iteratively refine evidence while improving decision transparency, representing a substantial advancement towards interpretable misinformation detection. The code will be released publicly after the official publication.

Figures

Figures reproduced from arXiv: 2505.18596 by the authors.

Figure 1
Figure 1. In Standard Multi-Agent Debate (SMAD), [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The D2D framework structures misinformation detection as a multi-agent debate, comprising two layers: [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Case Study – A Demonstration of the Structured MAD in the D2D Framework. The process reflects real [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance Comparison of Model Variants [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Effect of Debate Rounds on F1-Score Across [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Examples of low-quality samples in Weibo21. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Beyond a Joke: Multi-Angle Reasoning for Detecting and Explaining Harmful Humor in Memes

    cs.AI 2026-07 conditional novelty 6.0 of 10

    MAR-12 improves humor and hate detection in memes by prompting a VLM through twelve reasoning perspectives, attention-weighting them, and generating explanations from the weighted evidence.

Reference graph

Works this paper leans on

41 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Alfonso Amayuelas, Xianjun Yang, Antonis Antoniades, Wenyue Hua, Liangming Pan, and William Yang Wang. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.407 M ulti A gent collaboration attack: Investigating adversarial attacks in large language model collaborations via debate . In Findings of the Association for Computational Linguistics: EMNLP 2024, ...

    doi:10.18653/v1/2024.findings-emnlp.407 2024

  2. [2]

    Dickerson

    Wang Angelina, Morgenstern Jamie, and Dickerson P. Dickerson. 2025. https://doi.org/10.1038/s42256-025-00986-z Large language models that replace human participants can harmfully misportray and flatten identity groups . Nature Machine Intelligence, 7:400--411

    doi:10.1038/s42256-025-00986-z 2025

  3. [3]

    Do, Yan Xu, and Pascale Fung

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. https://doi.org/10.18653/v1/2023.ijcnlp-main.45 A multitask, multilingual, multimodal evaluation of C hat GPT on reasoning, hallucination, and interactivity . In Proceedings of the 13...

    doi:10.18653/v1/2023.ijcnlp-main.45 2023

  4. [4]

    Pan, Shuyi Yang, Lakshya A

    Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. 2025. https://arxiv.org/abs/2503.13657 Why do multi-agent llm systems fail? Preprint, arXiv:2503.13657

    arXiv 2025

  5. [5]

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2024. https://openreview.net/forum?id=FQepisCUWu Chateval: Towards better LLM -based evaluators through multi-agent debate . In The Twelfth International Conference on Learning Representations

    work page 2024

  6. [6]

    Sunhao Dai, Chen Xu, Shicheng Xu, Liang Pang, Zhenhua Dong, and Jun Xu. 2024. https://doi.org/10.1145/3637528.3671458 Bias and unfairness in information retrieval systems: New challenges in the llm era . KDD '24, page 6437–6447, New York, NY, USA. Association for Computing Machinery

    arXiv 2024

  7. [7]

    Rupak Kumar Das and Jonathan Dodge. 2025. https://arxiv.org/abs/2501.18649 Fake news detection after llm laundering: Measurement and explanation . Preprint, arXiv:2501.18649

    work page Pith review arXiv 2025

  8. [8]

    Zehang Deng, Yongjian Guo, Changzhou Han, Wanlun Ma, Junwu Xiong, Sheng Wen, and Yang Xiang. 2025. https://doi.org/10.1145/3716628 Ai agents under threat: A survey of key security challenges and future pathways . ACM Comput. Surv., 57(7)

    doi:10.1145/3716628 2025

Show all 41 references
  1. [9]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2024. Improving factuality and reasoning in language models through multiagent debate. ICML'24. JMLR.org

    2024

  2. [10]

    A \"i meur Esma, Amri Sabrine, and Brassard Gilles. 2023. https://doi.org/10.1007/s13278-023-01028-5 Fake news, disinformation and misinformation in social media: a review . Social Network Analysis and Mining, 13(1):30

    2023 doi

  3. [11]

    Zhibin Gou, Zhihong Shao, Yeyun Gong, yelong shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2024. https://openreview.net/forum?id=Sx038qxjek CRITIC : Large language models can self-correct with tool-interactive critiquing . In The Twelfth International Conference on Learning Rep...

    2024

  4. [12]

    Bing He, Mustaque Ahamad, and Srijan Kumar. 2023. https://doi.org/10.1145/3543507.3583388 Reinforcement learning-based counter-misinformation response generation: A case study of covid-19 vaccine misinformation . In Proceedings of the ACM Web Conference 2023, WWW '23, page 269...

    2023

  5. [13]

    Zhitao He, Pengfei Cao, Chenhao Wang, Zhuoran Jin, Yubo Chen, Jiexin Xu, Huaijun Li, Kang Liu, and Jun Zhao. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.549 A gents C ourt: Building judicial decision-making agents with court debate simulation and legal knowledge augm...

    2024 doi

  6. [14]

    Tianyi Huang, Jingyuan Yi, Peiyang Yu, and Xiaochuan Xu. 2025. https://arxiv.org/abs/2503.00724 Unmasking digital falsehoods: A comparative analysis of llm-based misinformation detection strategies . Preprint, arXiv:2503.00724

    2025 arXiv

  7. [15]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc

    2022

  8. [16]

    Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.427 Improving multi-agent debate with sparse communication topology . In Findings of the Association for Computational Linguistics: EM...

    2024 doi

  9. [17]

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.992 Encouraging divergent thinking in large language models through multi-agent debate . In Proceedings of the 2024 ...

    2024 doi

  10. [18]

    Aiwei Liu, Qiang Sheng, and Xuming Hu. 2024 a . https://doi.org/10.1145/3626772.3661377 Preventing and detecting misinformation generated by large language models . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retriev...

    2024

  11. [19]

    Yanchen Liu, Mingyu Derek Ma, Wenna Qin, Azure Zhou, Jiaao Chen, Weiyan Shi, Wei Wang, and Diyi Yang. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.846 Decoding susceptibility: Modeling misbelief to misinformation through a computational approach . In Proceedings of the...

    2024 doi

  12. [20]

    Xiaoxiao Ma, Yuchen Zhang, Kaize Ding, Jian Yang, Jia Wu, and Hao Fan. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.31 On fake news detection with LLM enhanced semantics mining . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pa...

    2024 doi

  13. [21]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. https://openrev...

    2023

  14. [22]

    Tula Masterman, Sandi Besen, Mason Sawtell, and Alex Chao. 2024. https://arxiv.org/abs/2404.11584 The landscape of emerging ai agent architectures for reasoning, planning, and tool calling: A survey . Preprint, arXiv:2404.11584

    2024 arXiv

  15. [23]

    Debjyoti Mondal, Suraj Modi, Subhadarshi Panda, Rituraj Singh, and Godawari Sudhakar Rao. 2024. https://doi.org/10.1609/aaai.v38i17.29844 Kam-cot: Knowledge augmented multimodal chain-of-thoughts reasoning . Proceedings of the AAAI Conference on Artificial Intelligence, 38(17)...

    2024 doi

  16. [24]

    M. F. Mridha, Ashfia Jannat Keya, Md. Abdul Hamid, Muhammad Mostafa Monowar, and Md. Saifur Rahman. 2021. https://doi.org/10.1109/ACCESS.2021.3129329 A comprehensive review on fake news detection with deep learning . IEEE Access, 9:156151--156170

    2021

  17. [25]

    Qiong Nan, Juan Cao, Yongchun Zhu, Yanyan Wang, and Jintao Li. 2021. https://doi.org/10.1145/3459637.3482139 Mdfend: Multi-domain fake news detection . In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, CIKM '21, page 3343–3347, New ...

    2021

  18. [26]

    Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Wang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.97 On the risk of misinformation pollution with large language models . In Findings of the Association for Computational Linguistics: EMNL...

    2023 doi

  19. [27]

    Someen Park, Jaehoon Kim, Seungwan Jin, Sohyun Park, and Kyungsik Han. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1166 PREDICT : Multi-agent-based debate simulation for generalized hate speech detection . In Proceedings of the 2024 Conference on Empirical Methods in Nat...

    2024 doi

  20. [28]

    Ver \'o nica P \'e rez-Rosas, Bennett Kleinberg, Alexandra Lefevre, and Rada Mihalcea. 2018. https://aclanthology.org/C18-1287/ Automatic detection of fake news . In Proceedings of the 27th International Conference on Computational Linguistics, pages 3391--3401, Santa Fe, New ...

    2018

  21. [29]

    Sougata Saha and Rohini Srihari. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.622 Integrating argumentation and hate-speech-based techniques for countering misinformation . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11...

    2024 doi

  22. [30]

    Upasna Sharma and Jaswinder Singh. 2024. https://doi.org/10.1007/s13278-024-01280-3 A comprehensive overview of fake news detection on social networks . Social Network Analysis and Mining, 14(1):120

    2024 doi

  23. [31]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: language agents with verbal reinforcement learning. In Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS '23, Red Hook, NY, USA. Cu...

    2023

  24. [32]

    Noam Slonim, Yonatan Bilu, Carlos Alzate, and 1 others. 2021. https://doi.org/10.1038/s41586-021-03215-w An autonomous debating system . Nature, 591:379--384

    2021 doi

  25. [33]

    Tump, Nina Ehmann, Philipp Lorenz-Spreen, Ralph Hertwig, Anton Gollwitzer, and Ralf H

    Mubashir Sultan, Alan N. Tump, Nina Ehmann, Philipp Lorenz-Spreen, Ralph Hertwig, Anton Gollwitzer, and Ralf H. J. M. Kurvers. 2024. https://doi.org/10.1073/pnas.2409329121 Susceptibility to online misinformation: A systematic meta-analysis of demographic and psychological fac...

    2024 doi

  26. [34]

    Spampatti Tobia, Hahnel Ulf J.J., Trutnevyte Evelina, and Tobias Brosch. 2024. https://doi.org/10.1038/s41562-023-01736-0 Psychological inoculation strategies to fight climate disinformation across 12 countries . Nature Human Behaviour, 8:380--398

    2024 doi

  27. [35]

    Herun Wan, Shangbin Feng, Zhaoxuan Tan, Heng Wang, Yulia Tsvetkov, and Minnan Luo. 2024. https://doi.org/10.18653/v1/2024.findings-acl.155 DELL : Generating reactions and explanations for LLM -based misinformation detection . In Findings of the Association for Computational Li...

    2024 doi

  28. [36]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processi...

    2022

  29. [37]

    Shuai Xu, Jianqiu Xu, Shuo Yu, and Bohan Li. 2024. https://doi.org/10.1145/3627673.3679788 Identifying disinformation from online social media via dynamic modeling across propagation stages . In Proceedings of the 33rd ACM International Conference on Information and Knowledge ...

    2024

  30. [38]

    Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, Nitesh V Chawla, and Xiangliang Zhang. 2024. https://arxiv.org/abs/2410.02736 Justice or prejudice? quantifying biases in llm-as-a-judge . Preprint, arXiv:...

    2024 arXiv

  31. [39]

    Yiqun Zhang, Xiaocui Yang, Shi Feng, Daling Wang, Yifei Zhang, and Kaisong Song. 2024. https://arxiv.org/abs/2408.04472 Can llms beat humans in debating? a dynamic multi-agent framework for competitive debate . Preprint, arXiv:2408.04472

    2024 arXiv

  32. [40]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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