Veracity: An Open-Source AI Fact-Checking System

Jean-Fran\c{c}ois Godbout; Kellin Pelrine; Li Wei Wang; Luda Cohen; Manon Gruaz; Maximilian Puelma Touzel; Mike Pinder; Reihaneh Rabbany; Sukanya Krishna; Taylor Lynn Curtis

REVIEW 4 major objections 4 minor 21 references

Veracity: An Open-Source AI Fact-Checking System

T0 review · 4 major / 4 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 This demo paper describes Veracity, an open-source fact-checking system in which an LLM teams with web retrieval agents to return a 0-100 reliability score, a plain-language explanation, and the sources behind the verdict.

desk verdict A clear demo paper describing a real open-source fact-checking tool, but with unsupported reliability claims and an uncalibrated 60% share threshold. read the letter →

arxiv 2506.15794 v1 pith:UITHABTV submitted 2025-06-18 cs.CL cs.AIcs.HC

Taylor Lynn Curtis , Maximilian Puelma Touzel , William Garneau , Manon Gruaz , Mike Pinder , Li Wei Wang , Sukanya Krishna , Luda Cohen

show 3 more authors

Jean-François Godbout Reihaneh Rabbany Kellin Pelrine

This is my paper · ORCID

classification cs.CLcs.AIcs.HC

keywords fact-checkingmisinformationlargelanguagemodelswebretrievalagentsopen-sourcesoftwarereliabilityscoringtransparentAImedialiteracy

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 presents Veracity, an open-source fact-checking application that combines a large language model with web retrieval agents to assess user-submitted claims. The authors' central claim is that this design produces transparent, grounded veracity assessments: users see a 0-100 reliability score, an explanation of the reasoning, and the sources used. The system is aimed at the general public and at expert users such as journalists, who get an aggregate dashboard. If the system works as described, it would give individuals an accessible alternative to platform moderation, and give researchers a test bed for fact-checking strategies.

What carries the argument

The central mechanism is the LLM/web-search-agent loop. Given a claim, the LLM decides whether retrieval is needed and issues queries through a web agent; retrieved sources are passed to the LLM as evidence, and the user sees the selected sources with credibility ratings. The score-based analysis is the second piece: the LLM must assign a 0-100 reliability score and justify it in text, and the interface maps the score to an actionable message, including a share recommendation when the score exceeds 60 percent.

What would settle it

Run Veracity on a set of claims whose truth has already been established by professional fact-checkers and compare the system's score to the known labels; if the share recommendation (score above 60 percent) does not distinguish true from false claims better than chance, the central claim that the tool provides reliable fact-checking would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that an open-source, production-ready factuality assessment application can be built by teaming an LLM with a web retrieval agent: the LLM decides whether to search, reads documents returned by a web search, and produces a reliability score plus a natural-language explanation, while the user is shown the sources and their average credibility ranking. The authors present this as the first tool of its kind to offer a numerical reliability score with an LLM justification, along with multilingual support, a messaging-app-style interface, and an expert dashboard. As a demo paper, its claim is about the existence and design of the system rather than a measured accuracy result; the force of the claim is that such a tool can be open, transparent, and immediately usable.

Load-bearing premise

The paper's value rests on the assumption that the LLM's reliability score—and the 60 percent cutoff for recommending a share—correctly tracks whether a claim is actually true, since no calibration, benchmark, or evaluation is reported to support it.

Editorial extensions

If this is right

A user can paste a claim and receive a scored, explained verdict without needing to judge the sources on their own.
The open-source code can be downloaded and run locally, so fact-checking does not depend on a single proprietary service.
Experts such as journalists can see aggregate claim trends through the dashboard, which could help them spot emerging misinformation narratives.
Researchers can treat Veracity as a test bed for new retrieval, scoring, and explanation strategies.
Because multilingual support is built in, non-English claims can be checked through the same interface.

Reading between the lines

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

The paper leaves open whether the LLM's 0-100 score is calibrated; a natural extension is to run the open-source system on a labeled benchmark to test whether the 60 percent share threshold separates true from false claims.
Because the interface turns the score into a share recommendation, an uncalibrated score could amplify rather than reduce misinformation at the boundary; a calibration study is the logical next step.
The same architecture could plausibly be extended to image, video, or audio claims by swapping the retrieval agent and using a multimodal LLM, though the paper does not discuss this.
The built-in user feedback mechanism (star ratings plus tags) could be mined as a human calibration signal, turning the app into a data-collection instrument for fact-checking research.

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

Colleague, here's my read on Veracity. The paper is an honest demo paper. The core architecture is directly inherited from Tian et al. 2024 (shared authors), and the paper says so up front. What's actually new is the packaging: the source display panel, the LLM-justified numerical score, and the chat-style interface. That is real, concrete, and useful as a community resource. The system may well work as described; the architecture and screenshots are internally consistent, and the authors should get credit for open-sourcing it and framing it as a test-bed.

The soft spot is exactly what the stress-test highlights. The 60% threshold converts the LLM reliability score into a user-facing share recommendation, and there is no calibration, accuracy measurement, or comparison behind it. The paper repeatedly calls assessments "grounded" and the application "production ready," but offers no evaluation at all. For a demo paper that is tolerable, but those claims need to be softened or backed by at least a sanity-check experiment. Also, the "first tool of its kind" novelty claim is unsupported, and the GitHub links appear only as footnote placeholders, not as actual URLs with commit hashes, which hurts reproducibility.

I don't think this is a load-bearing flaw in the sense that the paper's main deliverable is a working system, not a measured result. The stress-test concern is fair, though: if the share recommendation is meant to be taken seriously, an arbitrary 60% cutoff is not obviously safer than any other threshold. That said, the paper itself does not present the threshold as a validated scientific result; it describes it as a design choice. The real problem is the gap between this modest system description and the confidence in the abstract and conclusion.

Who is this for? People building open-source fact-checking tools, and demo-track reviewers who care about artifact claims. It deserves a serious referee because the artifact is real and the gap between claims and evidence is precisely the kind of thing reviewers should probe. I wouldn't cite it for accuracy results, but I might cite it as an example of an open-source deployment.

Recommendation: send it to peer review, ideally a demo track, with the request that the authors either add a small calibration evaluation or explicitly label the 60% threshold as an arbitrary product choice, not a validated decision rule.

Referee Report

4 major / 4 minor

Summary. The paper presents Veracity, an open-source AI fact-checking system that combines large language models (LLMs) with web retrieval agents to analyze user-submitted claims and provide grounded veracity assessments. The system displays retrieved sources, a numerical reliability score, an explanation of the model's reasoning, and a share recommendation based on a 60% threshold. The paper describes the architecture (frontend/backend, Model-View-Controller), the AI techniques (LLM plus web search, following Tian et al. 2024), and interactive elements including user feedback and an expert dashboard. The central claim is that Veracity is a production-ready, transparent, and accessible fact-checking tool, with open-source repositories available.

Significance. If the system works as described, Veracity could be a valuable open-source contribution to the fact-checking ecosystem, offering a test-bed for research and a practical tool for the public. The paper's strengths are its clear architecture description, its grounding in prior work on retrieval-augmented misinformation detection, and its emphasis on transparency and user-facing explainability. However, the paper provides no quantitative evaluation, calibration, or comparison against existing systems, so the central claims of effectiveness, groundedness, and production readiness are currently unverified. The open-source availability and detailed system description are useful, but the lack of empirical evidence is a significant gap.

major comments (4)

[Section 3, 'Score-based analysis'] The reliability score and the 60% share-recommendation threshold are presented as core innovations, but the paper provides no evidence that the LLM-generated score is calibrated or that the threshold produces correct recommendations. This is load-bearing because the share recommendation is the most consequential user-facing output. Without a calibration study on a standard fact-checking benchmark, the claim of 'grounded veracity assessments' and the production-readiness claim in Section 5 are unsupported.
[Section 2.1, item 5] The choice of 60% as the threshold for a positive share recommendation is unexplained and untested. If the score distribution is miscalibrated, the system may recommend sharing false claims. The authors should either justify the threshold empirically or clearly present it as a configurable placeholder with a planned calibration procedure.
[Section 5, Conclusion] The paper describes Veracity as 'production ready,' but no evaluation of accuracy, latency, robustness, or user experience is provided. For a demo paper, a description of the system is valuable, but the production-readiness claim requires at least some quantitative evidence or a clear statement that this is a prototype.
[Section 2.3] The paper states that the system is an implementation of Tian et al. (2024), but it does not compare Veracity's performance against that baseline or any other misinformation-detection system. Without such a comparison, it is unclear what the system adds beyond the existing method, especially regarding the claimed innovation of a reliability score.

minor comments (4)

[Section 2.2] The frontend and backend wikis are cited as '[link]' with no actual URLs; please provide the links or remove the references.
[Section 3] The claim 'This is the first tool of its kind to present a reliability score' is strong and unsubstantiated; consider softening it or providing a survey of existing tools to support it.
[References] Several references are incomplete or inconsistently formatted, e.g., the Zhou et al. (2024) entry ends with 'arxiv [preprint](2024)' without a full citation.
[Figure 1] The numerical annotations in Figure 1 are referenced in the text, but the figure itself is not included in the text review; ensure the figure is legible and the annotations match the described items.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: Veracity is a demo of an implemented system; the sole self-citation is attribution of the core method to prior work, not a load-bearing circular premise.

full rationale

The paper makes no fitted-parameter-then-prediction move and contains no equation or construction that reduces a claimed output to an input. The reliability score is an LLM output, not a derived quantity, and the 60% share-recommendation threshold is an unexplained design choice, not something the paper derives from data or theory. The only notable self-citation is in Section 3, where the system is described as 'an implementation of the LLM/web search engine teaming proposed by Tian et al. [2024]'. That prior paper shares three co-authors (Godbout, Rabbany, Pelrine) with the present work, and this paper provides no independent evaluation of that method in its own implementation. However, this is attribution and reliance on prior work rather than a logical reduction: the paper's contribution is the open-source application and interface, not a claim that the retriev-augmented architecture is newly derived or validated here. The unvalidated 60% threshold and the absence of calibration or benchmarks are real correctness and usefulness risks, but they are not circularity. The central claim of an open-source, runnable system does not depend on the self-citation for its logical force; it depends on the code and deployment described. Therefore the circularity score is low, reflecting only the minor load-bearing-adjacent self-citation.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central contribution is a system implementation rather than a derivation, so there are no fitted parameters in the mathematical sense. The main hand-chosen number is the 60 percent share threshold. Whether the system actually helps depends on unverified domain assumptions about LLM reliability with retrieved evidence and the validity of external credibility ratings.

free parameters (1)

share recommendation threshold = 60%
The interface states that a score greater than 60 percent maps to a positive share recommendation. The threshold is given without calibration, sensitivity analysis, or evidence about its effect on user decisions or accuracy.

assumptions (2)

domain assumption LLMs combined with retrieved web evidence can detect misinformation effectively.
The paper invokes external citations (Pelrine et al. 2023a; Chen and Shu 2024; Tian et al. 2024) but performs no new evaluation of this assumption for Veracity.
domain assumption Credibility ratings from Lin et al. 2023 are valid for scoring evidence in this system.
The sources panel uses documented credibility ratings, but the paper does not validate that these ratings transfer to the web retrieval setting or to the claims users submit.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Veracity: An Open-Source AI Fact-Checking System." pith.science (2026). https://pith.science/paper/UITHABTV

@misc{pith2026250615794,
  author       = {Pith},
  title        = {Pith review of: Veracity: An Open-Source AI Fact-Checking System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UITHABTV}},
  note         = {Machine review of arXiv:2506.15794}
}

read the original abstract

The proliferation of misinformation poses a significant threat to society, exacerbated by the capabilities of generative AI. This demo paper introduces Veracity, an open-source AI system designed to empower individuals to combat misinformation through transparent and accessible fact-checking. Veracity leverages the synergy between Large Language Models (LLMs) and web retrieval agents to analyze user-submitted claims and provide grounded veracity assessments with intuitive explanations. Key features include multilingual support, numerical scoring of claim veracity, and an interactive interface inspired by familiar messaging applications. This paper will showcase Veracity's ability to not only detect misinformation but also explain its reasoning, fostering media literacy and promoting a more informed society.

Figures

Figures reproduced from arXiv: 2506.15794 by the authors.

**Figure 1.** The main fact-checking page of Veracity open-source and available to anyone. The system is targeted at: • The general public, including both tech-savvy users and those less familiar with new technologies. • Expert users, such as journalists and professional factcheckers, who will have access to an expert dashboard. Our Contribution This paper describes the design and functionality of an open-source, claim-focused f… view at source ↗

**Figure 3.** The interactions between LLM, search engine, and user [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 2.** Veracity app architecture block diagram 3 AI Techniques and Innovations Core AI Methods This system uses AI to power its factchecking methodology, specifically LLM technology. Despite the challenges of misinformation detection, including the tendency of misinformation to contain a mix of both true and false information, LLMs have been shown to be effective tools for detecting misinformation online [Pelrine et al.… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Snippet of the expert dashboard 5 Conclusion This demo has showcased Veracity, an open-source AI system that combines LLMs and web retrieval agents to provide transparent and accessible fact-checking. While AI systems employing LLMs and web retrieval for fact-checking…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 15 canonical work pages

[1]

Factuality challenges in the era of large language models and opportunities for fact-checking.Nature Machine Intelligence, 6(8):852–863, Aug

[Augensteinet al., 2024 ] Isabelle Augenstein, Timothy Baldwin, Meeyoung Cha, Tanmoy Chakraborty, Gio- vanni Luca Ciampaglia, David Corney, Renee DiResta, Emilio Ferrara, Scott Hale, Alon Halevy, Eduard Hovy, Heng Ji, Filippo Menczer, Ruben Miguez, Preslav Nakov, Dietram Scheufele, Shivam Sharma, and Giovanni Zagni. Factuality challenges in the era of lar...

work page 2024
[5]

Chart.js: Simple yet flexible javascript charting library,

[Chart.js Contributors, 2025] Chart.js Contributors. Chart.js: Simple yet flexible javascript charting library,

work page 2025
[6]

Combat- ing misinformation in the age of llms: Opportunities and challenges.AI Magazine, 45(3):354–368,

[Chen and Shu, 2024] Canyu Chen and Kai Shu. Combat- ing misinformation in the age of llms: Opportunities and challenges.AI Magazine, 45(3):354–368,

work page 2024
[7]

Postgresql: The world’s most advanced open source re- lational database,

[Group, 2025] PostgreSQL Global Development Group. Postgresql: The world’s most advanced open source re- lational database,

work page 2025
[8]

Meta is ending its fact-checking program in favor of a ’community notes’ system similar to X’s

[Horvathet al., 2025 ] Bruna Horvath, Jason Abbruzzese, and Ben Goggin. Meta is ending its fact-checking program in favor of a ’community notes’ system similar to X’s. https://www.nbcnews.com/tech/social-media/ meta-ends-fact-checking-program-community-notes\ protect\penalty\z@-x-rcna186468, January

work page 2025
[9]

[Kondamudiet al., 2023 ] Medeswara Rao Kondamudi, Somya Ranjan Sahoo, Lokesh Chouhan, and Nandak- ishor Yadav

[Ac- cessed 03-02-2025]. [Kondamudiet al., 2023 ] Medeswara Rao Kondamudi, Somya Ranjan Sahoo, Lokesh Chouhan, and Nandak- ishor Yadav. A comprehensive survey of fake news in social networks: Attributes, features, and detection approaches.Journal of King Saud University-Computer and Information Sciences, 35(6):101571,

work page 2025
[11]

High level of correspon- dence across different news domain quality rating sets

[Linet al., 2023 ] Hause Lin, Jana Lasser, Stephan Lewandowsky, Rocky Cole, Andrew Gully, David G Rand, and Gordon Pennycook. High level of correspon- dence across different news domain quality rating sets. PNAS Nexus, 2(9):pgad286, 09

work page 2023
[12]

Towards reliable misinformation mitigation: Gen- eralization, uncertainty, and gpt-4.arXiv preprint arXiv:2305.14928,

[Pelrineet al., 2023a ] Kellin Pelrine, Anne Imouza, Camille Thibault, Meilina Reksoprodjo, Caleb Gupta, Joel Christoph, Jean-François Godbout, and Reihaneh Rab- bany. Towards reliable misinformation mitigation: Gen- eralization, uncertainty, and gpt-4.arXiv preprint arXiv:2305.14928,

arXiv

Show all 21 references

[13]

Exploiting novel gpt-4 apis.arXiv preprint arXiv:2312.14302,

[Pelrineet al., 2023b ] Kellin Pelrine, Mohammad Taufeeque, Michał Zaj ˛ ac, Euan McLean, and Adam Gleave. Exploiting novel gpt-4 apis.arXiv preprint arXiv:2312.14302,

arXiv
[14]

Credi- rag: Network-augmented credibility-based retrieval for misinformation detection in reddit.arXiv preprint arXiv:2410.12061,

[Ramet al., 2024 ] Ashwin Ram, Yigit Ege Bayiz, Arash Amini, Mustafa Munir, and Radu Marculescu. Credi- rag: Network-augmented credibility-based retrieval for misinformation detection in reddit.arXiv preprint arXiv:2410.12061,

2024 arXiv
[15]

Fastapi,

[Ramírez, 2025] Sebastián Ramírez. Fastapi,

2025
[16]

Web retrieval agents for evidence-based misinformation detection.arXiv preprint arXiv:2409.00009,

[Tianet al., 2024 ] Jacob-Junqi Tian, Hao Yu, Yury Orlovskiy, Tyler Vergho, Mauricio Rivera, Mayank Goel, Zachary Yang, Jean-Francois Godbout, Reihaneh Rabbany, and Kellin Pelrine. Web retrieval agents for evidence-based misinformation detection.arXiv preprint arXiv:2409.00009,

2024 arXiv
[17]

These are the 3 biggest emerging risks the world is fac- ing

[Torkington, 2024] Simon Torkington. These are the 3 biggest emerging risks the world is fac- ing. https://www.weforum.org/stories/2024/01/ ai-disinformation-global-risks/, January

2024
[18]

[Vercel, 2025] Vercel

[Ac- cessed 02-02-2025]. [Vercel, 2025] Vercel. Next.js,

2025
[19]

Long-form factuality in large language models.arXiv preprint arXiv:2403.18802,

[Weiet al., 2024 ] Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, et al. Long-form factuality in large language models.arXiv preprint arXiv:2403.18802,

2024 arXiv
[20]

A survey of fake news: Fundamental theories, detection methods, and opportunities.ACM Computing Surveys (CSUR), 53(5):1–40,

[Zhou and Zafarani, 2020] Xinyi Zhou and Reza Zafarani. A survey of fake news: Fundamental theories, detection methods, and opportunities.ACM Computing Surveys (CSUR), 53(5):1–40,

2020
[2020]

Correcting misinformation on social media with a large language model

[Zhouet al., 2024 ] Xinyi Zhou, Ashish Sharma, Amy X Zhang, and Tim Althoff. Correcting misinformation on social media with a large language model. arxiv [preprint](2024), 2024

2024
[2021]

Data poisoning in llms: Jailbreak-tuning and scaling laws.arXiv preprint arXiv:2408.02946,

1Links: frontend GitHub repository, backend GitHub repository [Bowenet al., 2024 ] Dillon Bowen, Brendan Murphy, Will Cai, David Khachaturov, Adam Gleave, and Kellin Pel- rine. Data poisoning in llms: Jailbreak-tuning and scaling laws.arXiv preprint arXiv:2408.02946,

2024 arXiv
[2023]

Free speech, fact checking, and the right to accurate information.Sci- ence, 387(6734),

[Lewandowsky, 2025] Stephan Lewandowsky. Free speech, fact checking, and the right to accurate information.Sci- ence, 387(6734),

2025
[2024]

Sqlalchemy: The database toolkit for python,

[Bayer, 2025] Mike Bayer. Sqlalchemy: The database toolkit for python,

2025
[2025]

A review on fact extraction and verification.ACM Computing Surveys (CSUR), 55(1):1–35,

[Bekouliset al., 2021 ] Giannis Bekoulis, Christina Papa- giannopoulou, and Nikos Deligiannis. A review on fact extraction and verification.ACM Computing Surveys (CSUR), 55(1):1–35,

2021

Pith tools

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

[1] [1]

Factuality challenges in the era of large language models and opportunities for fact-checking.Nature Machine Intelligence, 6(8):852–863, Aug

[Augensteinet al., 2024 ] Isabelle Augenstein, Timothy Baldwin, Meeyoung Cha, Tanmoy Chakraborty, Gio- vanni Luca Ciampaglia, David Corney, Renee DiResta, Emilio Ferrara, Scott Hale, Alon Halevy, Eduard Hovy, Heng Ji, Filippo Menczer, Ruben Miguez, Preslav Nakov, Dietram Scheufele, Shivam Sharma, and Giovanni Zagni. Factuality challenges in the era of lar...

work page 2024

[2] [5]

Chart.js: Simple yet flexible javascript charting library,

[Chart.js Contributors, 2025] Chart.js Contributors. Chart.js: Simple yet flexible javascript charting library,

work page 2025

[3] [6]

Combat- ing misinformation in the age of llms: Opportunities and challenges.AI Magazine, 45(3):354–368,

[Chen and Shu, 2024] Canyu Chen and Kai Shu. Combat- ing misinformation in the age of llms: Opportunities and challenges.AI Magazine, 45(3):354–368,

work page 2024

[4] [7]

Postgresql: The world’s most advanced open source re- lational database,

[Group, 2025] PostgreSQL Global Development Group. Postgresql: The world’s most advanced open source re- lational database,

work page 2025

[5] [8]

Meta is ending its fact-checking program in favor of a ’community notes’ system similar to X’s

[Horvathet al., 2025 ] Bruna Horvath, Jason Abbruzzese, and Ben Goggin. Meta is ending its fact-checking program in favor of a ’community notes’ system similar to X’s. https://www.nbcnews.com/tech/social-media/ meta-ends-fact-checking-program-community-notes\ protect\penalty\z@-x-rcna186468, January

work page 2025

[6] [9]

[Kondamudiet al., 2023 ] Medeswara Rao Kondamudi, Somya Ranjan Sahoo, Lokesh Chouhan, and Nandak- ishor Yadav

[Ac- cessed 03-02-2025]. [Kondamudiet al., 2023 ] Medeswara Rao Kondamudi, Somya Ranjan Sahoo, Lokesh Chouhan, and Nandak- ishor Yadav. A comprehensive survey of fake news in social networks: Attributes, features, and detection approaches.Journal of King Saud University-Computer and Information Sciences, 35(6):101571,

work page 2025

[7] [11]

High level of correspon- dence across different news domain quality rating sets

[Linet al., 2023 ] Hause Lin, Jana Lasser, Stephan Lewandowsky, Rocky Cole, Andrew Gully, David G Rand, and Gordon Pennycook. High level of correspon- dence across different news domain quality rating sets. PNAS Nexus, 2(9):pgad286, 09

work page 2023

[8] [12]

Towards reliable misinformation mitigation: Gen- eralization, uncertainty, and gpt-4.arXiv preprint arXiv:2305.14928,

[Pelrineet al., 2023a ] Kellin Pelrine, Anne Imouza, Camille Thibault, Meilina Reksoprodjo, Caleb Gupta, Joel Christoph, Jean-François Godbout, and Reihaneh Rab- bany. Towards reliable misinformation mitigation: Gen- eralization, uncertainty, and gpt-4.arXiv preprint arXiv:2305.14928,

arXiv