Pith. sign in

REVIEW 2 major objections 5 minor 28 references

A hybrid LLM-and-Semantic-Scholar checker flags 94 of 100 confirmed hallucinated citations from NeurIPS 2025, driven mainly by author mismatch.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 15:58 UTC pith:AQ4VY7JY

load-bearing objection Solid external recall eval of an open citation checker on the NeurIPS 2025 fakes; architecture is not new, but the numbers and author-mismatch mechanism are useful and honestly scoped. the 2 major comments →

arxiv 2607.09774 v1 pith:AQ4VY7JY submitted 2026-07-07 cs.DL

Hallucination Detector: A hybrid LLM and Semantic Scholar tool calling for detecting hallucination in scientific literature on AtomGPT.org

Harichandana Neralla , Jaehyung Lee , Aldo H. Romero , Kamal Choudhary This is my paper
classification cs.DL
keywords citation verificationreference hallucinationlarge language modelsSemantic Scholarbibliographic matchingscientific integrityrecord linkageNeurIPS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Fabricated citations generated by large language models are already reaching published literature, including top conferences, because peer review cannot keep up with their volume and surface plausibility. This paper presents an open web tool that parses free-form references with an LLM, retrieves the closest real papers from Semantic Scholar, and scores agreement on title, authors, and venue to grade each citation as trustworthy, partially supported, or likely fabricated. On an external set of 100 verified hallucinated citations that survived NeurIPS 2025 review, the tool flags 94 of them. Detection works because fake references typically pair a credible-looking title with an author list that matches no real paper; title similarity alone would miss most of them. The result matters because automated, interpretable verification at the speed of modern writing is now required to protect the scientific record.

Core claim

On the GPTZero NeurIPS 2025 set of 100 confirmed hallucinated citations, the AtomGPT reference checker classifies 66 as likely hallucinated, 28 as partial, and 6 as verified, for a 94 percent detection rate when both non-verified statuses count as flags. A per-field breakdown shows that near-zero author similarity is the dominant signal even when title similarity is high, while venue similarity contributes little; the few misses are fabrications that either recombine real fields or leave authors unparsed so a strong title alone pushes the composite over the verified threshold.

What carries the argument

The weighted composite score S = 0.60·title similarity + 0.25·author similarity + 0.15·journal similarity (missing fields replaced by the title score), computed against the best Semantic Scholar match after LLM field extraction and then thresholded into verified (≥80), partial (60–79), or likely hallucinated (<60). This score turns free-form citations into graded, field-interpretable judgments grounded in an external scholarly index.

Load-bearing premise

The graded judgment is reliable only if a single live scholarly index plus fixed weights that heavily favor title similarity (and substitute the title score for any missing author field) correctly separate fabrications from real papers.

What would settle it

A balanced test set containing both genuine and fabricated citations that produced a high false-positive rate on real references, or repeated runs on the same 100 known fakes that yielded substantially fewer than 94 detections because of LLM extraction variability, would falsify the reliability claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Editorial and submission systems can insert a lightweight open checker that catches the great majority of fabricated references before publication.
  • Author-list mismatch, not title plausibility, is the practical diagnostic signal for retrieval-grounded citation verification.
  • The tool’s automated status split (66 likely / 28 partial / 6 verified) closely tracks an independent manual taxonomy of the same citations, supporting the claim that most are wholesale inventions.
  • The six misses reveal a concrete failure mode: “Frankenstein” recombinations or unparsed author lists that let a strong title alone exceed the verified threshold.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Re-weighting the composite toward authorship or adding an explicit author-veto rule would likely cut the six false negatives, since three already arise from the missing-field substitution of title for author.
  • A balanced real-plus-fake benchmark is the single most important next measurement; without it the high recall could mask an unacceptable false-positive rate on genuine citations.
  • Cross-checking against multiple scholarly indices would harden the approach against coverage gaps and rate-limit zeros in any single live source.
  • The same author-driven signal could be run retrospectively over large published corpora to quantify how many already-accepted papers contain undetected fabrications.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper presents and evaluates the AtomGPT reference checker, a web tool that parses free-form citations with an open-weight LLM (gpt-oss-20b, with regex fallback), retrieves candidates from Semantic Scholar, and scores title/author/journal agreement via a fixed weighted composite (Eq. 1: 60/25/15) to assign verified / partial / likely-hallucinated labels. Evaluation is a citation-level recall test on the external GPTZero NeurIPS 2025 set of 100 human-verified hallucinated citations: 66 are labeled likely hallucinated, 28 partial, and 6 verified (94% detection when partial is counted as a flag). A per-field decomposition shows near-zero author similarity as the dominant signal despite high title similarity for fabrications; the six misses are analyzed in Table 2 and attributed mainly to the missing-field substitution rule and title-heavy weights. Limitations (positives-only design, LLM stochasticity, single-index dependence, weight misalignment) are stated explicitly, with public code and data.

Significance. Fabricated citations are already entering top-tier proceedings; an open, lightweight, retrieval-grounded checker with an external expert-verified benchmark and a clear mechanistic account (author mismatch as the decisive field) is a useful contribution to research integrity tooling. Strengths include the independent GPTZero ground truth, the alignment with Ansari’s manual taxonomy, the public web tool and evaluation notebook, and the honest scoping as recall-only rather than a precision claim. The work is incremental relative to CheckIfExist but supplies the quantitative external benchmark that system lacked. If the reported detection behavior holds under modest re-weighting and multi-source checks, the tool is a practical drop-in for editorial pipelines.

major comments (2)
  1. The evaluation is correctly scoped as recall on known positives, yet the abstract and conclusion still phrase the result as the tool “reliably flags the great majority” without always restating the operating point (partial counted as detection) or the absence of a false-positive estimate. A short, explicit statement in the abstract and §5 that precision remains unmeasured would prevent over-reading of the 94% figure.
  2. §2 Eq. (1) and the missing-field rule (substitute title score for unextracted author/venue) are load-bearing: Table 2 shows three of the six verified misses arise directly from this rule. Section 4.1 already diagnoses the misalignment and proposes re-weighting or an author veto testable on the existing data. Reporting at least one such re-weighting experiment (or an author-veto ablation) would convert a known failure mode into a quantified improvement and strengthen the central claim without requiring new data.
minor comments (5)
  1. Figure 3 and the status counts are clear; adding the composite-score histogram (or box-plots by band) would make the monotonic separation (means 33.3 / 69.4 / 88.5) immediately visible.
  2. Journal similarity is extractable for only 44/100 citations and does not separate bands; consider demoting or dropping the 15% journal weight in the main reported score, or reporting an ablated composite without it.
  3. The LLM extraction stage is stochastic; a short multi-run stability table (mean ± range of the 66/28/6 split over, e.g., 5 runs) would quantify the reproducibility caveat already noted in §4.1.
  4. Typographical: “Harichandaan” vs “Harichandana” in the acknowledgements/author list; “Frankenstein” references are well described but could be cross-referenced to Ansari’s compound-failure modes for consistency.
  5. Related-work placement of CheckIfExist and CiteAudit is appropriate; a one-sentence explicit statement that the present study supplies the external quantitative benchmark those works lacked would clarify novelty.

Circularity Check

0 steps flagged

No circularity: external recall test on independently curated positives; composite weights and thresholds are fixed a priori, not fitted then re-predicted.

full rationale

The paper’s load-bearing claim is an empirical recall measurement: of 100 GPTZero-verified hallucinated NeurIPS 2025 citations, the AtomGPT checker flags 94 (66 likely hallucinated + 28 partial), with a per-field decomposition showing author mismatch as the dominant signal. The ground-truth labels come from an external, human-verified set (GPTZero + Ansari’s independent taxonomy), not from the authors’ own judgments or from any quantity defined in terms of the tool’s outputs. The composite S = 0.60 st + 0.25 sa + 0.15 sj and the status thresholds (≥80 verified, 60–79 partial, <60 likely hallucinated) are stated as fixed design choices and applied once; they are not fitted to the benchmark and then re-reported as predictions. Self-citations ([11–16]) point to the AtomGPT platform that hosts the tool under test; they do not supply the fabrication labels, uniqueness theorems, or any ansatz that forces the reported detection rate. Limitations (positives-only design, stochastic LLM extraction, title-heavy weights, missing-field substitution) are acknowledged in Section 4.1 and constrain precision and production readiness, but they do not make any reported number equivalent to its inputs by construction. The evaluation is therefore self-contained against an external benchmark and exhibits no circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central recall claim rests on three hand-chosen scoring parameters, the assumption that Semantic Scholar is a sufficiently complete and stable index for the purpose, and the design rule that missing fields are replaced by the title score rather than penalized. No new physical or mathematical entities are postulated; the free parameters are the only quantities introduced by the authors that directly affect the reported detection rate.

free parameters (4)
  • title weight in composite = 0.60
    Fixed at 0.60 in Eq. 1; the authors later note this overweighting of the least discriminative field contributes to three of the six false negatives.
  • author weight in composite = 0.25
    Fixed at 0.25 in Eq. 1; per-field analysis shows authorship is actually the strongest signal, so the weight is misaligned with the data.
  • journal weight in composite = 0.15
    Fixed at 0.15 in Eq. 1; journal was extractable for only 44/100 citations and did not separate status bands.
  • verified / partial thresholds = 80 / 60
    Composite cut-offs of 80 and 60 that define the three status labels; chosen by the tool designers and not fitted to the NeurIPS set, but they determine the 94 % figure.
axioms (3)
  • domain assumption Semantic Scholar's top-k paper-search API returns a sufficiently complete and correct set of candidate real papers for matching.
    Invoked throughout Section 2; a genuine paper absent from the index or throttled by rate limits yields a zero score and would be counted as hallucinated.
  • ad hoc to paper When an author or venue field is missing from extraction, substituting the title score (rather than zero or a penalty) is an acceptable design choice.
    Explicitly stated in Section 2 after Eq. 1; this rule is responsible for two of the six verified false negatives.
  • ad hoc to paper Treating both 'likely hallucinated' and 'partial' as positive detections is a valid operating point for measuring recall.
    Stated in Section 3; the stricter 66 % figure that counts only 'likely hallucinated' is also reported, so the choice is transparent but still load-bearing for the headline 94 % claim.

pith-pipeline@v1.1.0-grok45 · 15566 in / 3151 out tokens · 32736 ms · 2026-07-14T15:58:41.244459+00:00 · methodology

0 comments
read the original abstract

Large language models are now commonly used as partners in scientific writing, and this shift has brought a subtler type of failure: made-up references. Fabricated authors, bogus DOIs, wrongly assigned identifiers, and citations that merge elements from multiple genuine articles are now being inserted into manuscripts at a volume that traditional peer review was never meant to handle. Recent audits reveal that such references have already slipped through the review process and made their way into the published literature, including leading journals and conferences. Automated verification that operates at the speed and scale of modern content production has therefore become a necessary safeguard rather than a convenience. This work presents and evaluates the AtomGPT reference checker (https://atomgpt.org/hallucination_detector), an open, web-accessible tool that verifies citations against the scholarly literature by combining large-language-model field extraction with structured retrieval from Semantic Scholar. For each reference, the tool extracts the bibliographic fields, retrieves the closest matching real papers, and scores the agreement across title, authorship, and venue to produce a graded judgment of whether a citation is trustworthy, partially supported, or likely fabricated. We benchmark the tool against an externally curated set of confirmed hallucinated citations from accepted NeurIPS 2025 papers and find that it reliably flags the great majority of them.

Figures

Figures reproduced from arXiv: 2607.09774 by Aldo H. Romero, Harichandana Neralla, Jaehyung Lee, Kamal Choudhary.

Figure 1
Figure 1. Figure 1: Evaluation pipeline. The GPTZero NeurIPS 2025 benchmark supplies 100 confirmed hallu [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Snapshot of the AtomGPT reference-checking web interface showing a representative result: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: AtomGPT detection status across the 100 known hallucinated citations from the GPTZero [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-field similarity score distributions across the scored citations. Each panel shows one component of the composite score, reported over the subset of citations for which that field was extracted, so a low score always reflects a genuine mismatch rather than a missing field. (a) Title similarity over the 100 scored citations (mean 62.4): despite every citation being fabricated, title scores cluster high … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 4 linked inside Pith

  1. [1]

    Walters and Esther I

    William H. Walters and Esther I. Wilder. Fabrication and errors in the bibliographic citations generated by ChatGPT.Scientific Reports, 13(1):14045, 2023

    2023

  2. [2]

    Miller, Debjani Bhattacharyya, and Larry E

    Mehul Bhattacharyya, Valerie M. Miller, Debjani Bhattacharyya, and Larry E. Miller. High rates of fabricated and inaccurate references in ChatGPT-generated medical content.Cureus, 15(5):e39238, 2023

    2023

  3. [3]

    Sai Anirudh Athaluri, Sandeep Varma Manthena, V. S. R. Krishna Manoj Kesapragada, Vineel Yarlagadda, Tirth Dave, and Rama Tulasi Siri Duddumpudi. Exploring the boundaries of reality: investigating the phenomenon of artificial intelligence hallucination in scientific writing through ChatGPT references.Cureus, 15(4):e37432, 2023

    2023

  4. [4]

    A preliminary investigation of fake peer-reviewed citations and references generated by ChatGPT.The Professional Geographer, 75(6):1024–1027, 2023

    Terence Day. A preliminary investigation of fake peer-reviewed citations and references generated by ChatGPT.The Professional Geographer, 75(6):1024–1027, 2023

    2023

  5. [5]

    Matthew Dahl, Varun Magesh, Mirac Suzgun, and Daniel E. Ho. Large legal fictions: profiling legal hallucinations in large language models.Journal of Legal Analysis, 16(1):64–93, 2024

    2024

  6. [6]

    Manning, and Daniel E

    Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, and Daniel E. Ho. Hallucination-free? assessing the reliability of leading AI legal research tools.Journal of Empirical Legal Studies, 2025

    2025

  7. [7]

    Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews: comparative analysis.Journal of Medical Internet Research, 26:e53164, 2024

    Mikaël Chelli, Jules Descamps, Vincent Lavoué, Christophe Trojani, Michel Azar, Marcel Deckert, Jean-Luc Raynier, Gilles Clowez, Pascal Boileau, and Caroline Ruetsch-Chelli. Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews: comparative analysis.Journal of Medical Internet Research, 26:e53164, 2024

    2024

  8. [8]

    Fabricated citations: an audit across 2.5 million biomedical papers.The Lancet, 407:1779–1781, 2026

    Maxim Topaz, Nir Roguin, Pallavi Gupta, Zhihong Zhang, and Laura-Maria Peltonen. Fabricated citations: an audit across 2.5 million biomedical papers.The Lancet, 407:1779–1781, 2026

    2026

  9. [9]

    NeurIPS, one of the world’s top AI conferences, accepted research papers with 100+ AI-hallucinated citations.https://gptzero.me/news/neurips/, 2026

    GPTZero. NeurIPS, one of the world’s top AI conferences, accepted research papers with 100+ AI-hallucinated citations.https://gptzero.me/news/neurips/, 2026. Accessed 2026

    2026

  10. [10]

    Compound deception in elite peer review: A failure mode taxonomy of 100 fabricated citations at NeurIPS 2025, 2026

    Mohammad Samar Ansari. Compound deception in elite peer review: A failure mode taxonomy of 100 fabricated citations at NeurIPS 2025, 2026. arXiv:2602.05930. 11

    arXiv 2025

  11. [11]

    AtomGPT: Atomistic generative pretrained transformer for forward and inverse materials design.The Journal of Physical Chemistry Letters, 15(27):6909–6917, 2024

    Kamal Choudhary. AtomGPT: Atomistic generative pretrained transformer for forward and inverse materials design.The Journal of Physical Chemistry Letters, 15(27):6909–6917, 2024

    2024

  12. [12]

    AGAPI-agents: An open-access agentic AI platform for accelerated materials design on Atom- GPT.org.The Journal of Physical Chemistry Letters, 2026

    Jaehyung Lee, Justin Ely, Kent Zhang, Akshaya Ajith, Charles Rhys Campbell, and Kamal Choud- hary. AGAPI-agents: An open-access agentic AI platform for accelerated materials design on Atom- GPT.org.The Journal of Physical Chemistry Letters, 2026

    2026

  13. [13]

    Kamal Choudhary. DiffractGPT: Atomic structure determination from X-ray diffraction patterns using a generative pretrained transformer.The Journal of Physical Chemistry Letters, 16(8):2110– 2119, 2025

    2025

  14. [14]

    Kamal Choudhary. MicroscopyGPT: Generating atomic-structure captions from microscopy images of 2D materials with vision-language transformers.The Journal of Physical Chemistry Letters, 16(27):7028–7035, 2025

    2025

  15. [15]

    The JARVIS infrastructure is all you need for materials design.Computational Materials Science, 259:114063, 2025

    Kamal Choudhary. The JARVIS infrastructure is all you need for materials design.Computational Materials Science, 259:114063, 2025

    2025

  16. [16]

    Romero, Kamal Choudhary, et al

    Jaehyung Lee, Harichandaan Neralla, Charles Campbell, Kent Zhang, Akshaya Ajith, Justin Ely, Aldo H. Romero, Kamal Choudhary, et al. Lessons learned from the 2025 agentic AI for science hackathon.Machine Learning: Science and Technology, 2026

    2025

  17. [17]

    The Se- mantic Scholar open data platform.arXiv preprint arXiv:2301.10140, 2023

    Rodney Kinney, Chloe Anastasiades, Russell Authur, Iz Beltagy, Jonathan Bragg, et al. The Se- mantic Scholar open data platform.arXiv preprint arXiv:2301.10140, 2023. arXiv:2301.10140

    Pith/arXiv arXiv 2023

  18. [18]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qiang- long Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.arXiv preprint arXiv:2311.05232, 2023. arXiv:2311.05232

    Pith/arXiv arXiv 2023

  19. [19]

    Hal- luMat: Detecting hallucinations in LLM-generated materials science content through multi-stage verification, 2025

    Bhanu Prakash Vangala, Sajid Mahmud, Pawan Neupane, Joel Selvaraj, and Jianlin Cheng. Hal- luMat: Detecting hallucinations in LLM-generated materials science content through multi-stage verification, 2025. arXiv:2512.22396

    arXiv 2025

  20. [20]

    Do language models know when they’re hallucinating references? InFindings of the Association for Computational Linguistics: EACL 2024, pages 912–928, 2024

    Ayush Agrawal, Mirac Suzgun, Lester Mackey, and Adam Tauman Kalai. Do language models know when they’re hallucinating references? InFindings of the Association for Computational Linguistics: EACL 2024, pages 912–928, 2024

    2024

  21. [21]

    Mesallam, Mohamed Farahat, and Khalid H

    FadiAljamaan, Mohamad-HaniTemsah, IbraheemAltamimi, AymanAl-Eyadhy, AmrJamal, Khalid Alhasan, Tamer A. Mesallam, Mohamed Farahat, and Khalid H. Malki. Reference hallucination score for medical artificial intelligence chatbots: development and usability study.JMIR Medical Informatics, 12:e54345, 2024

    2024

  22. [22]

    CheckIfExist: DetectingcitationhallucinationsintheeraofAI-generatedcontent,

    DilettaAbbonato. CheckIfExist: DetectingcitationhallucinationsintheeraofAI-generatedcontent,

  23. [23]

    Chawla, and Yanfang Ye

    Zhengqing Yuan, Kexin Shi, Zhaonan Zhang, Lichao Sun, Nitesh V. Chawla, and Yanfang Ye. CiteAudit: You cited it, but did you read it? a benchmark for verifying scientific references in the LLM era, 2026. arXiv:2602.23452

    Pith/arXiv arXiv 2026

  24. [24]

    Fellegi and Alan B

    Ivan P. Fellegi and Alan B. Sunter. A theory for record linkage.Journal of the American Statistical Association, 64(328):1183–1210, 1969

    1969

  25. [25]

    Mikhail Bilenko and Raymond J. Mooney. Adaptive duplicate detection using learnable string simi- larity measures. InProceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 39–48, 2003

    2003

  26. [26]

    arXiv API access.https://info.arxiv.org/help/api/index.html

    arXiv. arXiv API access.https://info.arxiv.org/help/api/index.html. Accessed 2026

    2026

  27. [27]

    Crossref REST API.https://www.crossref.org/documentation/retrieve-metadata/ rest-api/

    Crossref. Crossref REST API.https://www.crossref.org/documentation/retrieve-metadata/ rest-api/. Accessed 2026

    2026

  28. [28]

    gpt-oss-120b & gpt-oss-20b model card, 2025

    OpenAI. gpt-oss-120b & gpt-oss-20b model card, 2025. arXiv:2508.10925. 12

    Pith/arXiv arXiv 2025