Pith. sign in

REVIEW 2 major objections 3 minor 68 references

This paper claims that for LLM citation verifiers, false-positive rate—not recall—determines whether a tool is deployable, and backs the claim with a 2,526-entry benchmark that exposes three concrete failure modes.

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 · deepseek-v4-flash

2026-08-01 16:35 UTC pith:TZADY4KH

load-bearing objection A genuinely useful, openly shipped benchmark whose headline FPR-deployability quantification rests on two unrecoverable Anthropic operating points and an untested synthetic-to-real transfer. the 2 major comments →

arxiv 2607.18360 v1 pith:TZADY4KH submitted 2026-07-20 cs.CR cs.AIcs.LG

HALLMARK: Diagnosing Three Failure Modes in LLM Citation Verifiers

Patrik Reizinger , Wieland Brendel This is my paper
classification cs.CR cs.AIcs.LG
keywords citation hallucinationLLM verifierfalse positive ratebenchmarkprecision at low prevalencetraining cutoffBibTeXhallucination taxonomy
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.

The paper sets out to answer which citation-verification tools can be trusted when a venue audits its submissions for hallucinated references. It builds HALLMARK, a controlled benchmark of 2,526 BibTeX entries spanning 14 hallucination types, three difficulty tiers, and six diagnostic sub-tests, then runs thirteen verifiers under one protocol. Its central claim is that the false-positive rate, not recall, decides whether a verifier is usable: at a venue-realistic ~2% hallucination rate, verifiers span an order of magnitude in FPR, producing a ~7x precision gap between the best and most aggressive tools. The paper also diagnoses three failure modes: agentic lookups buy recall but inflate false positives, precision collapses at low base rates, and most LLMs over-flag papers published past their training cutoff. If correct, the paper redirects verifier design and procurement toward FPR and calibration rather than raw detection rate.

Core claim

Across the benchmark, measured false-positive rates span roughly 0.05 to 0.70 across verifiers, while detection rates all sit in a 48–91% band. At a venue-realistic 2% base rate, Bayes' rule turns that FPR spread into a ~7x precision gap: the lowest-FPR tools catch one true hallucination per 6–9 flags, whereas the most aggressive high-recall models catch fewer than one per 35 flags even at ≥87% detection. The paper argues this is a deployment law, not a curiosity: for a reviewer-bound venue audit, a few points of FPR decide whether a tool's flags are mostly true catches or mostly noise. It also isolates the mechanism behind the failure modes—agentic harnesses flag on any single missing datab

What carries the argument

The central object is HALLMARK, a benchmark whose load-bearing design combines a 14-type hallucination taxonomy, six diagnostic binary sub-tests per entry, and a contamination-resistant held-out split, so that each verifier is scored on prevalence-independent metrics (detection rate, false-positive rate, MCC). The mechanism that carries the argument is the closed-form relationship PPV = DR·prev / (DR·prev + FPR·(1−prev)), which turns measured FPR differences into precision gaps at deployment prevalence; against that identity the paper evaluates zero-shot LLMs, agentic tool-calling harnesses, a DOI-lookup baseline, and a co-designed rule-based verifier that pairs consensus-absence flagging wi

Load-bearing premise

The benchmark's synthetic and perturbed hallucination entries faithfully represent real-world hallucinations, so the measured false-positive rates and rankings transfer to deployment—an assumption the paper itself says remains untested at sample sizes of only 108 real-world entries.

What would settle it

Curate a few thousand hand-verified, real hallucinated citations from diverse venues, run the same verifier cohort under the same protocol, and check at a 2% base rate whether the precision spread (roughly 1-in-6 to 1-in-35 flags) and the FPR-based ranking reproduce on this authentic corpus; if low-FPR tools no longer separate from high-recall tools, the central claim fails.

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

If this is right

  • At low-prevalence venue audits, verifier selection should be driven by false-positive rate and calibration, not recall, because precision is bottlenecked by FPR.
  • Agentic retrieval harnesses should flag only on consensus absence across databases (plus positive metadata contradictions), not on any single missing record, to avoid a ~15x FPR inflation.
  • Most LLM verifiers degrade sharply on papers published after their training cutoff, so deployment on recent literature requires cutoff-aware prompting or retrieval augmentation, and the mitigation is model-dependent.
  • When missed fabrications are far costlier than triaging extra flags, or when downstream human review absorbs false alarms, high-recall tools plus human triage remain the right choice—the FPR rule is regime-conditional, not absolute.
  • A two-stage cascade (rule-based stage plus LLM diagnoser) can reach detection rates near 0.99 at an FPR near 0.11, providing a concrete high-recall, low-FPR reference design.

Where Pith is reading between the lines

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

  • If the FPR-ranking generalizes beyond synthetic entries, leaderboards that sort by recall or F1 mislead deployment decisions; the paper's logic implies that any verifier benchmark should report PPV at venue-realistic prevalence and precision-oriented metrics like FPR at 90% coverage.
  • The post-cutoff over-flagging result suggests a testable product implication: verifier trust expires, so tools could expose their training cutoff, route post-cutoff entries to retrieval, or watermark uncertainty—and models trained with abstention-oriented post-training may hold up better regardless of cutoff recency.
  • A natural extension would be a prospective live audit of a single conference's full submission pool, comparing each verifier's flags against author confirmations, to test whether the measured 1-in-6 to 1-in-35 precision spread reproduces on real traffic.
  • The benchmark's perturbation-heavy corpus is easier than LLM-generated hallucinations for at least one verifier (detection 0.846 vs 0.656), so building a much larger hand-validated real-world corpus is the key step to translating the paper's rankings into a deployment guarantee.

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 / 3 minor

Summary. This paper introduces HALLMARK, a benchmark of 2,526 BibTeX entries spanning 14 hallucination types across three difficulty tiers, with six diagnostic sub-tests per entry and a contamination-resistant split. Thirteen citation verifiers are evaluated under one protocol: a DOI-only baseline, twelve zero-shot LLMs, three agentic harnesses, and the authors' rule-based bibtex-updater. The central claim is that false-positive rate (FPR), not recall, governs deployability: at a venue-realistic ~2% base rate, Bayes' rule makes PPV depend on FPR, and the measured FPR spread (0.05–0.70) produces a reported ~7x precision gap, with low-FPR tools catching one true hallucination per 6–9 flags and aggressive high-recall tools falling below one per 35. Three failure modes are diagnosed: agentic lookups inflate FPR; base-rate precision drops; and most LLMs over-flag post-cutoff papers. The paper is unusually transparent about limitations, including the synthetic–real gap, the co-designed status of bibtex-updater, and endpoint drift on two Anthropic rows.

Significance. The benchmark itself is a valuable shared resource: the code and pinned dataset are released, per-entry verdicts are provided for bibtex-updater, bootstrap CIs accompany most aggregates, and the contamination canary and adaptive-data-analysis hygiene measures are thoughtful. The PPV analysis is a clean, correct application of Bayes' rule once DR and FPR are measured, and the qualitative finding that FPR ordering, not recall, dominates precision at low prevalence is useful for practitioners. However, the headline quantitative claims rest on two load-bearing measurements that are not reproducible (the Anthropic dev_public rows), and the deployment framing assumes that FPRs measured on perturbed entries transfer to real hallucinations—an assumption the paper itself states is untested. The benchmark retains substantial value, but the specific 'one-in-6-to-9' and '~7x gap' numbers should be treated as provisional.

major comments (2)
  1. [§7, §E.2, Tab. 3, Tab. 21] The two Anthropic dev_public rows (Opus 4.7 FPR 0.072, Sonnet 4.6 FPR 0.127) are summary-only, with no stored per-entry predictions, and §7 reports that a later OpenRouter snapshot roughly doubles both (0.162 and 0.165). These two rows anchor the low-FPR frontier that drives the abstract's and §6's quantitative claims: Tab. 21 gives Opus PPV 17.6% and Sonnet 11.2% at a 2% base rate, leading to 'one true hallucination per 6–9 flags' and the '~7x precision gap.' Using the drift-adjusted FPRs, Opus PPV falls to about 8.7% and the gap to DeepSeek-V3.2's 2.6% shrinks to roughly 3x. The paper's decision to report the internally consistent snapshot and to read rankings rather than absolute levels is defensible for the ordinal claim, but the abstract and §6 present these absolute PPV numbers unconditionally. This is load-bearing for the central claim and needs to be fixed, either by basing the h
  2. [§6, §7, Tab. 16, §C.5] The deployment claim assumes that FPR/DR measured on synthetic perturbed entries transfer to real hallucinations. The paper honestly states in §7 that 'distributional and semantic equivalence remains untested' and that 'whether rankings on synthetic hallucinations predict performance on real errors remains open.' The real-world anchor is only 108 entries, 55% of which are plausible_fabrication, and the KS-based equivalence tests are underpowered. In-paper evidence cuts against transfer: GPT-5.1's DR on LLM-generated entries is 0.656 vs 0.846 on perturbations (Tab. 16), and the cross-domain probe shows bibtex-updater's FPR rising from 0.092 to 0.375. Since FPR is the deployment-decisive metric, even a modest real-world FPR shift changes the 'one-in-N flags' numbers and could reorder the precision ranking. The Walters–Wilder supplement (§C.5) is a useful partial check, but it is articles-o
minor comments (3)
  1. [§5.2] 'the F1 score (F1) gap' is redundant; should be 'the F1 gap.'
  2. [§3.3] 'The three sources play complementary roles ... then stratified by generation method lets us assess' is a sentence fragment; consider splitting into two sentences.
  3. [§7] Missing space in 'Opus 4.7dev_public'; also consider defining 'pp' at first use for readers outside the subfield.

Circularity Check

1 steps flagged

No material circularity: FPR/DR are measured independently and PPV is closed-form Bayes; the only co-design overlap (bibtex-updater) is disclosed, excluded from ranking, and non-load-bearing.

specific steps
  1. other [§ 5.1 and § G.2]
    "We label bibtex-updater co-designed because its development overlapped the taxonomy: the typed sub-tests partly mirror the tool's verification stages, so it may score better here than on a novel hallucination distribution. We therefore read its row as an upper-bound reference, excluded from ranking (§ G.2)."

    The tool's verification stages and the benchmark's typed sub-tests were co-developed, so bibtex-updater's scores on those sub-tests are partly self-consistent by construction rather than independent measurements. The paper neutralizes this by excluding the tool from rankings and labeling it an upper-bound reference; no conclusion about FPR-vs-recall depends on its row, so this is a minor, disclosed circularity rather than a load-bearing one.

full rationale

The central deployability claim is a measurement-plus-arithmetic claim, not a fitted prediction. The paper measures DR and FPR for independent tools under one protocol (Tab. 3), then computes PPV = DR·prev / (DR·prev + FPR·(1−prev)) at an assumed 2% base rate (§ E.1), explicitly stating that the PPV-precision drop 'follows directly from Bayes’ rule ... and is no discovery of ours.' No parameter is fit to the target conclusion, and no tool's output is defined in terms of the ranking it is used to support. The observed FPR spread is an empirical finding. The endpoint-drift caveat (§ 7: Opus 0.072→0.162, Sonnet 0.127→0.165) and the untested synthetic-real equivalence (§§ 6–7: 'distributional and semantic equivalence remains untested at these sample sizes'; 'Whether rankings on synthetic hallucinations predict performance on real errors remains open') are external-validity and reproducibility limitations, not circular derivations. The only self-referential element—the authors' co-designed bibtex-updater—is explicitly labeled co-designed and excluded from ranking, so the deployability claim does not reduce to it. No self-citation chain, uniqueness theorem, or ansatz is load-bearing. Overall: no significant circularity; at most a minor, disclosed design-overlap caveat.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

This is an empirical benchmark, not a derivation, so the axiom ledger is short. The main hand-chosen quantities are the assumed base rate and metric weights. The scientific contribution rests on two domain assumptions—database-backed ground truth and synthetic-real representativeness—the second explicitly acknowledged as untested.

free parameters (3)
  • assumed entry-level hallucination prevalence (π) = 0.02
    Used in the PPV analysis to translate measured FPR/DR into precision. The authors call it an estimate and sweep 1–5%; the FPR ranking is prevalence-invariant, but absolute PPV depends on this choice.
  • tier weights for TW-F1 = {1,2,3}
    Hand-chosen linear weighting for tier-weighted F1; robustness is checked across four other weighting schemes, so the ranking is stable but the headline metric is a design choice.
  • title-similarity threshold in consensus aggregation proxy = ≥0.85 token-sort ratio
    Deterministic matcher used to isolate any-no-match vs consensus aggregation. The authors caution to read the ordering, not the absolute level.
axioms (4)
  • domain assumption DBLP, CrossRef, Semantic Scholar, OpenAlex, and arXiv records are trusted ground truth.
    Valid entries are verified against these databases and hallucination labels/sub-tests are derived from them (§3.3, §A.3). If these databases are incomplete or wrong, the labels, DR, and FPR are systematically wrong.
  • domain assumption Synthetic perturbation entries are representative of real-world hallucinations.
    Most hallucinated entries are perturbation-generated; only 108 real-world entries exist, with no human IAA. The authors explicitly leave equivalence untested (§6, §7). This is the paper's weakest load-bearing premise.
  • domain assumption The 'FPR decides deployability' claim is restricted to a low-prevalence, reviewer-bound regime.
    The authors state the framing is calibrated to ~2% prevalence, finite reviewer attention, and cheap false alarms; under high prevalence or costly misses, recall plus human triage dominates (Tab. 6, §7).
  • standard math Bayes' rule and the PPV formula.
    Eq. (3) in §E.1 is standard probability; no controversy.

pith-pipeline@v1.3.0-alltime-deepseek · 57405 in / 10454 out tokens · 94084 ms · 2026-08-01T16:35:23.662108+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) now routinely draft literature reviews and assist with academic writing, which means a higher risk of fabricated references: GPTZero found 53 papers with hallucinated citations among NeurIPS 2025's accepted set. Rule- and LLM-based verifiers are emerging, but no shared benchmark compares them and gives detailed failure diagnostics. We close that gap with HALLMARK (Hallucination benchmark): 2,526 BibTeX entries spanning 14 hallucination types, three difficulty tiers, six diagnostic sub-tests per entry, and a contamination-resistant held-out split. On it we evaluate a DOI-lookup baseline, frontier LLMs zero-shot, tool-augmented agents, and our own rule-based, co-designed verifier bibtex-updater. Across the benchmark one result is consistent: the false-positive rate, not recall, decides whether a verifier is deployable. HALLMARK makes it concrete through three failure modes: agentic lookups buy recall but inflate false positives; at a venue-realistic base rate, the order-of-magnitude spread in false-positive rates (FPRs) -- not recall -- governs whether a verifier's flags are mostly true catches or mostly noise; and most LLMs over-flag papers published past their training cutoff, where only the two latest-cutoff models hold their false-positive rate near in-distribution levels (a signal we report as descriptive, since it is confounded with possible recall of those entries). Thus FPR is the deployment bottleneck, but an undetected fabrication remains the costlier error for the scientific record.

Figures

Figures reproduced from arXiv: 2607.18360 by Patrik Reizinger, Wieland Brendel.

Figure 1
Figure 1. Figure 1: Overview of HALLMARK. Top: The benchmark pipeline: real papers sourced from DBLP are transformed via perturbation, LLM generation, and real-world collection into 2,526 annotated BibTeX entries, each with six diagnostic sub-tests. Thirteen full-coverage verification tools (1 citation￾database, 12 zero-shot LLMs) are evaluated using tier-weighted metrics. Bottom left: The three-tier hallucination taxonomy wi… view at source ↗
Figure 2
Figure 2. Figure 2: DR–FPR Pareto frontier on dev_public. Each point is a (tool, configuration) pair. The dotted line traces the Pareto front (DR↑, FPR↓). Independent zero-shot LLMs occupy the precision￾end of the front (Sonnet 4.6, Opus 4.7, Gemini 2.5 Pro); the recall-end is occupied by recall-aggressive open-weight models (DeepSeek-V3.2, Qwen3-VL-235B). Agentic harnesses sit above the LLM zero-shot points on DR but to the … view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation pipeline. Each citation entry passes through a shared pre-screening layer (DOI format, year bounds, name heuristics) that may emit a final verdict directly; otherwise the entry is forwarded to the tool. Metrics aggregate over all entries, including pre-screening overrides. The shared layer means every tool’s reported FPR includes pre-screening false positives; see Tab. 3. 2. Year bounds checking… view at source ↗
Figure 4
Figure 4. Figure 4: Per-type detection rates across all full-coverage tools. HaRC and verify-citations are excluded due to <7% effective coverage (§ B.2.2). author_mismatch (shown in the heatmap under its enum value “swapped authors”) and near_miss_title remain the hardest types across the cohort. 0.190, Tab. 3), though far from calibrated. For deployment pipelines that triage flags by confidence, the two open-weight models’ … view at source ↗
Figure 5
Figure 5. Figure 5: Cost–accuracy tradeoff. All sixteen evaluated tools with recorded throughput are shown; rate-limited tools (HaRC, verify-citations) are plotted at their sub-7%-coverage operating points (§ B.2.2). Rate-limited tools are impractical for venue-scale deployment. Per-entry cost is the dominant feasibility constraint at low prevalence (∼2%): even highly accurate verifiers misallocate reviewer effort because pre… view at source ↗
Figure 6
Figure 6. Figure 6: Detection rate by difficulty tier across all full-coverage tools. DOI-only concentrates in Tier 1; every LLM achieves broad coverage with graceful degradation as difficulty rises [PITH_FULL_IMAGE:figures/full_fig_p039_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Thinking-budget regime boundary. F1 vs. saturation ratio p95/cap for each (model, budget) cell, where p95 is the 95th-percentile output token count and cap is the configured max_completion_tokens; saturation → 1 means the reasoning trace is hitting the budget ceil￾ing and the structured output is likely truncated. Arrows trace each model’s trajectory from low to high budget; marker size is proportional to … 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

68 extracted references · 2 canonical work pages

  1. [1]

    CheckIfExist: Detecting citation hallucinations in the era of AI-generated content

    Diletta Abbonato. CheckIfExist: Detecting citation hallucinations in the era of AI-generated content. arXiv preprint arXiv:2602.15871, 2026

    arXiv 2026

  2. [2]

    Do language models know when they’re hallucinating references? InEACL, pages 912–928

    Ayush Agrawal, Mirac Suzgun, Lester Mackey, and Adam Kalai. Do language models know when they’re hallucinating references? InEACL, pages 912–928. Association for Computational Linguistics, 2024. doi: 10.18653/v1/2024.findings-eacl.62. URL https://aclanthology. org/2024.findings-eacl.62

    work page doi:10.18653/v1/2024.findings-eacl.62 2024

  3. [3]

    McFarlane

    Hussam Alkaissi and Samy I. McFarlane. Artificial hallucinations in ChatGPT: Implications in scientific writing.Cureus, 15(2), 2023. doi: 10.7759/cureus.35179. URL https://doi.org/ 10.7759/cureus.35179

    doi:10.7759/cureus.35179 2023

  4. [4]

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

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

    arXiv 2025

  5. [5]

    The case of the mysterious citations.arXiv preprint arXiv:2602.05867, 2026

    Amanda Bienz, Carl Pearson, and Simon Garcia de Gonzalo. The case of the mysterious citations.arXiv preprint arXiv:2602.05867, 2026

    Pith/arXiv arXiv 2026

  6. [6]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Pith/arXiv arXiv 2021

  7. [7]

    dblp computer science bibliography — monthly snapshot XML release of october 2025, 2025

    dblp Team. dblp computer science bibliography — monthly snapshot XML release of october 2025, 2025. URLhttps://dblp.org

    2025

  8. [8]

    DeepSeek-V3 technical report.arXiv preprint arXiv:2412.19437, 2024

    DeepSeek-AI et al. DeepSeek-V3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    The reusable holdout: Preserving validity in adaptive data analysis.Science, 349(6248):636–638,

    Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Aaron Roth. The reusable holdout: Preserving validity in adaptive data analysis.Science, 349(6248):636–638,

  10. [10]

    Tibshirani.An Introduction to the Bootstrap

    Bradley Efron and Robert J. Tibshirani.An Introduction to the Bootstrap. Chapman & Hall/CRC, 1994

    1994

  11. [11]

    On the foundations of noise-free selective classification

    Ran El-Yaniv and Yair Wiener. On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11:1605–1641, 2010. URLhttps://www.jmlr.org/ papers/v11/el-yaniv10a.html

    2010

  12. [12]

    Hal- luHard: A hard multi-turn hallucination benchmark

    Dongyang Fan, Sebastien Delsad, Nicolas Flammarion, and Maksym Andriushchenko. Hal- luHard: A hard multi-turn hallucination benchmark. arXiv:2602.01031, 2026. URL https://arxiv.org/abs/2602.01031

    arXiv 2026

  13. [13]

    Joseph L. Fleiss. Measuring nominal scale agreement among many raters.Psychological Bulletin, 76(5):378–382, 1971. doi: 10.1037/h0031619

    doi:10.1037/h0031619 1971

  14. [14]

    RARR: Researching and revising what language models say, using language models

    Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. RARR: Researching and revising what language models say, using language models. InAnnual Meeting of the Association for Computational Linguistics, pages 16477–16508. Association for Computational Lingu...

    doi:10.18653/v1/2023.acl-long.910 2023

  15. [15]

    Datasheets for datasets.Communications of the ACM, 64(12):86–92, 2021

    Timnit Gebru, Jamie Morgenstern, Brenda Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets.Communications of the ACM, 64(12):86–92, 2021. doi: 10.1145/3458723. URL https://doi.org/10.1145/ 3458723

    doi:10.1145/3458723 2021

  16. [16]

    Selective classification for deep neural net- works

    Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural net- works. InAdvances in Neural Information Processing Systems (NeurIPS), volume 30, pages 4878–4887, 2017. URL https://proceedings.neurips.cc/paper/2017/hash/ 4a8423d5e91fda00bb7e46540e2b0cf1-Abstract.html. 15

    2017

  17. [17]

    Onebench to test them all: Sample-level benchmarking over open-ended capabilities

    Adhiraj Ghosh, Sebastian Dziadzio, Ameya Prabhu, Vishaal Udandarao, Samuel Albanie, and Matthias Bethge. Onebench to test them all: Sample-level benchmarking over open-ended capabilities. InACL, pages 32445–32481. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.acl-long.1560. URL https://aclanthology.org/2025.acl-long. 1560/

    doi:10.18653/v1/2025.acl-long.1560 2025

  18. [18]

    Gemini 2.5: Our most intelligent AI model, 2025

    Google DeepMind. Gemini 2.5: Our most intelligent AI model, 2025. URL https:// deepmind.google/technologies/gemini/

    2025

  19. [19]

    HaRC: Hallucinated reference checker, 2024

    HaRC Contributors. HaRC: Hallucinated reference checker, 2024. URL https://pypi.org/ project/harcx/

    2024

  20. [20]

    Princeton University Press, 2026

    Moritz Hardt.The Emerging Science of Machine Learning Benchmarks. Princeton University Press, 2026. Forthcoming; manuscript available athttps://mlbenchmarks.org/

    2026

  21. [21]

    The dataset nutrition label

    Sarah Holland, Ahmed Hosny, Sarah Newman, Joshua Joseph, and Kasia Chmielinski. The dataset nutrition label. InData Protection and Privacy, pages 1–25. Hart Publishing, 2020. doi: 10.5040/9781509932771.ch-001. URL https://doi.org/10.5040/9781509932771. ch-001

    work page doi:10.5040/9781509932771.ch-001 2020

  22. [22]

    RefChecker: Reference-based fine-grained hallucination checker and benchmark for large language models.arXiv preprint arXiv:2405.14486, 2024

    Xiangkun Hu, Dongyu Ru, Lin Qiu, Qipeng Guo, Tianhang Zhang, Yang Xu, Yun Luo, Pengfei Liu, Yue Zhang, and Zheng Zhang. RefChecker: Reference-based fine-grained hallucination checker and benchmark for large language models.arXiv preprint arXiv:2405.14486, 2024. URLhttps://github.com/amazon-science/RefChecker

    Pith/arXiv arXiv 2024

  23. [23]

    A survey on hallu- cination in large language models: Principles, taxonomy, challenges, and open questions

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qian- glong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallu- cination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1–55, 2025. doi: 10.1145/3703155. URL https://doi.or...

    doi:10.1145/3703155 2025

  24. [24]

    Livecodebench: Holistic and con- tamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and con- tamination free evaluation of large language models for code. InICLR, 2025. URL https://openreview.net/forum?id=chfJJYC3iL

    2025

  25. [25]

    Survey of hallucination in natural language generation

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1–38, 2023. doi: 10.1145/3571730. URL https://doi. org/10.1145/3571730

    doi:10.1145/3571730 2023

  26. [26]

    Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

    Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can language models resolve real-world GitHub issues? In Proceedings of ICLR, 2024. URLhttps://openreview.net/forum?id=VTF8yNQM66

    2024

  27. [27]

    Dynabench: Rethinking benchmarking in NLP

    Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, et al. Dynabench: Rethinking benchmarking in NLP. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4110–

    2021

  28. [28]

    Richard Landis and Gary G

    J. Richard Landis and Gary G. Koch. The measurement of observer agreement for categorical data.Biometrics, 33(1):159–174, 1977. doi: 10.2307/2529310

    doi:10.2307/2529310 1977

  29. [29]

    Halueval: A large- scale hallucination evaluation benchmark for large language models

    Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large- scale hallucination evaluation benchmark for large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449–6464. Association for Computational Linguistics, 2023. doi: 10.18653/v1/2023.emnlp-main.397. URLhttps:/...

    doi:10.18653/v1/2023.emnlp-main.397 2023

  30. [30]

    FActScore: Fine-grained atomic evaluation of factual precision in long form text generation

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. InConference on Empirical Methods in Natural Language Processing, pages 12076–12100. Association for Computational Linguistics,

  31. [31]

    Large enough: Mistral Large 2, 2024

    Mistral AI. Large enough: Mistral Large 2, 2024. URL https://mistral.ai/news/ mistral-large-2407/

    2024

  32. [32]

    Cooper, and Milos Hauskrecht

    Mahdi Pakdaman Naeini, Gregory F. Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using Bayesian binning. InProceedings of AAAI, 2015

    2015

  33. [33]

    GPT-5.1 Instant and GPT-5.1 Thinking system card addendum, 2025

    OpenAI. GPT-5.1 Instant and GPT-5.1 Thinking system card addendum, 2025. URLhttps: //openai.com/index/gpt-5-system-card-addendum-gpt-5-1/

    2025

  34. [34]

    PostTrainBench: Can LLM agents automate LLM post-training? arXiv preprint arXiv:2603.08640, 2026

    Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, and Maksym Andriushchenko. PostTrainBench: Can LLM agents automate LLM post-training? arXiv preprint arXiv:2603.08640, 2026. URLhttps://arxiv.org/abs/2603.08640

    arXiv 2026

  35. [35]

    BibTeX citation hallucinations in scientific publishing agents: Evaluation and mitigation

    Delip Rao and Chris Callison-Burch. BibTeX citation hallucinations in scientific publishing agents: Evaluation and mitigation. arXiv preprint arXiv:2604.03159, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Detecting and correcting reference hallu- cinations in commercial LLMs and deep research agents

    Delip Rao, Eric Wong, and Chris Callison-Burch. Detecting and correcting reference hallu- cinations in commercial LLMs and deep research agents. arXiv preprint arXiv:2604.03173, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    HALoGEN: Fantastic LLM hallucinations and where to find them.arXiv preprint arXiv:2501.08292, pages 1402– 1425, 2025

    Abhilasha Ravichander, Shrusti Ghela, David Wadden, and Yejin Choi. HALoGEN: Fantastic LLM hallucinations and where to find them.arXiv preprint arXiv:2501.08292, pages 1402– 1425, 2025. doi: 10.18653/v1/2025.acl-long.71. URL https://doi.org/10.18653/v1/ 2025.acl-long.71

    Pith/arXiv arXiv 2025

  38. [38]

    bibtex-updater: Automated BibTeX verification and updating, 2025

    Patrik Reizinger. bibtex-updater: Automated BibTeX verification and updating, 2025. URL https://github.com/rpatrik96/bibtexupdater

    2025

  39. [39]

    HalluCitation matters: Revealing the impact of hallucinated references with 300 hallucinated papers in ACL conferences.arXiv preprint arXiv:2601.18724, 2026

    Yusuke Sakai, Hidetaka Kamigaito, and Taro Watanabe. HalluCitation matters: Revealing the impact of hallucinated references with 300 hallucinated papers in ACL conferences.arXiv preprint arXiv:2601.18724, 2026

    arXiv 2026

  40. [40]

    HalluCiteChecker: A lightweight toolkit for hallucinated citation detection and verification in the era of AI scientists

    Yusuke Sakai, Hidetaka Kamigaito, and Taro Watanabe. HalluCiteChecker: A lightweight toolkit for hallucinated citation detection and verification in the era of AI scientists. arXiv preprint arXiv:2604.26835, 2026

    Pith/arXiv arXiv 2026

  41. [41]

    Chawla, and Yanfang Ye

    Kaiwen Shi, Weixiang Sun, Zheyuan 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.arXiv preprint arXiv:2602.23452, 2026. URL https://arxiv.org/abs/2602. 23452

    Pith/arXiv arXiv 2026

  42. [42]

    GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers, 2026

    Nazar Shmatko, Alex Adam, and Paul Esau. GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers, 2026. URL https://gptzero.me/news/neurips/. GPTZero analy- sis of 4,841 accepted NeurIPS 2025 papers, published January 21, 2026

    2025

  43. [43]

    verify-citations: Automated citation verification tool, 2025

    verify-citations Contributors. verify-citations: Automated citation verification tool, 2025. URL https://pypi.org/project/verify-citations/

    2025

  44. [44]

    Fact or fiction: Verifying scientific claims

    David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. Fact or fiction: Verifying scientific claims. InConference on Empirical Methods in Natural Language Processing, pages 7534–7550. Association for Computational Linguistics, 2020. doi: 10.18653/v1/2020.emnlp-main.609. URL https: //doi.org/10.18653...

    doi:10.18653/v1/2020.emnlp-main.609 2020

  45. [45]

    Walters and Esther Isabelle Wilder

    William H. Walters and Esther Isabelle Wilder. Fabrication and errors in the bibliographic citations generated by ChatGPT.Scientific Reports, 13:14045, 2023. doi: 10.1038/ s41598-023-41032-5. 17

    2023

  46. [46]

    GhostCite: A large-scale analysis of citation validity in the age of large language models.arXiv preprint arXiv:2602.06718, 2026

    Zuyao Xu, Yuqi Qiu, Lu Sun, Fasheng Miao, Fubin Wu, Xinyi Wang, Xiang Li, et al. GhostCite: A large-scale analysis of citation validity in the age of large language models.arXiv preprint arXiv:2602.06718, 2026

    Pith/arXiv arXiv 2026

  47. [47]

    flag everything unfamiliar

    An Yang et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025. 18 Appendix Table of Contents A Dataset and taxonomy 21 A.1 Full taxonomy details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 A.2 Real-world incident mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 A.3 Construction details . . . . . . . . . ....

    Pith/arXiv arXiv 2025

  48. [52]

    DOI format validation: Checks that DOI strings match the expected format (10.XXXX/...) and that the DOI prefix corresponds to a known registrant. 26 Citation BibTeX entry Pre-screening DOI format year bounds name heuristics Tool call API / LLM Metrics DR, FPR, TW-F1, ECE Override direct verdict pass flag shared layer Figure 3:Evaluation pipeline.Each cita...

    1900

  49. [53]

    John Doe,

    Author name heuristics: Detects common placeholder patterns (“John Doe,” “A. Author,” single- word author names, repeated names). Pre-screening results are tagged with[Pre-screening override] in the reason string to maintain transparency about which detections come from the pre-screening layer vs. the external tool. B.2.2 HaRC and verify-citations: rate-l...

  50. [54]

    Is the title p l a u s i b l e and does it match known work by these authors ?

  51. [55]

    Are the authors real r e s e a r c h e r s in this field ?

  52. [56]

    Is the venue ( journal / c o n f e r e n c e ) real ?

  53. [57]

    Does the year make sense ?

  54. [58]

    Author1

    If a DOI is present , does it look pr op erl y f o r m a t t e d ? When the entry is HALLUCINATED , cl as si fy the h a l l u c i n a t i o n mode using exactly one of : ‘ fabricated_doi ‘ , ‘ n o n e x i s t e n t _ v e n u e ‘ , ‘ p l a c e h o l d e r _ a u t h o r s ‘ , ‘ future_date ‘ , ‘ chimeric_title ‘ , ‘ wrong_venue ‘ , ‘ swapped_authors ‘ , ‘ p...

    1900

  55. [59]

    Inspect the entry for obvious red flags ( fake DOI prefix , future year , p l a c e h o l d e r authors )

  56. [60]

    Use tools to cross - re fe re nc e : resolve the DOI , or search by title / author

  57. [61]

    label ":

    After ga th er ing evidence ( or after finding s u f f i c i e n t signal ) , emit your verdict . When you are ready to give your final answer , output ONLY valid JSON -- no prose , no markdown fences : { " label ": " VALID " or " H A L L U C I N A T E D " , " c o n f i d e n c e ": 0.0 to 1.0 , " reason ": " concise e x p l a n a t i o n citing the evide...

  58. [62]

    For almost every entry , call ‘ v e r i f y _ w i t h _ b i b t e x _ u p d a t e r ‘ once with the exact BibTeX string you were given

  59. [63]

    I nt erp re t the returned ‘ status ‘ field : statuses like ‘ verified ‘ , ‘ url_verified ‘ , or ‘ p u b l i s h e d _ v e r s i o n _ e x i s t s ‘ suggest VALID ; statuses like ‘ not_found ‘ , ‘ title_mismatch ‘ , ‘ author_mismatch ‘ , ‘ hallucinated ‘ , ‘ future_date ‘ , ‘ doi_not_found ‘ , or ‘ venue_mismatch ‘ suggest H A L L U C I N A T E D

  60. [64]

    If the tool returns ‘ api_error ‘ or you suspect the tool is wrong ( e . g . it reports ‘ verified ‘ but the entry still looks s u s p i c i o u s on i n s p e c t i o n ) , apply your own judgment -- the tool is evidence , not an oracle

  61. [65]

    label ":

    If the first call is unambiguous , do NOT call again . Extra calls waste budget . When you are ready to give your final answer , output ONLY valid JSON -- no prose , no markdown fences : { " label ": " VALID " or " H A L L U C I N A T E D " , " c o n f i d e n c e ": 0.0 to 1.0 , " reason ": " concise e x p l a n a t i o n r e f e r e n c i n g the tool s...

    2021

  62. [66]

    This rules out approaches requiring human-in-the-loop validation or expensive API calls to language models

    Zero human effort.Verification must be fully automated: no manual review, no prompt engineer- ing, no LLM inference costs. This rules out approaches requiring human-in-the-loop validation or expensive API calls to language models

  63. [67]

    A tool that requires a separate platform or manual invocation will not be adopted

    Workflow integration.The tool must integrate into existing pipelines: CI/CD (GitHub Actions), pre-commit hooks, Overleaf builds, and one-off command-line checks. A tool that requires a separate platform or manual invocation will not be adopted

  64. [68]

    Co-designed (reference upper bound)

    Graceful degradation.When APIs are unavailable or rate-limited, the tool should return partial results rather than fail silently. Venues processing hundreds of submissions cannot tolerate flaky infrastructure. 50 0.2 0.4 0.6 0.8 1.0 Saturation ratio p95/cap (lower is healthier) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8F1-Hallucination saturated budget capmarker siz...

    2021

  65. [2015]

    URL https://doi.org/10.1126/science.aaa9375

    doi: 10.1126/science.aaa9375. URL https://doi.org/10.1126/science.aaa9375

    doi:10.1126/science.aaa9375

  66. [2023]

    URL https://doi.org/10.18653/v1/ 2023.emnlp-main.741

    doi: 10.18653/v1/2023.emnlp-main.741. URL https://doi.org/10.18653/v1/ 2023.emnlp-main.741

    doi:10.18653/v1/2023.emnlp-main.741 2023

  67. [2026]

    The 448-entry temporal supplement (§ F.1) validates the probe’s findings at scale; the probe additionally covers 2026 arXiv submissions

    combines 21 perturbation and 9 adversarial entries. The 448-entry temporal supplement (§ F.1) validates the probe’s findings at scale; the probe additionally covers 2026 arXiv submissions. The cross-domain split(test_crossdomain, 500 entries: 200 valid / 300 hallucinated) probes transfer outside the ML-venue regime the main splits sample: 299 biomedical e...

    2026

  68. [4124]

    doi: 10.18653/v1/2021.naacl-main.324

    Association for Computational Linguistics, 2021. doi: 10.18653/v1/2021.naacl-main.324. URLhttps://doi.org/10.18653/v1/2021.naacl-main.324

    doi:10.18653/v1/2021.naacl-main.324 2021