When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree

Shinwoo Park; Yo-Sub Han

REVIEW 4 major objections 4 minor 6 references

When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A 16-criterion rubric, built after annotators saw the answers, raised their majority-vote accuracy at spotting AI essays from 60% to 90% and shows LLM essays outscore students on formal polish while losing on creativity.

desk verdict Korean pilot showing 60→90% detection accuracy is transparent and useful, but the causal claim about rubric scaffolding is not supported by a single-arm bundled design. read the letter →

arxiv 2601.19913 v4 pith:VJ65XLVW submitted 2026-01-06 cs.CL cs.AI

Shinwoo Park , Yo-Sub Han This is my paper · ORCID

classification cs.CLcs.AI

keywords LREADLLM-generatedtextdetectionhumanattributionKoreanargumentativeessayswritingrubriccalibrationfluencytrapassessment

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

This paper is trying to establish that human readers can be calibrated to distinguish student essays from LLM-generated Korean text, and that a criterion-anchored rubric is the mechanism that makes this work. Three linguistically trained readers first guessed at 60% accuracy using intuition alone; after jointly building a 16-criterion, 100-point writing rubric from national standards and their own errors, they reached 90% on a disjoint 30-essay set and 10/10 on a small held-out set. The rubric also explains the scoring gap between humans and machines: LLM essays earn near-perfect marks on orthographic norms and genre-appropriate register, while student essays are rated higher on creativity and natural Korean phrasing. The authors' claim is that attribution is a learnable, auditable skill, and that surface well-formedness is a misleading cue for human authorship.

What carries the argument

The load-bearing instrument is LREAD, a 16-criterion, 100-point rubric for Korean argumentative essays, built in two moves: it starts from the national writing standards and is then adapted using the annotators' Phase-1 errors and reflections after gold-label disclosure. The protocol is a three-phase longitudinal design — intuitive blind screening, rubric-guided scoring with explicit justifications, and internalized blind application on a held-out subset — with majority voting, uncertainty marking, and inclusive Fleiss' kappa used to track agreement. The rubric's Expression category (50 of 100 points) is where the action is: it contains detection-oriented micro-diagnostics such as orthograph

What would settle it

Run the same protocol with a new, independent panel of Korean linguists who receive only the written LREAD rubric, no gold-label feedback and no prior exposure to GPT-4o, Solar, Qwen2, or Llama3.1, on a fresh 30-essay set; if their majority-vote accuracy does not substantially exceed the 60% intuition baseline, the central calibration claim is not robust.

Watch

Extended reading notes

Core claim

The central claim is that the 'fluency trap' — readers over-trusting the polished local well-formedness of LLM text — can be broken by moving from holistic intuition to explicit, criterion-anchored scoring. The paper grounds this in a three-phase study: Phase 1 blind intuition (majority-vote accuracy 0.60, not significantly above chance, with agreement near zero); Phase 2, after gold-label disclosure and collaborative rubric construction, a disjoint set scored with the LREAD rubric reaches 0.90 with false negatives on AI essays dropping from 12 to 2; Phase 3, a 10-essay elementary-persona subset, is classified perfectly, though with a wide confidence interval. The rubric includes content, or

Load-bearing premise

The improvement is credited to the rubric, but the rubric was built after the annotators saw the gold labels and was tested by the same three annotators on the same four generator models, so the 90% gain may be an artifact of that particular panel and set rather than a portable property of rubrics.

Editorial extensions

If this is right

Rubric-guided evaluation can complement automated detectors, producing explicit, auditable reasons for attributing a text to a human or an AI.
Formal-conformity criteria push overall scores toward LLM essays: the pooled LLM mean is 18.35 points above the student mean, with orthography and register accounting for 8.19 of those points.
Readers can be trained out of the fluency trap: the calibration gain comes mainly from reducing false negatives on AI essays, not from blanket over-detection of human work.
Creativity and natural Korean phrasing are where student writing outperforms machine writing, suggesting that rubrics that reward those traits would change the human–AI ranking.
The released rubric and cue taxonomy provide a starting point for adapting the same calibration idea to other languages, genres, and generator populations.

Reading between the lines

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

If adopted as a grading rubric in classrooms, this kind of formal-conformity weighting would systematically favor AI-written essays; the paper's small sample (6 student essays) is enough to pose the hypothesis, not to settle it.
The 90% result is a within-panel result. A fresh set of annotators who did not help build the rubric, and newer models not in the original four, are the natural stress test; until then, the generalizability is unproven.
A direct extension would be to re-weight the rubric toward creativity and natural phrasing and see whether the 18.35-point gap reverses — a testable prediction from the paper's criterion-level scores.
The rubric's reliance on specific micro-cues (spacing, optional commas, translationese) means model updates that smooth those habits would erode the signal; the authors acknowledge this as a temporal-generalization risk.

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 — quick read of arXiv:2601.19913.

Punchline: the descriptive result is credible and the reporting is unusually honest, but the paper's central causal claim — that the LREAD rubric scaffolded humans to substantially better attribution — is not supported by the design. The abstract and conclusion overstate.

What's actually new: a 16-criterion Korean-specific rubric for distinguishing LLM from human argumentative essays, a three-phase protocol with three linguistically trained annotators, and a transparently released dataset and rubric (GitHub). The authors show majority-vote accuracy rising from 60% to 90% on disjoint sets, with agreement going from negative Fleiss' kappa to 0.82, and they report exact CIs, conservative voting rules, and a detailed appendix. That is a legitimate pilot result for a non-English language.

What's genuinely good: the paper acknowledges its own limitations in a way you don't always see. The bundle of interventions is named; the post-hoc construction of the rubric is admitted; Phase 3's tiny sample is flagged. The cue taxonomy in Appendix D is useful as an inventory to test.

Soft spots, in proportion. The biggest is the bundled intervention. Gold-label disclosure, collaborative rubric construction, model identity knowledge, and the requirement to produce justifications all happened together. The rubric was built by the same three annotators after seeing Phase 1 gold labels, so Phase 2 is partly measuring self-consistency with their own instrument. The individual trajectories show the majority gain is mainly one annotator (A2 46.7→96.7; A1 50→66.7). No control separates the rubric from simpler calibration effects like seeing the answers. The stress-test note is right: the 90% is not attributable to the rubric per se. Also, set difficulty across phases is unmeasured; the Phase 2 set could be easier.

A separate problem is the framing mismatch. The title and the top abstract describe a secondary audit of rubric scores (LLM essays outscore student essays by 18.35 points). The full text, though, is a detection-calibration study with its own abstract. That mismatch is jarring and would confuse any reader; it needs to be fixed before anything else.

The descriptive finding that LLM essays receive higher formal-expression scores and humans higher creativity scores is plausible but unsurprising; the rubric was built to encode those exact signals, so the score comparison is partly circular.

Who this is for: people studying human detection of AI text, educational assessment, and Korean NLP. It is a useful pilot, not a validated method. A serious referee would get value from it, mainly to push the authors to soften the causal claim or add ablations/independent raters. I'd send it to review, not desk reject, because the artifact and reporting are real contributions. But the causal framing needs a major revision.

Referee Report

4 major / 4 minor

Summary. The paper reports a three-phase longitudinal study in which three Korean-language majors first performed intuition-only human-vs-AI attribution on 30 Korean argumentative essays (Phase 1), then co-designed a 100-point, 16-criterion rubric (LREAD) grounded in national writing standards and their own Phase 1 errors, and used it to score and attribute a disjoint 30-essay set (Phase 2), before a 10-essay elementary-persona held-out test (Phase 3). Majority-vote accuracy rises from 60% to 90% to 100%, Fleiss' kappa from -0.088 to 0.243 to 0.820, and false negatives on AI essays drop from 12/24 to 2/24 to 0/8. The paper also compares calibrated human cues against zero-shot LLM judges. The authors position the work as pilot evidence that rubric-scaffolded judgment improves attribution and release the rubric and dataset.

Significance. If the causal claim were established, the paper would make a useful contribution: an auditable, language-specific rubric procedure for human-centered LLM-text attribution, with practical relevance for Korean educational assessment. The descriptive trajectories, the released rubric, the exact statistical reporting, and the transparent limitations are genuine strengths. However, the central causal inference — that the rubric, rather than bundled feedback, rubric construction, model familiarity, or practice, caused the accuracy gain — is not supported by the current single-arm design. The paper's value at this stage is as a hypothesis-generating pilot, not as evidence that rubric scaffolding per se improves human attribution. The abstract and title overstate the strength of the conclusion.

major comments (4)

[§4.1–4.2, Fig. 1, Table 4] The claimed Phase 1→2 improvement (0.60→0.90) is attributed to 'criterion-anchored evaluation,' but the intervention bundles at least four components: (a) gold-label disclosure for the Phase 1 essays, (b) collaborative construction of LREAD after that disclosure, (c) knowledge of the four generator identities, and (d) a requirement to produce explicit rubric-justified scores. No control condition separates these. The Limitations section 'Ablation of Calibration Components' concedes this. Consequently, the 0.90 result is consistent with feedback, rubric construction, or practice effects, not specifically the rubric. The abstract's 'substantially improves performance' is too strong for a single-arm longitudinal pilot. Please either add ablation conditions (feedback-only, rubric-without-gold) or reframe the claim as descriptive within-panel improvement.
[§4.1, Limitations 'Post-hoc Rubric Design'] LREAD was constructed by the same three annotators after they received Phase 1 gold labels, and then used by those annotators in Phase 2. This is circular for measuring their attribution accuracy: the rubric is an explicit encoding of the annotators' own post-hoc rationalizations of their errors. The disjointness of the essay sets does not address overfitting to the annotators' idiosyncratic cue preferences, nor to the four specific models. Independent annotators who did not participate in rubric construction, or a pre-registered rubric defined before any labels, are needed to establish that the rubric rather than annotator adaptation drives the gain. The current design cannot rule out self-consistency effects.
[Table 6, §5.1] The panel-level gain is fragile. A2 improves from 46.7% to 96.7% between Phases 1 and 2, while A1 improves only from 50.0% to 66.7% and A3 from 80.0% to 86.7%. The majority vote's jump from 60% to 90% is therefore largely driven by one annotator's idiosyncratic improvement. This undercuts the interpretation of a stable rubric-induced calibration effect. Please report majority accuracy without each annotator (or use a robust aggregation), and discuss the heterogeneity explicitly; a random-effects analysis would also clarify whether the improvement is systematic.
[Abstract vs. body] The title and the abstract provided with the manuscript describe a rubric-score audit (pooled LLM mean exceeds student mean by 18.35 points; orthographic/register criteria contribute 8.19 points; all 72 LLM ratings at orthographic maximum). The body's own abstract and the main narrative, by contrast, focus on detection accuracy (60%→90%→100%) and do not present the scoring-gap analysis except indirectly in Appendix B (Table 11). This mismatch makes the paper's central contribution ambiguous and leaves key abstract claims (e.g., 'all 72 LLM ratings reach the orthographic maximum') unsupported in the main text. The authors must reconcile the abstract with the body or explicitly frame the scoring-gap analysis as a secondary contribution.

minor comments (4)

[Fig. 1] The y-axis starts at 40, which visually inflates the Phase 1→2 difference. Please start at 0 or clearly indicate a broken axis. Also, individual annotator points are plotted without error bars; this is acceptable for descriptive purposes but should be stated.
[§5.3, Table 4] The inclusive definition of Fleiss' kappa (mapping uncertainty to a third category) is described, but the main text should explicitly note that this choice inflates agreement relative to a binary-only treatment. Please add a sentence acknowledging that the kappa values are not directly comparable to standard binary kappa.
[§4.2] Phase 3 is described as 'blind,' but the protocol says annotators receive a second disclosure of ground-truth labels before the phase. Please clarify what 'blind' means here — presumably they do not see labels for the 10 held-out essays during attribution, but the preceding disclosure may carry over. This affects interpretation.
[Appendix C.1] The conservative majority-vote rule (any three-way tie is resolved to Human) may bias accuracy upward when the true label is Human and downward when it is AI. Please report how many tie votes occurred in each phase and whether the results change under an alternative tie-breaking rule.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Phase 2 is an out-of-sample empirical measurement; post-hoc rubric design is a stated limitation, not a circular reduction.

full rationale

The paper's central claim is an empirical accuracy comparison across phases, not a derivation. Phase 2 uses a disjoint 30-essay set with withheld labels (Section 4.2: 'We then conduct a second blind evaluation using a new set of 30 essay samples... deliberately withhold the label distribution'), so the 0.90 accuracy is an out-of-sample measurement that could have failed; it is not forced by the Phase 1 rubric construction. The rubric is admittedly post hoc ('the rubric itself is developed only after the completion of the Phase 1 blind evaluation'), and the protocol bundles gold-label feedback with rubric construction, but the paper explicitly flags these limitations in 'Post-hoc Rubric Design' and 'Ablation of Calibration Components.' These are threats to causal attribution and generalizability, not circular reductions. The self-citations (KatFish benchmark, WaterMod) are dataset/related-work references, not used to justify the calibration claim via a uniqueness theorem or an imported ansatz. No equation is defined in terms of its output, no fitted parameter is renamed as a prediction. The gap decomposition (orthographic/register contributing 8.19 points) is descriptive of the Phase 2 scores, and its dependence on the post-hoc weighting is acknowledged as an overfitting concern rather than hidden. Hence no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the ground-truth labels of the KatFish benchmark, the external validity of the national writing standards, the competence of three non-randomly selected annotators, and — most importantly — the assumption that post-hoc rubric construction by the same judges does not overfit the measured accuracy gain. The rubric weights and the percentage anchors are ad hoc design choices fitted to Phase 1 observations. No new physical or linguistic entities are introduced; LREAD is a protocol, not an entity.

free parameters (3)

Rubric weight allocation (100 points across Content/Organization/Expression) = Content 40, Organization 10, Expression 50
Section 4.1: the weighting 'reflects our empirical observation from the Phase 1 blind study'; Expression is deliberately over-weighted because orthography/register were seen as diagnostic. This is an ad hoc design choice fitted to the observed failure modes, not an externally validated standard.
Percentage anchors for error thresholds = e.g., orthographic errors <1%, 1–3%, ≥3% of words; sentence errors <10%, 10–20%, 20–30%, >30%
Appendix A: these anchors are chosen by the rubric developers without derivation from an external norm. They directly map to point scores and influence which essays receive maximum scores on expression criteria.
Conservative deduplication rule for error counting = Repeated identical error patterns counted once
Appendix A: 'We apply conservative deduplication when the same error pattern recurs... treating repeated spelling errors with the same erroneous form as a single case.' This choice affects both orthographic and sentence-construction scores and is not externally benchmarked.

assumptions (5)

domain assumption KatFish ground-truth labels are accurate (essays are correctly labeled human vs. AI)
The entire calibration measurement compares annotator judgments to these labels (Section 3.1). If any label is wrong, accuracy figures are wrong. The paper cites Park et al. (2025) but does not independently audit the labels.
domain assumption The National Institute of Korean Language standards for argumentative writing are a valid external basis for the rubric
Section 4.1 states LREAD is 'derived from the 2025 national standards' issued by NIKL. The rubric's content and weights inherit the validity of those standards.
domain assumption Three advanced undergraduates in Korean Language and Literature constitute a meaningfully 'linguistically trained' panel whose judgments are competent measures
The paper recruits three annotators and treats their judgments as the object of study. The small homogeneous panel limits generalization, as the Limitations note, but the analysis still depends on this assumption.
ad hoc to paper Phase 1 gold-label disclosure and rubric construction do not overfit the specific annotators and models
The central causal claim requires that Phase 2 improvement reflects a transferable calibration effect, not memorization of the four generators' styles or the judges' own rubric-building confabulations. The paper asserts disjoint sets mitigate this but provides no independent replication.
domain assumption LLM judge baselines (GPT-5.2 Thinking, Gemini 3 Flash, Claude Sonnet 4.5) under provider-default web-interface settings are a fair zero-shot comparison
Section C.7 states no temperature/top-p tuning was done and the comparison is 'not a fairness-controlled contest'. Still, the 96.67% LLM baseline is used to motivate the human–LLM gap, so the baseline's validity matters.

how reviews work

0 comments

Cite this review

Pith. "Pith review of When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree." pith.science (2026). https://pith.science/paper/VJ65XLVW

@misc{pith2026260119913,
  author       = {Pith},
  title        = {Pith review of: When LLM Essays Outscore Student Essays: What a Korean Writing Rubric Rewards and Where Readers Disagree},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJ65XLVW}},
  note         = {Machine review of arXiv:2601.19913}
}

read the original abstract

LLMs now help students plan, draft, and revise essays. Educational assessment therefore faces a basic question: how should student and LLM writing be compared? Rubrics assign points to content, organization, and expression. Their total can still hide which criteria drive the comparison, where ratings approach the maximum, and where readers disagree. We therefore conducted a secondary, post hoc audit of a Korean writing study with source-informed scoring. Three Korean language and literature majors used a 16-criterion, 100-point rubric to score six student essays and 24 essays from four LLMs prompted for three school levels. After source disclosure, they helped develop the rubric and, according to the protocol, scored the Phase 2 essays without gold labels while recording source judgments. The pooled LLM mean exceeds the student mean by 18.35 points. Orthographic norms and genre-appropriate register contribute 8.19 points, or 44.7% of the gap. All 72 LLM ratings reach the orthographic maximum, and 70 reach the register maximum. The student source mean is highest or tied highest on both creativity criteria and natural Korean phrasing, where reader agreement is weak. Criterion-level auditing therefore offers a deeper account of human and LLM writing than the total alone.

Figures

Figures reproduced from arXiv: 2601.19913 by the authors.

**Figure 1.** Detection accuracy across the three phases for [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

**Figure 3.** Phase 2 rubric profile by generator. The radar [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: The full LREAD rubric with anchored scoring criteria (Korean original). B.1 Individual Annotator Accuracy Trajectories [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

**Figure 5.** Figure 5: The full LREAD rubric with anchored scoring criteria (English translation). and human-versus-LLM majority-vote comparison. B.5 Rubric-Score Diagnostics [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Distribution of mean Phase 2 LREAD scores by source group, aggregated over three annotators for each essay. Higher scores indicate stronger formal conformity to rubric criteria, not necessarily stronger evidence of human authorship. is a meaningful part of the calibrat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 1 linked inside Pith

[3]

InProceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (ACL)

People who frequently use ChatGPT for writ- ing tasks are accurate and robust detectors of AI- generated text. InProceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (ACL). Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov
[5]

InProceedingsofthe2023Con- ference on Empirical Methods in Natural Language Processing (EMNLP)

Ontheautomaticgenerationandsimplification ofchildren’sstories. InProceedingsofthe2023Con- ference on Empirical Methods in Natural Language Processing (EMNLP). ZhengxiangWang,NafisIrtizaTripto,SolhaPark,Zhen- zhen Li, and Jiawei Zhou. 2025a. Catch me if you can?notyet:LLMsstillstruggletoimitatetheimplicit writing styles of everyday authors. InFindings of t...

2025
[6]

A FullLREADRubric and Detection-Oriented Adaptations This section documents the fullLREAD rubric as used in this study

Embracing ai in education: Understanding the surge in large language model use by secondary students.arXiv preprint arXiv:2411.18708. A FullLREADRubric and Detection-Oriented Adaptations This section documents the fullLREAD rubric as used in this study. We distinguish between (i) sta- ble, framework-level dimensions that are likely to transfer across sett...

arXiv
[2023]

InFindings of the Association for Computational Linguistics (EMNLP)

DetectLLM:LeveragingLogRankInformation for Zero-Shot Detection of Machine-Generated Text. InFindings of the Association for Computational Linguistics (EMNLP). Maria Valentini, Jennifer Weber, Jesus Salcido, Téa Wright,ElianaColunga,andKatharinavonderWense
[2024]

InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)

Pretraining language models using transla- tionese. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP). LiamDugan,DaphneIppolito,ArunKirubarajan,Sherry Shi, and Chris Callison-Burch. 2023. Real or fake text?:Investigatinghumanabilitytodetectboundaries between human-written and machine-generated text. InProceed...

arXiv 2024
[2025]

InProceedingsofthe 63rd Annual Meeting of the Association for Compu- tational Linguistics (ACL)

Lost in literalism: How supervised training shapestranslationeseinLLMs. InProceedingsofthe 63rd Annual Meeting of the Association for Compu- tational Linguistics (ACL). YijianLu,AiweiLiu,DianzhiYu,JingjingLi,andIrwin King. 2024. An entropy-based text watermarking detection method. InProceedings of the 62nd An- nual Meeting of the Association for Computati...

2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [3]

InProceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (ACL)

People who frequently use ChatGPT for writ- ing tasks are accurate and robust detectors of AI- generated text. InProceedings of the 63rd Annual Meeting of the Association for Computational Lin- guistics (ACL). Jinyan Su, Terry Zhuo, Di Wang, and Preslav Nakov

[2] [5]

InProceedingsofthe2023Con- ference on Empirical Methods in Natural Language Processing (EMNLP)

Ontheautomaticgenerationandsimplification ofchildren’sstories. InProceedingsofthe2023Con- ference on Empirical Methods in Natural Language Processing (EMNLP). ZhengxiangWang,NafisIrtizaTripto,SolhaPark,Zhen- zhen Li, and Jiawei Zhou. 2025a. Catch me if you can?notyet:LLMsstillstruggletoimitatetheimplicit writing styles of everyday authors. InFindings of t...

2025

[3] [6]

A FullLREADRubric and Detection-Oriented Adaptations This section documents the fullLREAD rubric as used in this study

Embracing ai in education: Understanding the surge in large language model use by secondary students.arXiv preprint arXiv:2411.18708. A FullLREADRubric and Detection-Oriented Adaptations This section documents the fullLREAD rubric as used in this study. We distinguish between (i) sta- ble, framework-level dimensions that are likely to transfer across sett...

arXiv

[4] [2023]

InFindings of the Association for Computational Linguistics (EMNLP)

DetectLLM:LeveragingLogRankInformation for Zero-Shot Detection of Machine-Generated Text. InFindings of the Association for Computational Linguistics (EMNLP). Maria Valentini, Jennifer Weber, Jesus Salcido, Téa Wright,ElianaColunga,andKatharinavonderWense

[5] [2024]

InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)

Pretraining language models using transla- tionese. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP). LiamDugan,DaphneIppolito,ArunKirubarajan,Sherry Shi, and Chris Callison-Burch. 2023. Real or fake text?:Investigatinghumanabilitytodetectboundaries between human-written and machine-generated text. InProceed...

arXiv 2024

[6] [2025]

InProceedingsofthe 63rd Annual Meeting of the Association for Compu- tational Linguistics (ACL)

Lost in literalism: How supervised training shapestranslationeseinLLMs. InProceedingsofthe 63rd Annual Meeting of the Association for Compu- tational Linguistics (ACL). YijianLu,AiweiLiu,DianzhiYu,JingjingLi,andIrwin King. 2024. An entropy-based text watermarking detection method. InProceedings of the 62nd An- nual Meeting of the Association for Computati...

2024