Pith. sign in

REVIEW 4 major objections 6 minor 64 references

Pangram 4 detects frontier AI text at AUROC 0.9916 with about one false positive in 24,000 documents, and labels human, AI-assisted, and AI-generated spans in a single pass.

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

2026-07-30 11:00 UTC pith:2PYFFFJZ

load-bearing objection Strong closed-product binary detector numbers; the mixed-authorship “novelty” is real engineering but partly graded against the authors’ own Soft N-Grams oracle. the 4 major comments →

arxiv 2607.27183 v1 pith:2PYFFFJZ submitted 2026-07-29 cs.CL

Pangram 4 Technical Report

Ben Glickenhaus , Katherine Thai , Jenna Russell , Elyas Masrour , Yue Han , Max Spero , Bradley Emi This is my paper
classification cs.CL
keywords AI-text detectionmixed authorshiptokenwise provenanceSoft N-Grams labelinghumanizer robustnessfalse positive rateEditLensCRF decoding
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.

This technical report introduces Pangram 4, a deep-learning AI-text classifier built to tell fully human writing, fully AI writing, and mixed co-authorship apart at high precision. On more than a million English human texts it reports a false positive rate of 0.0041% (about one in 24,300), and on 520,000 generations from 26 frontier models a false negative rate of 0.3396%, with AUROC 0.9916. Beyond binary detection, the model predicts three labels per token—human, AI-assisted, and AI-generated—so it can mark both clean block boundaries in interleaved documents and homogeneous edits where human ideas and AI wording are entangled. Training rests on synthetic mirrors of licensed human sources, AI-assisted edits labeled by a clause-level Soft N-Grams procedure, a shared mixture-of-experts backbone with segment, tokenwise, mixed-authorship, and humanizer heads, and a calibrated sliding-window CRF decoder. The authors argue this combination raises accuracy, mixed-authorship resolution, and resistance to humanizers and other attacks relative to Pangram 3, which matters wherever readers need to know how much of a document an LLM actually wrote.

Core claim

Pangram 4 is presented as a single model that simultaneously achieves very low document-level error on fully human and fully AI text and, for the first time in this line of work, joint token-level detection of both homogeneous and heterogeneous mixed authorship, with improved boundary resolution and humanizer robustness over Pangram 3 across internal and public benchmarks.

What carries the argument

Soft N-Grams labeling (clause-level lexical and semantic match of an AI-edited target against a hidden human source, with AI-assisted weighted at half in the AI fraction) supplies token supervision; a causal MoE backbone with Repeat2, a three-class token head, segment and mixed heads, a stop-gradient humanizer probe, and linear-chain CRF postprocessing turns those labels into calibrated spans.

Load-bearing premise

The clause-level Soft N-Grams rules—and weighting AI-assisted text as half AI—are treated as faithful ground truth for training and mixed-authorship evaluation even though the original human source is invisible at test time and per-token “AI-assisted” is acknowledged as fuzzy.

What would settle it

Hold out a large set of real human–AI co-writing sessions with independent, human-agreed span labels (not Soft N-Grams), score Pangram 4’s token and document metrics against those labels, and check whether production FPR stays near 0.0041% and mixed recall holds; large systematic disagreement would overturn the central mixed-authorship claim.

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

If this is right

  • Readers and platforms can demand span-level provenance rather than a single AI/human bit for long documents.
  • Light AI proofreading can be separated from heavy rewrite at much lower false-alarm rates than prior Pangram releases.
  • Humanizer and commercial evasion tools become less reliable ways to hide frontier-model prose under this operating point.
  • Multilingual and non-native English corpora can be screened at FPR comparable to native English on the reported sets.
  • Benchmark tables that only report binary AUROC understate the mixed-authorship task the model is built to solve.

Where Pith is reading between the lines

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

  • If Soft N-Grams remains the main supervision source, detectors may overfit to LLM edit styles present in training mirrors and under-detect novel co-writing workflows that never look like those edits.
  • Very low FPR at production threshold will push adversaries toward out-of-scope formats (terse notes, heavy context-copying) rather than open prose, shifting the practical attack surface.
  • Tokenwise mixed labels create a natural product surface for editors and classrooms: highlight only the AI-generated clauses instead of rejecting whole submissions.
  • Consistency across isolated excerpts versus full-document context, flagged as a limitation, will become a user-trust bottleneck as span UI ships.

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

4 major / 6 minor

Summary. Pangram 4 is a technical report on a commercial AI-text detector built on an anonymized open-weight MoE backbone with LoRA, trained in two stages and decoded via sliding-window Repeat2 token heads, segment heads, a mixed-authorship head, and a linear-chain CRF. The model outputs per-token labels in {human, AI-assisted, AI-generated} plus a document-level humanizer probe. The central empirical claims are production-point binary performance (AUROC 0.9916, FPR 0.0041% on >1M English human texts, FNR 0.3396% on ~520k frontier generations), improved length/multilingual/ESL behavior versus Pangram 3, adversarial and humanizer robustness, and—as the main novelty—simultaneous homogeneous and heterogeneous mixed-authorship detection with finer boundary resolution. Mixed supervision and much of the mixed evaluation rest on Soft N-Grams labeling (Algorithm 1) and the weighted AI fraction f_AI.

Significance. If the binary operating-point numbers hold under independent audit, the work is highly significant for practical AI-text detection: sub-0.01% FPR at sub-0.5% FNR on large, multi-generator suites, plus strong public-benchmark and humanizer results, would set a high bar relative to GPTZero, Originality, and zero-shot baselines. The architectural package (tokenwise provenance, Repeat2, CRF postprocessing, stop-gradient humanizer probe) is a concrete engineering contribution. The mixed-authorship claim would also matter if grounded in independent provenance, because real writing is increasingly co-authored. Strengths that should be credited include large-N FPR/FNR with Wilson intervals, per-generator and length breakdowns, FineWeb2 multilingual FPR, ESL corpora, BLADER/commercial humanizer tests, and broad public-benchmark coverage. The closed model and Soft N-Grams-dependent mixed metrics currently limit how much of the novelty can be treated as settled science.

major comments (4)
  1. [§3.5, Algorithm 1; §5.5.1 Tables 5–6] §3.5 Algorithm 1 and §5.5.1 Table 6: The headline mixed-authorship novelty (simultaneous homogeneous/heterogeneous detection and high AI-Assisted recall) is not independently grounded. Soft N-Grams defines clause labels via Claude Haiku splits plus lexical/semantic match to a hidden source S, with AI-assisted weighted 0.5 in f_AI. The model is trained on those labels, and the key WildChat Edits set (n=4,826) is filtered to Soft N-Grams f_AI ∈ [0.25, 0.75]. Reporting 65.7% Mixed recall on that set largely measures recovery of the same oracle the paper already calls nebulous per token (§3.3) and unavailable at test time (§3.5). This does not invalidate binary FPR/FNR, but it does make the mixed SOTA/novelty claim load-bearing-circular. Please either (i) evaluate homogeneous recall only on external intent/tool sets with human or multi-annotator provenance, or (ii) clearly separate “agreemen
  2. [§5.8, Table 29 (OpAI-Bench)] §5.8 / Appendix Table 29 (OpAI-Bench): Character-level ground truth is available, yet Pangram 4’s mean AI+Assisted fraction badly under-tracks true AI coverage at high-edit stages (e.g., mean ~0.70 vs true ~0.99 at v8; large gaps also at v5–v7). This is the strongest external continuous mixed test in the paper and weakens the claim of calibrated fine-grained mixed detection. Either improve calibration on this trajectory, report error stratified by true AI fraction with confidence intervals, or qualify the mixed claim to boundary/heterogeneous settings where Table 7 is stronger.
  3. [§5.1; §4.3.3–4.3.4] §5.1 decode rule and §4.3 CRF/calibration: Document labels use fixed free thresholds (f_human≥0.90→Human, f_AI≥0.80→AI, else Mixed), and Mixed counts against both FPR and FNR. Combined with fitted unary/transition parameters (α_c, w_c, δ_c, λ, γ, L_min) and the 15% mixed-head rule, operating-point metrics are not uniquely determined by the backbone. For the central production claims (FPR 0.0041%, FNR 0.3396%), report sensitivity to these thresholds, or release the exact calibration fit and a frozen decode config so results are reproducible in principle. Without that, cross-paper “SOTA” comparisons remain partly postprocessing choices.
  4. [§2; §4.1 Table 1; §5.2–5.3] Reproducibility and auditability of the strongest binary claim: Training data composition is only categorical (§2, Figure 2), the backbone is anonymized (Table 1), weights/API internals are closed, and user/customer data are excluded by assertion. Large-N internal suites and public benchmarks are valuable, but a journal-level claim of 0.0041% FPR / 0.3396% FNR and broad SOTA needs either open evaluation artifacts (full scored ID lists, prompts, generator versions, decode outputs) or an independent third-party re-score protocol. As written, the binary results are impressive but not scientifically auditable at the precision advertised.
minor comments (6)
  1. [Abstract; §5.1] Abstract AUROC 0.9916 is not clearly tied to a named corpus/split in §5; state the exact evaluation set used for that single number.
  2. [Figure 1; §4.1] Figure 1 is referenced as the architecture overview but the textual description in §4.1 is denser than the figure caption; label heads (segment 15-way, token 3-way, mixed 2-way, humanizer 4-way) explicitly on the figure.
  3. [§5.8 Table 12] Table 12 is very long and mixes incompatible native metrics; a short normalized summary (e.g., wins/ties vs next-best at each paper’s primary metric) would help readers without replacing the detailed appendix.
  4. [§5.6] §5.6 hypothesizes tokenization effects for Urdu/Persian FNR but provides no tokenizer analysis; either add a brief experiment or mark as speculation.
  5. [Abstract; §5] Typographical inconsistencies: “W e achieve”, “F alse”, “Pangram 4exhibits”, and similar spacing artifacts appear in the front matter and §5; clean for camera-ready.
  6. [AI Disclosure] AI Disclosure and Figure 6 (self-detection of the report) are fine for transparency but should not be read as validation; keep them out of the results narrative.

Circularity Check

2 steps flagged

Binary FPR/FNR/AUROC claims are externally grounded; mixed-authorship “AI-Assisted recall” partly measures recovery of the authors’ Soft N-Grams labeler used as training supervision.

specific steps
  1. self definitional [§3.5 Soft N-Grams Labeling; Algorithm 1; §4.1 segment f_AI buckets]
    "We do not have ground-truth tokenwise labels for mixed text a priori. Therefore, we compute tokenwise labels for AI-assisted text using the method described in Algorithm 1, which we call Soft N-Grams Labeling. ... f_AI ← (0.5 C_AA + C_AG) / D ... We discretize f_AI ∈ [0,1] into 15 ordered buckets and train a segment-level classification head"

    Token/segment supervision for the mixed classes is defined by the authors’ clause-match procedure (lexical n-grams then embeddings vs hidden source S; AI-assisted weighted 0.5). The model’s mixed-authorship heads are trained to reproduce that definition. Claiming the heads “detect” AI-assisted provenance therefore partly restates the labeling rule rather than an external ground truth—especially since the paper notes per-token AI-assisted is nebulous and S is unavailable at test time.

  2. fitted input called prediction [§5.5.1 Homogeneous Mixed Eval, WildChat Edits, Table 6]
    "Finally, we score all resulting edited texts using Soft N-Grams Labeling and filter our dataset to 4,826 examples with an AI fraction of between 0.25 and 0.75. We deem texts with this score to be AI-assisted by any reasonable standard... Pangram 4 classifies the majority of these texts as Mixed, achieving an AI-Assisted Recall of 65.7%"

    The evaluation set membership and the positive class for “AI-Assisted Recall” are determined by the same Soft N-Grams f_AI used to supervise training. Reporting high Mixed recall on texts pre-selected so the labeler says 0.25–0.75 AI fraction measures recovery of that synthetic target on similarly constructed edits, not independent verification of co-authorship. (Prompt/tool polish sets and OpAI-Bench are less circular; this table is the load-bearing mixed-recall number.)

full rationale

This is an empirical detector report, not a first-principles derivation. Overall FPR (human corpora), FNR (520k generations from named frontier models), public binary benchmarks, and adversarial/humanizer tests do not reduce to fitted parameters renamed as predictions: they compare model outputs to external human/AI provenance. The only material circularity is in the novel mixed-authorship half of the claim. Soft N-Grams (Alg. 1) is explicitly invented because “we do not have ground-truth tokenwise labels,” defines Human / AI-Assisted / AI-Generated and f_AI = (0.5 C_AA + C_AG)/D, and supplies both segment-bucket and tokenwise training targets. The key homogeneous recall set (WildChat Edits, Table 6) is then filtered to the same Soft N-Grams f_AI ∈ [0.25, 0.75], so “65.7% Mixed recall” is agreement with that labeler on held-out synthetic edits, not an independent oracle. External polish/edit-prompt sets and OpAI-Bench character fractions are more independent and do not fully close the loop. EditLens self-citation is normal prior architecture/data method, not a load-bearing uniqueness theorem. Score 3: one real self-definitional evaluation loop on the mixed novelty claim; central binary SOTA numbers remain non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

Headline detector metrics rest on operational definitions of AI text, synthetic training distributions, a custom mixed-label algorithm, and many postprocessing/calibration choices. The ledger separates standard ML practice from paper-specific labeling axioms and fitted decode knobs that load-bear on reported FPR/FNR and mixed recall.

free parameters (4)
  • Document decode thresholds (f_human≥0.90 → Human; f_AI≥0.80 → AI; else Mixed) = 0.90 / 0.80
    Rule-based map from continuous fractions to evaluation labels; directly defines reported FP/FN (§5.1).
  • AI-assisted weight in f_AI (0.5·C_AA + C_AG)/D and 15 ordered buckets = 0.5 weight; 15 buckets; S=512
    Hand-chosen half-credit for assisted characters and discretization used as segment supervision (§3.5, §4.1).
  • CRF/calibration parameters (α_c, w_c, δ_c, λ, γ, L_min) and mixed-head 15% rule = held-out fitted; L_min configurable; mixed if >15% non-dominant tokens
    Fitted on held-out calibration data; control smoothness, class priors, minimum segment length, and mixed-window target (§4.3, §4.1).
  • Humanizer loss weight 0.25 with stop-gradient; Stage-1/2 LoRA schedule = loss weight 0.25; stop-gradient probe
    Training hyperparameters chosen after ablations claiming FPR concentration in academic/ESL if unblocked (§4.1–4.2, Table 1–2).
axioms (5)
  • domain assumption AI-generated text := original substantial NL prose from an LLM on open-ended tasks, with output tokens > input tokens and length ≥50 words; factual short answers and pure compression are out of scope.
    Scopes positive class and excludes many real LLM uses (§3.1); all FNR claims inherit this scope.
  • ad hoc to paper Clause-level Soft N-Grams: lexical match⇒Human, semantic-only match⇒AI-Assisted, no match⇒AI-Generated; deletions/reorders are non-edits; labels transferable without source at test time.
    Core supervision for mixed text (Algorithm 1, §3.5); not a community standard ground truth.
  • domain assumption Synthetic mirrors preserve topic while changing authorship enough that detectors should learn process not topic; near-verbatim mirrors discarded.
    Training distribution construction from Emi & Spero 2024 (§2.2).
  • domain assumption Under causal LMs, Repeat2 makes every supervised token see full window context without breaking next-token pretraining inductive bias enough to invalidate token labels.
    Justifies tokenwise head design via Leviathan et al. 2025 (§4.1).
  • standard math Standard supervised fine-tuning / LoRA on a pretrained MoE yields calibrated generalization to held-out generators, languages, and attacks at reported rates.
    Ordinary deep-learning inductive assumption underlying all metrics.
invented entities (2)
  • Soft N-Grams Labeling (clause provenance labeler) no independent evidence
    purpose: Create tokenwise human / AI-assisted / AI-generated targets for homogeneously edited text without human span annotation.
    Introduced as Pangram 4’s labeling method distinct from EditLens embedding distance; depends on Haiku clause splits and embedding/n-gram thresholds.
  • Three-way token provenance + document humanizer 4-way probe as joint product taxonomy no independent evidence
    purpose: Express heterogeneous boundaries and homogeneous co-authorship in one pass; flag evasion-style rewrites.
    Product ontology claimed as first simultaneous homo/hetero detector (§3.3–3.4); external validity only via their eval sets.

pith-pipeline@v1.2.0-daily-grok45 · 36105 in / 4180 out tokens · 83152 ms · 2026-07-30T11:00:48.870953+00:00 · methodology

0 comments
read the original abstract

We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs. We achieve an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%. In addition to its increased overall accuracy compared with Pangram 3, Pangram 4 exhibits superior out-of-distribution generalization and robustness to adversarial attacks. Another novel contribution of Pangram 4 is its improved ability to distinguish fine-grained edits and mixed AI-human co-authored text. We demonstrate improvements to both boundary detection tasks and the detection of interleaved AI assistance. Finally, we report metrics on standard AI detection benchmarks showing that Pangram 4 achieves state-of-the-art performance on the AI text detection task across a wide variety of settings and domains.

Review history (2 revisions) →

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

64 extracted references · 10 canonical work pages

  1. [1]

    and Zhang, Hao and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric P. and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , title =. 2023 , publisher =

    2023

  2. [2]

    Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Papers)

    Russell, Jenna and Karpinska, Marzena and Akinode, Destiny and Zhou, James and Thai, Katherine and Emi, Bradley and Spero, Max and Iyyer, Mohit. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Papers). 2026. doi:10.18653/v1/2026.acl-long.663

    work page doi:10.18653/v1/2026.acl-long.663 2026

  3. [3]

    Findings of the Association for Computational Linguistics: ACL 2024

    Thompson, Brian and Dhaliwal, Mehak and Frisch, Peter and Domhan, Tobias and Federico, Marcello. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.103

    work page doi:10.18653/v1/2024.findings-acl.103 2024

  4. [4]

    2604.26965 , archivePrefix=

    Jonas Dolezal and Sawood Alam and Mark Graham and Maty Bohacek , year=. 2604.26965 , archivePrefix=

    Pith/arXiv arXiv

  5. [5]

    2025 , month = jan, url =

    2025

  6. [6]

    and Zhang, Edwin , title =

    Kalai, Adam Tauman and Nachum, Ofir and Vempala, Santosh S. and Zhang, Edwin , title =. Nature , year =. doi:10.1038/s41586-026-10549-w , url =

    doi:10.1038/s41586-026-10549-w

  7. [7]

    , title =

    Bai, Xuechunzi and Wang, Angelina and Sucholutsky, Ilia and Griffiths, Thomas L. , title =. Proceedings of the National Academy of Sciences , year =. doi:10.1073/pnas.2416228122 , url =

    doi:10.1073/pnas.2416228122

  8. [8]

    and Seamans, Rob , title =

    Raj, Manav and Berg, Justin M. and Seamans, Rob , title =. Journal of Experimental Psychology: General , year =. doi:10.1037/xge0001889 , url =

    work page doi:10.1037/xge0001889

  9. [9]

    2025 , month = nov, url =

    2025

  10. [10]

    2026 , month = jul, url =

    2026

  11. [11]

    2026 , month = jun, note =

    2026

  12. [12]

    2026 , month = jul, note =

    2026

  13. [13]

    2025 , month = oct, note =

    2025

  14. [14]

    2026 , month = feb, note =

    2026

  15. [15]

    2026 , month = may, note =

    2026

  16. [16]

    2026 , month = apr, note =

    2026

  17. [17]

    2024 , month = dec, note =

    2024

  18. [18]

    2025 , month = aug, note =

    2025

  19. [19]

    2026 , month = mar, note =

    2026

  20. [20]

    Scientific Reports , volume =

    Porter, Brian and Machery, Edouard , title =. Scientific Reports , volume =. 2024 , doi =

    2024

  21. [21]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Dugan, Liam and Hwang, Alyssa and Trhl \'i k, Filip and Zhu, Andrew and Ludan, Josh Magnus and Xu, Hainiu and Ippolito, Daphne and Callison-Burch, Chris. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.674

    doi:10.18653/v1/2024.acl-long.674 2024

  22. [22]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Li, Yafu and Li, Qintong and Cui, Leyang and Bi, Wei and Wang, Zhilin and Wang, Longyue and Yang, Linyi and Shi, Shuming and Zhang, Yue. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.3

    doi:10.18653/v1/2024.acl-long.3 2024

  23. [23]

    Guo, Biyang and Zhang, Xin and Wang, Ziyuan and Jiang, Minqi and Nie, Jinran and Ding, Yuxuan and Yue, Jianwei and Wu, Yupeng. 2023

    2023

  24. [24]

    Paetzold , year=

    Chenjun Li and Cheng Wan and Johannes C. Paetzold , year=. 2605.06903 , archivePrefix=

    Pith/arXiv arXiv

  25. [25]

    Oliveira and Lei Li , booktitle=

    Andre Vicente Duarte and Brian Tufts and Aditya Oke and Fei Fang and Arlindo L. Oliveira and Lei Li , booktitle=. 2026 , url=

    2026

  26. [26]

    2026 , url=

    Sungduk Yu and Man Luo and Avinash Madasu and Vasudev Lal and Phillip Howard , booktitle=. 2026 , url=

    2026

  27. [27]

    2026 , url=

    Rounak Saha and Gurusha Juneja and Dayita Chaudhuri and Naveeja Sajeevan and Nihar B Shah and Danish Pruthi , booktitle=. 2026 , url=

    2026

  28. [28]

    Findings of the Association for Computational Linguistics: NAACL 2024

    Zhang, Qihui and Gao, Chujie and Chen, Dongping and Huang, Yue and Huang, Yixin and Sun, Zhenyang and Zhang, Shilin and Li, Weiye and Fu, Zhengyan and Wan, Yao and Sun, Lichao. Findings of the Association for Computational Linguistics: NAACL 2024. 2024. doi:10.18653/v1/2024.findings-naacl.29

    doi:10.18653/v1/2024.findings-naacl.29 2024

  29. [29]

    Wong and Shu Yang and Xinyi Yang and Yulin Yuan and Lidia S

    Junchao Wu and Runzhe Zhan and Derek F. Wong and Shu Yang and Xinyi Yang and Yulin Yuan and Lidia S. Chao , booktitle=. 2024 , url=

    2024

  30. [30]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)

    Verma, Vivek and Fleisig, Eve and Tomlin, Nicholas and Klein, Dan. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.naacl-long.95

    doi:10.18653/v1/2024.naacl-long.95 2024

  31. [31]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =

    Macko, Dominik and Moro, Robert and Uchendu, Adaku and Lucas, Jason and Yamashita, Michiharu and Pikuliak, Mat. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =. doi:10.18653/v1/2023.emnlp-main.616 , url =

    doi:10.18653/v1/2023.emnlp-main.616 2023

  32. [32]

    Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , pages =

    He, Xinlei and Shen, Xinyue and Chen, Zeyuan and Backes, Michael and Zhang, Yang , title =. Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security , pages =. 2024 , isbn =. doi:10.1145/3658644.3670344 , abstract =

    arXiv 2024

  33. [33]

    Patterns , volume =

    Liang, Weixin and Yuksekgonul, Mert and Mao, Yining and Wu, Eric and Zou, James , title =. Patterns , volume =. 2023 , doi =

    2023

  34. [35]

    2026 , eprint =

    Quaremba, Gerrit and Black, Elizabeth and Vrande. 2026 , eprint =

    2026

  35. [36]

    Wallace , year=

    Chantal Shaib and Tuhin Chakrabarty and Diego Garcia-Olano and Byron C. Wallace , year=. 2509.19163 , archivePrefix=

    arXiv

  36. [37]

    2402.14873 , archivePrefix=

    Bradley Emi and Max Spero , year=. 2402.14873 , archivePrefix=

    Pith/arXiv arXiv

  37. [38]

    2026 , url=

    Thai, Katherine and Emi, Bradley and Masrour, Elyas and Iyyer, Mohit , booktitle=. 2026 , url=

    2026

  38. [39]

    Findings of the Association for Computational Linguistics: ACL 2025

    Saha, Shoumik and Feizi, Soheil. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.1303

    work page doi:10.18653/v1/2025.findings-acl.1303 2025

  39. [40]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Drayson, George and Yilmaz, Emine and Lampos, Vasileios. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.1506

    work page doi:10.18653/v1/2025.emnlp-main.1506 2025

  40. [41]

    Proceedings of the ACM Web Conference 2026 , pages =

    Yu, Hongyeon and Kim, Dongchan and Kim, Young-Bum , title =. Proceedings of the ACM Web Conference 2026 , pages =. 2026 , isbn =. doi:10.1145/3774904.3792955 , abstract =

    arXiv 2026

  41. [42]

    Claessens, Scott and Veitch, Pierce and Everett, Jim A. C. , title =. Computers in Human Behavior , volume =. 2026 , pages =. doi:10.1016/j.chb.2025.108894 , url =

    arXiv 2026

  42. [43]

    Proksch, S. and Sch. Frontiers in Artificial Intelligence , volume =. 2024 , pages =. doi:10.3389/frai.2024.1412710 , url =

    arXiv 2024

  43. [44]

    2604.03136 , archivePrefix=

    Jenna Russell and Rishanth Rajendhran and Chau Minh Pham and Mohit Iyyer and John Wieting , year=. 2604.03136 , archivePrefix=

    Pith/arXiv arXiv

  44. [45]

    Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)

    Artemova, Ekaterina and Lucas, Jason S and Venkatraman, Saranya and Lee, Jooyoung and Tilga, Sergei and Uchendu, Adaku and Mikhailov, Vladislav. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.naacl-long.357

    doi:10.18653/v1/2025.naacl-long.357 2025

  45. [46]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Russell, Jenna and Karpinska, Marzena and Iyyer, Mohit. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.267

    doi:10.18653/v1/2025.acl-long.267 2025

  46. [47]

    Second Conference on Language Modeling , year=

    Guilherme Penedo and Hynek Kydl. Second Conference on Language Modeling , year=

  47. [48]

    Advances in Neural Information Processing Systems , volume=

    Guilherme Penedo and Hynek Kydl. Advances in Neural Information Processing Systems , volume=. 2024 , url=

    2024

  48. [49]

    , title =

    Dice, Lee R. , title =. Ecology , year =. doi:10.2307/1932409 , url =

    doi:10.2307/1932409

  49. [50]

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing

    Ng, Jun-Ping and Abrecht, Viktoria. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. 2015. doi:10.18653/v1/D15-1222

    work page doi:10.18653/v1/d15-1222 2015

  50. [51]

    2019 , pages =

    Reimers, Nils and Gurevych, Iryna , booktitle =. 2019 , pages =. doi:10.18653/v1/D19-1410 , url =

    doi:10.18653/v1/d19-1410 2019

  51. [52]

    2023 , url =

    Krishna, Kalpesh and Song, Yixiao and Karpinska, Marzena and Wieting, John and Iyyer, Mohit , booktitle =. 2023 , url =

    2023

  52. [53]

    2512.14982 , archivePrefix=

    Yaniv Leviathan and Matan Kalman and Yossi Matias , year=. 2512.14982 , archivePrefix=

    arXiv

  53. [54]

    2501.03437 , archivePrefix=

    Elyas Masrour and Bradley Emi and Max Spero , year=. 2501.03437 , archivePrefix=

    Pith/arXiv arXiv

  54. [55]

    2025 , month = sep, doi =

    Jabarian, Brian and Imas, Alex , institution =. 2025 , month = sep, doi =

    2025

  55. [56]

    2025 , eprint =

    Gehring, Lukas and Paa. 2025 , eprint =

    2025

  56. [57]

    and Postma, Darius and Hickerson, Don and McGaughran, James and Khuat, Huy Q

    Perkins, Mike and Roe, Jasper and Vu, Binh H. and Postma, Darius and Hickerson, Don and McGaughran, James and Khuat, Huy Q. , title =. International Journal of Educational Technology in Higher Education , year =. doi:10.1186/s41239-024-00487-w , url =

    work page doi:10.1186/s41239-024-00487-w

  57. [58]

    International Journal for Educational Integrity , year =

    Van Vlasselaer, Marijke and Van Droogenbroeck, Filip and Spruyt, Bram , title =. International Journal for Educational Integrity , year =. doi:10.1007/s40979-026-00226-w , url =

    work page doi:10.1007/s40979-026-00226-w

  58. [59]

    Al Khatib and Hao Li and Salman Khan and Zhiqiang Shen , year=

    Sondos Mahmoud Bsharat and Jiacheng Liu and Xiaohan Zhao and Tianjun Yao and Xinyi Shang and Yi Tang and Jiacheng Cui and Ahmed Elhagry and Salwa K. Al Khatib and Hao Li and Salman Khan and Zhiqiang Shen , year=. 2606.06481 , archivePrefix=

    Pith/arXiv arXiv

  59. [60]

    2606.04906 , archivePrefix=

    Nils Dycke and Marina Sakharova and Nico Daheim and Iryna Gurevych , year=. 2606.04906 , archivePrefix=

    Pith/arXiv arXiv

  60. [61]

    2024 , url=

    Wenting Zhao and Xiang Ren and Jack Hessel and Claire Cardie and Yejin Choi and Yuntian Deng , booktitle=. 2024 , url=

    2024

  61. [62]

    International Journal of Learner Corpus Research , year =

    Crossley, Scott and Tian, Yu and Baffour, Perpetual and Franklin, Alex and Kim, Youngmeen and Morris, Wesley and Benner, Meg and Picou, Aigner and Boser, Ulrich , title =. International Journal of Learner Corpus Research , year =. doi:10.1075/ijlcr.22026.cro , note =

    doi:10.1075/ijlcr.22026.cro

  62. [63]

    2020 , version =

    Juffs, Alan and Han, Na-Rae and Naismith, Ben , title =. 2020 , version =. doi:10.5281/zenodo.3991977 , url =

    work page doi:10.5281/zenodo.3991977 2020

  63. [64]

    English Teaching , year =

    Ishikawa, Shin'ichiro , title =. English Teaching , year =. doi:10.15858/engtea.74.4.201912.153 , url =

    work page doi:10.15858/engtea.74.4.201912.153

  64. [65]

    2025 , howpublished=

    2025