Pith. sign in

REVIEW 4 major objections 4 minor 34 references

The paper claims that source-conditioned description-length gain—the drop in a frozen language model's codelength for a suspicious document when a candidate source is added—is a unified, training-free signal for detecting generative plagiar

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

2026-08-05 10:56 UTC pith:3MNIMN2R

load-bearing objection A genuinely useful empirical package built on a simple likelihood-ratio identity, with a real overstatement about what is and isn't training-free. the 4 major comments →

arxiv 2608.03859 v1 pith:3MNIMN2R submitted 2026-08-04 cs.CL cs.AI

Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking

Peijia Guo , Wenxuan Xie , ZiGuang Li , Ming Li This is my paper
classification cs.CL cs.AI
keywords generative plagiarism detectionsource-conditioned description-length gaincodelengthtoken-level log-likelihood ratiocandidate-source rerankingLLM plagiarismsame-topic confoundingtraining-free detection
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 paper tries to establish that the right way to detect generative plagiarism is to ask a different question: not 'does the suspicious text resemble a source?' but 'does the source help a fixed language model predict the suspicious text?' The proposed measure, Source-Conditioned Description-Length Gain (SCDG), subtracts the model's codelength for the document conditioned on the source from its unconditional codelength; a larger drop means the source supplies information that makes the text more predictable. On pairwise detection and multi-source retrieval benchmarks, this training-free signal outperforms lexical overlap, embedding similarity, and prompted-LLM judgments, reaching about 0.94 F1 in pairwise tests and about 0.83 ranking quality (nDCG@10) in retrieval. A same-topic stress test shows that the gain distribution, once calibrated, stays close to zero for independently written documents about the same event. The paper's contribution is therefore a directional, token-decomposable measure of source dependence that works where surface similarity fails.

Core claim

The central claim is that the quantity Dθ(P←S) = L0θ(P) − LSθ(P) = log[pθ(P|S)/pθ(P)] measures how much predictive evidence a candidate source S contributes to a suspicious document P. Here L0 and LS are the negative log-likelihoods (codelengths) assigned by one frozen autoregressive language model to the identical target token sequence P, with or without S in the context. Because this difference decomposes exactly into per-token log-likelihood ratios, it can be aggregated over all tokens or over the top-gain tokens, making it sensitive to sparse, localized reuse. The paper validates the mechanism with a controlled intervention in which the proportion of retained source evidence is varied an

What carries the argument

The central object is Eq. (3): Dθ(P←S) = log[pθ(P|S)/pθ(P)], a model-relative conditional log-likelihood ratio, interpreted as a reduction in description length. In an autoregressive language model, expected codelength equals negative log-probability, so conditioning on S produces a per-token gain gt = log[pθ(pt | S, p<t)/pθ(pt | p<t)]; these gains sum exactly to the document-level gain. The mechanism does the work of turning 'source reuse' into an operational quantity: reused content becomes easier to predict when S is in the context, yielding positive gains at precisely the positions where source-derived evidence survives rewriting. The same machinery supports sparse aggregation (averaging

Load-bearing premise

The drop in codelength when a candidate source is added to the model's context is caused mainly by genuine reuse of that source, not merely by shared topic, style, or event.

What would settle it

Run the calibrated SCDG classifier on pairs of documents that are independently written but cover the same event (for example, two news agency reports of the same game from different desks). If a substantial fraction of such pairs receive positive source-reuse decisions, the selective-sensitivity claim fails. A sharper test: in the controlled source-retention setting, replace the retained source evidence with equal-length, same-topic distractors and check whether the gain still rises monotonically; if it does, the gain is tracking topical overlap rather than source reuse.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • One scalar, computed without any training, serves both as a plagiarism classifier for document pairs and as a reranker for candidate sources retrieved from a large corpus.
  • Because the gain decomposes per token, the signal can show where in the suspicious document the source contributes predictive evidence, supporting passage-level attribution.
  • The signal works across three different frozen 8-billion-parameter models, so it does not depend on one model family.
  • Reranking fixed first-stage candidates raises ranking quality at the head of the list by over 0.13 on the 2026 benchmark, without changing the candidate set.
  • Calibrating the token-gain distribution rather than using the mean brings the same-topic false-positive proxy from 62.9% down to 0.125% on a news-domain stress set.

Where Pith is reading between the lines

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

  • The token-level gain profile could be converted into an alignment map between a suspicious document and a candidate source, giving an interpretable, passage-level plagiarism report; the paper lists localization only as future work.
  • Because the gain is relative to a particular frozen model, the ranking may inherit that model's priors; in low-resource or specialized domains the source-specific component could shrink, so domain-adapted backends are a natural stress test.
  • A testable extension is to check whether SCDG can distinguish a document paraphrased from S from one merely inspired by S's topic: tuning the calibration on pairs that share only an abstract or a citation, rather than reused passages, would probe the boundary of 'source reuse'.
  • The 0.125% same-topic false-positive figure depends on a logistic calibrator trained on a cluster-disjoint news split; whether this calibration transfers across corpora and languages is an open question that determines practical deployment.

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 / 4 minor

Summary. The paper introduces Source-Conditioned Description-Length Gain (SCDG), defined in Eq. (3) as the difference between a frozen language model's unconditional and source-conditioned description lengths: Dθ(P←S) = L0θ(P) − LSθ(P) = log pθ(P|S)/pθ(P). The claimed contribution is a training-free, token-decomposable, directional signal that identifies generative source reuse and supports both pairwise plagiarism detection and candidate-source reranking. The authors evaluate SCDG on PAN 2025-derived pairwise detection, PAN 2026 multi-source retrieval, a controlled source-retention intervention, and a Multi-News same-topic stress test. Reported results include PAN25 F1 0.9433 and PAN26 nDCG@10 0.8315, both exceeding the selected baselines. The paper also introduces DAAC, a rank-based retrieval fusion method, and a 'gain-distribution logistic' variant that achieves a low proxy false-positive rate on the same-topic stress set.

Significance. If the central claims held, SCDG would be a valuable new tool for generative plagiarism detection: it operationalizes the intuitive idea that a reused source should make a suspicious document more predictable, provides a principled token-level decomposition, and is evaluated under controlled partitions and matched candidate sets. The source-retention intervention is a commendable attempt to test sensitivity to genuine source evidence, and the reranking comparisons on fixed Top-1000 candidates are well controlled. However, the same-topic confounding experiments reveal a load-bearing gap between the raw SCDG signal and the robustness claim: the raw average score has a 62.9% proxy false-positive rate on same-topic independent pairs, and only a trained logistic over the token-gain distribution reaches 0.125%. This undermines the paper's 'training-free' and 'selectively sensitive' framing, and the central contribution needs substantial reframing or additional principled debiasing before the claims can be accepted as stated.

major comments (4)
  1. [Robustness to Same-Topic Confounding, Table 3 and Eq. (3)] The raw Average SCDG row reports a 62.876% proxy false-positive rate on same-topic, same-event Multi-News pairs, while the gain-distribution logistic reaches 0.1250%. This is direct evidence that the raw likelihood-ratio signal in Eq. (3) does not distinguish source reuse from shared topic or event information. The controlled source-retention study (Figure 3) demonstrates sensitivity to retained source evidence but does not test specificity against topical overlap. Since the abstract and contributions describe SCDG as a 'training-free' and 'source-specific content reuse' signal, this discrepancy is load-bearing. The paper needs either a principled way to remove topical confounding from the raw score, or a clear statement that the robust claim applies only to the trained logistic variant, with the raw signal characterized as topic-sensitive.
  2. [Pairwise Detection and Candidate-Source Reranking, Eq. (7) and Contributions] The 'training-free' claim is not precise. Pairwise detection in Eq. (7) requires a decision threshold τ calibrated on validation data, and the robust same-topic performance requires full logistic training on the token-gain distribution. This is more than hyperparameter selection or a trivial threshold. The contribution bullet says SCDG works 'without task-specific parameter updates', but the paper's own most robust variant is trained. The authors should clearly separate the raw score, the threshold-calibrated detector, and the trained logistic calibrator, and state which claims apply to which variant.
  3. [Table 1: Pair-Level Binary Classification] The main pairwise results are reported as per-metric medians over 20 hash-locked test partitions, but no confidence intervals, variance, or significance tests are provided. The claim that SCDG 'outperforms all baselines' would be more convincing with some measure of stability, especially because the reported differences between some configurations (e.g., Llama Average vs. Top-q F1) are small. I recommend adding error bars or a distribution summary.
  4. [Table 3: Multi-News Stress Protocol] The proxy false-positive-rate comparison requires specification of how each method's decision threshold is set. For Average SCDG, what threshold was used and how was it calibrated? Without this, the 62.876% figure could be the result of an arbitrarily chosen operating point, and the comparison between raw SCDG and the logistic is confounded by the amount of supervision each method receives. The paper should report the threshold-selection procedure and, ideally, ROC-style curves for the Multi-News pairs.
minor comments (4)
  1. [Methodology, Eq. (5)–(6)] The same symbol Dθ(P←S) is used for the total token-sum gain in Eq. (5) and the length-normalized mean gain in Eq. (6). This is potentially confusing; a distinct notation (e.g., Δθ or an overbar) would clarify the subsequent threshold and ranking equations.
  2. [Experimental Setup, Implementation Details] The text refers to 'the supplement' for full implementation and computing-environment details, but the submitted manuscript does not include the supplement. Key hyperparameters mentioned in Table 2 (N=1000, α=.75) are not fully defined in the main text. Please either include the supplement or define these quantities in the main body.
  3. [Abstract] The abstract reports '0.92 Precision, 0.97 Recall, and 0.94 F1', which are rounded values consistent with the Llama back-end in Table 1, but the table also shows precision and recall vary across backends. It would help to identify which configuration the abstract numbers refer to.
  4. [Conclusions] The limitations paragraph mentions computational cost and English-centric evaluation, but it does not acknowledge the same-topic confounding issue visible in Table 3. A sentence noting that raw SCDG is topic-sensitive and that the robust classifier requires calibration would be appropriate.

Circularity Check

0 steps flagged

No significant circularity: SCDG is an external benchmark evaluation of a likelihood-ratio contrast; thresholds are validation-calibrated and self-citations are not load-bearing.

full rationale

The paper's central quantity Dθ(P←S)=L0θ(P)−LSθ(P)=log pθ(P|S)/pθ(P) (Eq. 3) is an explicit algebraic definition of a conditional log-likelihood ratio/PMI, and the authors openly state the contribution is operationalization, not the identity itself. All detection and reranking uses are evaluated against external benchmarks (PAN25 pairwise, PAN26 retrieval, Multi-News stress set), with thresholds τ, top-q aggregation, and DAAC fusion weights selected on validation/development partitions and fixed for test evaluation—standard practice, not fitting a prediction to the test set. The controlled source-retention study is an empirical sensitivity check, not a circular derivation. Self-citations (Li & Vitányi 2008 for Kolmogorov-complexity bounds; Li et al. 2025 for compression by LMs) are motivational context and do not carry the load; the likelihood-ratio identity is mathematically self-contained. One non-circular correctness caveat: Table 3 shows raw Average SCDG has a 62.9% proxy FPR on same-topic Multi-News pairs, and the 0.125% robustness result comes from a trained gain-distribution logistic, so the 'training-free' signal's topical robustness is weaker than the headline suggests; this is an overclaim/correctness issue, not circularity. No step in the derivation chain reduces by construction to its own input.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

SCDG and DAAC are scoring procedures, not new theoretical entities. The paper introduces no new particles, forces, dimensions, or ontological commitments. Its free parameters are standard calibration choices, and its axioms are largely background information theory plus an empirical premise about conditional likelihood and source reuse.

free parameters (5)
  • decision threshold tau = calibrated on validation splits, not stated numerically
    Used in Eq. (7) to convert SCDG score into a binary source-reuse decision. Selected on training portion of each PAN25 split, fixed for test.
  • top-q aggregation proportion q = top-2% on PAN26; selected per split on PAN25
    Eq. (6) aggregation set A is either all tokens or the top-q gains; the proportion is a free hyperparameter tuned on development data.
  • fusion weight lambda = selected on development queries
    Eq. (13) balances DAAC and SCDG scores in final ranking; value chosen on dev folds.
  • DAAC constants k and exponents = k=60; exponents 2
    Eqs. (9-11) use reciprocal rank with k=60 and squared BM25/abstract terms; these are hand-chosen without derivation.
  • gain-distribution logistic coefficients = fitted on Multi-News training pairs
    The 0.125% proxy FPR robustness result comes from a supervised logistic classifier on gain-distribution features, not from the raw SCDG score.
axioms (5)
  • standard math The negative log-probability of a sequence under a fixed model is a valid model-relative codelength (Shannon/MDL).
    Invoked in 'Theoretical Foundation' as the basis for interpreting likelihood differences as description-length gain.
  • domain assumption A frozen autoregressive LM's token probabilities provide a reliable generative model for the target document.
    All SCDG scores depend on p_theta from Qwen3, Llama-3.1, or Ministral; no validation that the model's probability estimates are well-calibrated outside these benchmarks.
  • ad hoc to paper Lower conditional codelength when S is supplied is caused by source reuse, not merely shared topic or style.
    This is the design premise of SCDG. The controlled retention study supports it on 300 pairs, but the Multi-News stress test shows raw mean gain is not robust, so the premise is only partially validated.
  • domain assumption Same-cluster Multi-News article pairs are treated as non-reuse cases.
    Used to define proxy false-positive rate in Table 3; if any same-cluster pairs contain genuine source reuse, the stated FPR is biased.
  • domain assumption PAN25 labels and PAN26 qrels are accurate ground truth for source reuse.
    All headline numbers are measured against these benchmarks; errors in qrels would propagate.

pith-pipeline@v1.4.0-daily-deepseek-medium · 12101 in / 12668 out tokens · 118278 ms · 2026-08-05T10:56:41.248906+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking." pith.science (2026). https://pith.science/paper/3MNIMN2R

@misc{pith2026260803859,
  author       = {Pith},
  title        = {Pith review of: Beyond Representational Similarity: Source-Conditioned Description-Length Gain for Generative Plagiarism Detection and Candidate Source Reranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MNIMN2R}},
  note         = {Machine review of arXiv:2608.03859}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) pose challenges to academic integrity and peer review. Yet generative plagiarism detection remains an underexplored and largely unresolved challenge. Prior work on LLM-generated-text detection targets AI involvement, which may be permissible, rather than source reuse, while similarity-based methods struggle after extensive rewriting and multi-source synthesis. Motivated by the description-length view of probabilistic prediction, in which relevant side information can reduce a target sequence's code length, we introduce Source-Conditioned Description-Length Gain (SCDG), a directional, training-free framework that contrasts a frozen language model's description length of a suspicious document $P$ with and without a candidate source $S$. This contrast yields token-level log-likelihood gains that measure the incremental predictive evidence supplied by $S$. We evaluate SCDG on the PAN at CLEF benchmarks for generative plagiarism. On a PAN 2025-derived pairwise benchmark, SCDG achieves 0.92 Precision, 0.97 Recall, and 0.94 F1, outperforming all baselines; on PAN 2026's multi-source retrieval task, it reaches 0.83 nDCG@10 and 0.96 Recall@100, surpassing all baselines. On a same-topic, same-event Multi-News test, the calibrated gain-distribution SCDG classifier predicts source reuse for only $0.125\%$ of pairs, supporting robustness to topical overlap under this evaluation protocol. These results establish SCDG as a unified and token-decomposable signal for source-specific content reuse under extensive transformation.

Figures

Figures reproduced from arXiv: 2608.03859 by Ming Li, Peijia Guo, Wenxuan Xie, Ziguang Li.

Figure 1
Figure 1. Figure 1: SCDG distinguishes source-conditioned predictive [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Source-Conditioned Description-Length Gain (SCDG). Given a suspicious document [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean length-normalized description-length gain [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] 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

34 extracted references · 21 canonical work pages

  1. [1]

    , title =

    Shannon, Claude E. , title =. The Bell System Technical Journal , year =

  2. [2]

    Automatica , year =

    Rissanen, Jorma , title =. Automatica , year =

    work page

  3. [3]

    MacKay, David J. C. , title =. 2003 , isbn =

    work page 2003

  4. [4]

    , title =

    Mahoney, Matthew V. , title =. Proceedings of the Sixteenth National Conference on Artificial Intelligence , year =

    work page

  5. [5]

    Language Modeling Is Compression , booktitle =

    Del. Language Modeling Is Compression , booktitle =. 2024 , url =

    work page 2024

  6. [6]

    An Introduction to Kolmogorov Complexity and Its Applications , edition =

    Li, Ming and Vit. An Introduction to Kolmogorov Complexity and Its Applications , edition =. 2008 , doi =

    work page 2008

  7. [7]

    An Evaluation Framework for Plagiarism Detection , booktitle =

    Potthast, Martin and Stein, Benno and Barr. An Evaluation Framework for Plagiarism Detection , booktitle =. 2010 , address =

    work page 2010

  8. [8]

    Overview of the 3rd International Competition on Plagiarism Detection , booktitle =

    Martin Potthast and Andreas Eiselt and Alberto Barr. Overview of the 3rd International Competition on Plagiarism Detection , booktitle =. 2011 , address =

    work page 2011

  9. [9]

    Working Notes for CLEF 2013 Conference , series =

    Martin Potthast and Matthias Hagen and Tim Gollub and Martin Tippmann and Johannes Kiesel and Paolo Rosso and Efstathios Stamatatos and Benno Stein , title =. Working Notes for CLEF 2013 Conference , series =. 2013 , publisher =

    work page 2013

  10. [10]

    Working Notes for CLEF 2014 Conference , series =

    Martin Potthast and Matthias Hagen and Anna Beyer and Matthias Busse and Martin Tippmann and Paolo Rosso and Benno Stein , title =. Working Notes for CLEF 2014 Conference , series =. 2014 , publisher =

    work page 2014

  11. [11]

    Academic Plagiarism Detection: A Systematic Literature Review , journal =

    Tom. Academic Plagiarism Detection: A Systematic Literature Review , journal =. 2019 , doi =

    work page 2019

  12. [12]

    Proceedings of the 2021 ACM/IEEE Joint Conference on Digital Libraries , pages =

    Jan Philip Wahle and Terry Ruas and Norman Meuschke and Bela Gipp , title =. Proceedings of the 2021 ACM/IEEE Joint Conference on Digital Libraries , pages =. 2021 , publisher =. doi:10.1109/JCDL52503.2021.00065 , url =

    arXiv 2021

  13. [13]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages =

    Jan Philip Wahle and Terry Ruas and Frederic Kirstein and Bela Gipp , title =. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages =. 2022 , address =. doi:10.18653/v1/2022.emnlp-main.62 , url =

    work page doi:10.18653/v1/2022.emnlp-main.62 2022

  14. [14]

    Proceedings of the ACM Web Conference 2023 , pages =

    Jooyoung Lee and Thai Le and Jinghui Chen and Dongwon Lee , title =. Proceedings of the ACM Web Conference 2023 , pages =. 2023 , publisher =. doi:10.1145/3543507.3583199 , url =

    arXiv 2023

  15. [15]

    Jooyoung Lee and Toshini Agrawal and Adaku Uchendu and Thai Le and Jinghui Chen and Dongwon Lee , title =. 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) , pages =. 2025 , address =. doi:10.18653/v1/2025.naacl-long.384 , url =

    doi:10.18653/v1/2025.naacl-long.384 2025

  16. [16]

    Overview of the Plagiarism Detection Task at

    Andr. Overview of the Plagiarism Detection Task at. Working Notes of the Conference and Labs of the Evaluation Forum (CLEF 2025) , series =. 2025 , publisher =

    work page 2025

  17. [17]

    Overview of

    Janek Bevendorff and Maik Fr. Overview of. 2026 , eprint =. doi:10.48550/arXiv.2602.09147 , url =

    work page doi:10.48550/arxiv.2602.09147 2026

  18. [18]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    Xianjun Yang and Liangming Pan and Xuandong Zhao and Haifeng Chen and Linda Ruth Petzold and William Yang Wang and Wei Cheng , title =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , address =. doi:10.18653/v1/2024.findings-emnlp.572 , url =

    doi:10.18653/v1/2024.findings-emnlp.572 2024

  19. [19]

    Manning and Chelsea Finn , title =

    Eric Mitchell and Yoonho Lee and Alexander Khazatsky and Christopher D. Manning and Chelsea Finn , title =. Proceedings of the 40th International Conference on Machine Learning , series =. 2023 , publisher =

    work page 2023

  20. [20]

    The Twelfth International Conference on Learning Representations , year =

    Guangsheng Bao and Yanbin Zhao and Zhiyang Teng and Linyi Yang and Yue Zhang , title =. The Twelfth International Conference on Learning Representations , year =

    work page

  21. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Jiaqi Chen and Xiaoye Zhu and Tianyang Liu and Ying Chen and Xinhui Chen and Yiwen Yuan and Chak Tou Leong and Zuchao Li and Long Tang and Lei Zhang and Chenyu Yan and Guanghao Mei and Jie Zhang and Lefei Zhang , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    work page 2025

  22. [22]

    Working Notes for CLEF 2012 Conference , editor =

    Overview of the 4th International Competition on Plagiarism Detection , author =. Working Notes for CLEF 2012 Conference , editor =. 2012 , address =

    work page 2012

  23. [23]

    The Probabilistic Relevance Framework:

    Robertson, Stephen and Zaragoza, Hugo , journal =. The Probabilistic Relevance Framework:. 2009 , publisher =. doi:10.1561/1500000019 , url =

    doi:10.1561/1500000019 2009

  24. [24]

    doi:10.48550/arXiv.2412.03223 , url =

    Choi, Chanyeol and Kim, Junseong and Lee, Seolhwa and Kwon, Jihoon and Gu, Sangmo and Kim, Yejin and Cho, Minkyung and Sohn, Jy-yong , year =. doi:10.48550/arXiv.2412.03223 , url =. 2412.03223 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.03223

  25. [25]

    Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =

    Chen, Jianlyu and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng , editor =. Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =. doi:10.18653/v1/2024.findings-acl.137 , url =

    doi:10.18653/v1/2024.findings-acl.137 2024

  26. [26]

    The Second Text REtrieval Conference (

    Combination of Multiple Searches , author =. The Second Text REtrieval Conference (. 1994 , address =

    work page 1994

  27. [27]

    Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval , series =

    Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods , author =. Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval , series =. 2009 , address =. doi:10.1145/1571941.1572114 , url =

    arXiv 2009

  28. [28]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , month = jul, year =

    Fabbri, Alexander and Li, Irene and She, Tianwei and Li, Suyi and Radev, Dragomir , editor =. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , month = jul, year =. doi:10.18653/v1/P19-1102 , url =

    doi:10.18653/v1/p19-1102

  29. [29]

    2026 , month = jun, publisher =

    Greiner-Petter, Andr. 2026 , month = jun, publisher =. doi:10.5281/zenodo.20590774 , url =

    work page doi:10.5281/zenodo.20590774 2026

  30. [30]

    2025 , month = mar, publisher =

    Greiner-Petter, Andr. 2025 , month = mar, publisher =. doi:10.5281/zenodo.14969012 , url =

    work page doi:10.5281/zenodo.14969012 2025

  31. [31]

    arXiv preprint arXiv:2505.09388 , year =

    Qwen3 Technical Report , author =. arXiv preprint arXiv:2505.09388 , year =

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2407.21783 , year =

    The Llama 3 Herd of Models , author =. arXiv preprint arXiv:2407.21783 , year =

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2601.08584 , year =

    Ministral 3 , author =. arXiv preprint arXiv:2601.08584 , year =

    Pith/arXiv arXiv

  34. [34]

    Nature Machine Intelligence , volume =

    Lossless data compression by large models , author =. Nature Machine Intelligence , volume =. 2025 , doi =

    work page 2025