pith. sign in

arxiv: 2606.17582 · v1 · pith:BK4EH7MOnew · submitted 2026-06-16 · 💻 cs.DB

Collaborative Large and Small Language Models for Accurate and Scalable Data Repair

Qian Chen , Jianwei Wang , Wenjie Zhang This is my paper

Pith reviewed 2026-06-26 22:22 UTC · model grok-4.3

classification 💻 cs.DB
keywords data repairdata cleaninglarge language modelssmall language modelsexpectation-maximisationmodel collaborationconfidence calibration
0
0 comments X

The pith

LasRepair++ pairs an LLM instructor with an SLM corrector, refines context via EM, and applies column-calibrated confidence to improve data repair.

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

The paper addresses data repair by having a large language model select a global repair context that guides a smaller language model in correcting errors. It extends this into an EM procedure that alternates between updating the corrector and refining the context, then adds weighting to reduce the impact of uncertain repairs. Experiments show the full method raises average F1 scores by 18.1 percent over the best prior approach on real datasets. The work aims to combine the reasoning strength of large models with the efficiency of small ones while limiting noise from raw context or uncertain outputs.

Core claim

LasRepair uses an LLM to select global repair context that guides an SLM corrector; LasRepair+ casts the process as an EM procedure that alternates E-step parameter updates with M-step context refinement; LasRepair++ further weights rows by column-calibrated model confidence to down-weight unreliable repairs. Theoretical analysis proves the EM procedure and weighting are effective, and experiments on real-world datasets show LasRepair++ achieves an average 18.1% F1-score improvement over the strongest baseline.

What carries the argument

The LasRepair framework, in which the LLM instructor selects a global repair context to guide the SLM corrector, extended by an EM alternation for context refinement and column-calibrated confidence weighting to suppress unreliable rows.

If this is right

  • The EM-style alternation between corrector updates and context refinement raises repair quality.
  • Column-calibrated confidence weighting reduces the effect of uncertain model outputs on the final repairs.
  • Using the SLM for the actual correction step improves efficiency while the LLM supplies only high-level guidance.
  • The theoretical guarantees on the EM procedure and weighting hold under the stated model assumptions.

Where Pith is reading between the lines

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

  • The same instructor-corrector split could be tested on other data-cleaning subtasks such as duplicate detection or schema matching.
  • If context selection quality varies across domains, the method may need an explicit verification step before EM begins.
  • The approach suggests that hybrid large-small model pipelines can trade off accuracy and compute in settings where full LLM inference on every row is too costly.

Load-bearing premise

The LLM can reliably choose a global repair context that improves the SLM corrector's learning signal rather than adding noise or bias.

What would settle it

A controlled test in which the LLM supplies deliberately poor or random contexts and the resulting SLM repair F1 score after EM updates is lower than when no context is supplied.

Figures

Figures reproduced from arXiv: 2606.17582 by Jianwei Wang, Qian Chen, Wenjie Zhang.

Figure 1
Figure 1. Figure 1: An illustrated example of common data errors. The dirty table [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Conceptual comparison of representative data repair methods. (a) Baran relies on user-provided labels to train correctors and classifiers [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of LasRepair++. Starting from the dirty table, the LLM instructor constructs a column influence graph and extracts the target [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Repair performance under varying cell-level error rates on Flight [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance under different numbers of iterations on datasets [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: F1-score under different LLM–SLM combinations on Beers, [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scalability of different repair methods under increasing data [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Running time of different repair methods on seven datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation study of LasRepair++ on seven benchmark datasets. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Case study demonstrating how the column-specific context sets [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
read the original abstract

We study the problem of data repair, a key task in data cleaning that corrects erroneous entries in raw datasets to improve overall data quality. Although recent data-driven methods, especially those based on large language models (LLMs), achieve remarkable performance, we observe that: (i) they directly repair data in the raw and low-quality context, which may compromise learning signals, and (ii) they directly use uncertain model outputs as repairs, potentially introducing unreliable corrections and compromising repair quality. Motivated by the efficiency of small language models (SLMs) and the capabilities of LLMs, and aiming to address the above limitations, we propose LasRepair, a framework that collaborates Large and small language models for data repair. LasRepair employs an LLM as an instructor, which selects a global repair context to guide the SLM. The SLM acts as a corrector, using the selected context to repair erroneous data more efficiently. Moreover, to further improve context quality, we extend LasRepair to LasRepair+, which formulates data repair as an Expectation-Maximisation (EM) procedure that alternates between an E-step for updating the corrector parameters and an M-step for refining the repair context. Furthermore, to mitigate model uncertainty, we propose LasRepair++, which uses column-calibrated model confidence to down-weight unreliable repaired rows when updating the corrector, thereby enhancing repair quality. Theoretical analysis and empirical evaluation demonstrate the superiority of our methods. We theoretically prove the effectiveness of the EM-style procedure and the confidence-based weighting. Experiments on real-world datasets show that LasRepair++~ achieves an average F1-score improvement of 18.1% over the strongest baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes LasRepair, a framework collaborating LLMs (as instructors selecting global repair contexts) with SLMs (as correctors) for data repair. It extends to LasRepair+ via an EM procedure alternating E-step (corrector parameter updates) and M-step (context refinement), and to LasRepair++ adding column-calibrated confidence weighting to down-weight unreliable rows. The authors claim theoretical proofs for the EM procedure and weighting scheme, plus empirical results where LasRepair++ achieves 18.1% average F1 improvement over the strongest baseline on real-world datasets.

Significance. If the central claims hold, the work offers a promising direction for scalable data repair by leveraging LLM contextual capabilities with SLM efficiency while mitigating uncertainty via EM and weighting. The provision of theoretical analysis alongside empirical evaluation on real datasets is a strength, though the load-bearing premise on context selection requires further substantiation for the gains to be fully attributable to the proposed components.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (Framework): The 18.1% F1 improvement is credited to the LLM instructor's selection of global repair context improving the SLM corrector's learning signal, yet no ablation isolating this selection (e.g., vs. random/fixed contexts or raw data) is reported; without it, alternative explanations such as introduction of correlated bias or noise cannot be ruled out, undermining attribution of the gain to LasRepair++.
  2. [§4] §4 (Theoretical Analysis): The claimed proofs for the EM-style procedure and confidence weighting do not explicitly address the potential circular dependence, where both E-step updates and M-step refinements operate on outputs from the same model being optimized; this risks making the analysis non-independent, as noted in the abstract's assertion of 'independent theoretical analysis'.
  3. [§5] §5 (Experiments): The reported average F1 lift lacks accompanying details on data splits, variance across runs, statistical significance, or error analysis, which are necessary to evaluate robustness of the 18.1% claim against post-hoc choices or dataset-specific effects.
minor comments (2)
  1. [Abstract] The progression from LasRepair to LasRepair+ to LasRepair++ could be more clearly delineated in the abstract and introduction to avoid notation confusion for readers.
  2. [§4] Variable definitions in the EM formulation and confidence weighting equations would benefit from explicit cross-references to avoid ambiguity in the derivations.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point-by-point below, indicating planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (Framework): The 18.1% F1 improvement is credited to the LLM instructor's selection of global repair context improving the SLM corrector's learning signal, yet no ablation isolating this selection (e.g., vs. random/fixed contexts or raw data) is reported; without it, alternative explanations such as introduction of correlated bias or noise cannot be ruled out, undermining attribution of the gain to LasRepair++.

    Authors: We agree that an ablation isolating the contribution of the LLM instructor's global context selection is needed to rule out alternative explanations. The current manuscript does not report such an ablation. In the revision we will add experiments comparing LasRepair++ against variants that use random contexts, fixed contexts, and raw data without instructor selection. revision: yes

  2. Referee: [§4] §4 (Theoretical Analysis): The claimed proofs for the EM-style procedure and confidence weighting do not explicitly address the potential circular dependence, where both E-step updates and M-step refinements operate on outputs from the same model being optimized; this risks making the analysis non-independent, as noted in the abstract's assertion of 'independent theoretical analysis'.

    Authors: The proofs in §4 follow standard EM convergence arguments and weighting analysis. We acknowledge that the section does not explicitly discuss potential circular dependence arising from E- and M-steps operating on outputs of the same model. We will revise §4 to clarify this point and adjust the independence claim in the abstract accordingly. revision: yes

  3. Referee: [§5] §5 (Experiments): The reported average F1 lift lacks accompanying details on data splits, variance across runs, statistical significance, or error analysis, which are necessary to evaluate robustness of the 18.1% claim against post-hoc choices or dataset-specific effects.

    Authors: We agree that robustness details are required. The revised manuscript will report data-split methodology, variance across multiple runs, statistical significance tests, and error analysis to support the 18.1% average F1 improvement. revision: yes

Circularity Check

0 steps flagged

No circularity: framework and EM procedure are standard and self-contained

full rationale

The abstract describes a collaborative LLM-SLM framework, an EM-style alternation for context refinement, and column-calibrated weighting, with a claimed theoretical proof of effectiveness. No equations, self-citations, or derivations are supplied that reduce any claimed result to a fitted input or self-defined quantity by construction. The EM loop follows the classic E/M structure without evidence that the 'prediction' of improvement is forced by the same outputs being optimized. The 18.1% F1 claim is presented as empirical, not as a derived identity. This is the normal case of an independent engineering contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so no concrete free parameters, axioms, or invented entities can be extracted or audited from the full manuscript.

pith-pipeline@v0.9.1-grok · 5833 in / 1100 out tokens · 32610 ms · 2026-06-26T22:22:44.435865+00:00 · methodology

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

63 extracted references · 7 canonical work pages

  1. [1]

    Cleanml: A study for evaluating the impact of data cleaning on ml classification tasks,

    P. Li, X. Rao, J. Blase, Y . Zhang, X. Chu, and C. Zhang, “Cleanml: A study for evaluating the impact of data cleaning on ml classification tasks,” 2021. [Online]. Available: https://arxiv.org/abs/1904.09483

    arXiv 2021

  2. [2]

    On llm- enhanced mixed-type data imputation with high-order message passing,

    J. Wang, K. Wang, Y . Zhang, W. Zhang, X. Xu, and X. Lin, “On llm- enhanced mixed-type data imputation with high-order message passing,”

  3. [3]

    Available: https://arxiv.org/abs/2501.02191

    [Online]. Available: https://arxiv.org/abs/2501.02191

    arXiv

  4. [4]

    Datacomp-lm: In search of the next gen- eration of training sets for language models,

    J. Li, A. Fang, G. Smyrnis, M. Ivgi, M. Jordan, S. Gadre, H. Bansal, E. Guha, S. Keh, K. Arora, S. Garg, R. Xin, N. Muennighoff, R. Heckel, J. Mercat, M. Chen, S. Gururangan, M. Wortsman, A. Albalak, Y . Bitton, M. Nezhurina, A. Abbas, C.-Y . Hsieh, D. Ghosh, J. Gardner, M. Kilian, H. Zhang, R. Shao, S. Pratt, S. Sanyal, G. Ilharco, G. Daras, K. Marathe, ...

    Pith/arXiv arXiv 2025

  5. [5]

    Data cleaning: Problems and current approaches,

    E. Rahm and H. Do, “Data cleaning: Problems and current approaches,” IEEE Data Eng. Bull., vol. 23, pp. 3–13, 01 2000

    2000

  6. [6]

    I. F. Ilyas and X. Chu,Data Cleaning. New York, NY , USA: Association for Computing Machinery, 2019

    2019

  7. [7]

    Ensembling llm-induced decision trees for explainable and robust error detection,

    M. Wang, J. Wang, Q. Liu, X. Xu, Z. Xing, L. Zhu, and W. Zhang, “Ensembling llm-induced decision trees for explainable and robust error detection,” 2025. [Online]. Available: https://arxiv.org/abs/2512.07246

    arXiv 2025

  8. [8]

    Automatic data repair: Are we ready to deploy?

    W. Ni, X. Miao, X. Zhao, Y . Wu, S. Liang, and J. Yin, “Automatic data repair: Are we ready to deploy?”Proc. VLDB Endow., vol. 17, no. 10, pp. 2617–2630, 2024. [Online]. Available: https://www.vldb.org/pvldb/ vol17/p2617-miao.pdf

    2024

  9. [9]

    Holistic data cleaning: Putting violations into context,

    X. Chu, I. F. Ilyas, and P. Papotti, “Holistic data cleaning: Putting violations into context,” in2013 IEEE 29th International Conference on Data Engineering (ICDE), 2013, pp. 458–469

    2013

  10. [10]

    Bigdansing: A system for big data cleansing,

    Z. Khayyat, I. F. Ilyas, A. Jindal, S. Madden, M. Ouzzani, P. Papotti, J. Quian´e-Ruiz, N. Tang, and S. Yin, “Bigdansing: A system for big data cleansing,” inProceedings of the 2015 ACM SIGMOD International Conference on Management of Data, Melbourne, Victoria, Australia, May 31 - June 4, 2015, T. K. Sellis, S. B. Davidson, and Z. G. Ives, Eds. ACM, 2015,...

    work page doi:10.1145/2723372.2747646 2015

  11. [11]

    A hybrid data cleaning framework using markov logic networks,

    Y . Gao, C. Ge, X. Miao, H. Wang, B. Yao, and Q. Li, “A hybrid data cleaning framework using markov logic networks,” 2019. [Online]. Available: https://arxiv.org/abs/1903.05826

    Pith/arXiv arXiv 2019

  12. [12]

    Horizon: scalable dependency-driven data cleaning,

    E. K. Rezig, M. Ouzzani, W. G. Aref, A. K. Elmagarmid, A. R. Mahmood, and M. Stonebraker, “Horizon: scalable dependency-driven data cleaning,”Proc. VLDB Endow., vol. 14, no. 11, p. 2546–2554, Jul

  13. [13]

    Available: https://doi.org/10.14778/3476249.3476301

    [Online]. Available: https://doi.org/10.14778/3476249.3476301

    work page doi:10.14778/3476249.3476301

  14. [14]

    Nadeef: a commodity data cleaning system,

    M. Dallachiesa, A. Ebaid, A. Eldawy, A. Elmagarmid, I. F. Ilyas, M. Ouzzani, and N. Tang, “Nadeef: a commodity data cleaning system,” inProceedings of the 2013 ACM SIGMOD International Conference on Management of Data, ser. SIGMOD ’13. New York, NY , USA: Association for Computing Machinery, 2013, p. 541–552. [Online]. Available: https://doi.org/10.1145/2...

    work page doi:10.1145/2463676.2465327 2013

  15. [15]

    Holoclean: Holistic data repairs with probabilistic inference,

    T. Rekatsinas, X. Chu, I. F. Ilyas, and C. R ´e, “Holoclean: Holistic data repairs with probabilistic inference,”Proc. VLDB Endow., vol. 10, no. 11, pp. 1190–1201, 2017. [Online]. Available: http://www.vldb.org/ pvldb/vol10/p1190-rekatsinas.pdf

    2017

  16. [16]

    A unified model for data and constraint repair,

    F. Chiang and R. J. Miller, “A unified model for data and constraint repair,” inProceedings of the 27th International Conference on Data Engineering, ICDE 2011, April 11-16, 2011, Hannover, Germany, S. Abiteboul, K. B ¨ohm, C. Koch, and K. Tan, Eds. IEEE Computer Society, 2011, pp. 446–457. [Online]. Available: https://doi.org/10.1109/ ICDE.2011.5767833

    arXiv 2011

  17. [17]

    On the relative trust between inconsistent data and inaccurate constraints,

    G. Beskales, I. F. Ilyas, L. Golab, and A. Galiullin, “On the relative trust between inconsistent data and inaccurate constraints,” 2012. [Online]. Available: https://arxiv.org/abs/1207.5226

    Pith/arXiv arXiv 2012

  18. [18]

    Conditional functional dependencies for data cleaning,

    P. Bohannon, W. Fan, F. Geerts, X. Jia, and A. Kementsietsidis, “Conditional functional dependencies for data cleaning,” in2007 IEEE 23rd International Conference on Data Engineering, 2007, pp. 746–755

    2007

  19. [19]

    Don’t be scared: use scalable automatic repairing with maximal likelihood and bounded changes,

    M. Yakout, L. Berti- ´Equille, and A. K. Elmagarmid, “Don’t be scared: use scalable automatic repairing with maximal likelihood and bounded changes,” inProceedings of the ACM SIGMOD International Confer- ence on Management of Data, SIGMOD 2013, New York, NY, USA, June 22-27, 2013, K. A. Ross, D. Srivastava, and D. Papadias, Eds. ACM, 2013, pp. 553–564. [O...

    arXiv 2013

  20. [20]

    Baran: Effective error correction via a unified context representation and transfer learning,

    M. Mahdavi and Z. Abedjan, “Baran: Effective error correction via a unified context representation and transfer learning,”Proc. VLDB Endow., vol. 13, no. 11, pp. 1948–1961, 2020. [Online]. Available: http://www.vldb.org/pvldb/vol13/p1948-mahdavi.pdf

    1948

  21. [21]

    GIDCL: A graph- enhanced interpretable data cleaning framework with large language models,

    M. Yan, Y . Wang, Y . Wang, X. Miao, and J. Li, “GIDCL: A graph- enhanced interpretable data cleaning framework with large language models,”Proc. ACM Manag. Data, vol. 2, no. 6, pp. 236:1–236:29,

  22. [22]

    Available: https://doi.org/10.1145/3698811

    [Online]. Available: https://doi.org/10.1145/3698811

    work page doi:10.1145/3698811

  23. [23]

    Jellyfish: A large language model for data preprocessing,

    H. Zhang, Y . Dong, C. Xiao, and M. Oyamada, “Jellyfish: A large language model for data preprocessing,” 2024. [Online]. Available: https://arxiv.org/abs/2312.01678

    arXiv 2024

  24. [24]

    Stochastic gradient descent as approximate bayesian inference,

    S. Mandt, M. D. Hoffman, and D. M. Blei, “Stochastic gradient descent as approximate bayesian inference,” 2018. [Online]. Available: https: //arxiv.org/abs/1704.04289

    Pith/arXiv arXiv 2018

  25. [25]

    Uncertainty estimation in autoregressive structured prediction,

    A. Malinin and M. Gales, “Uncertainty estimation in autoregressive structured prediction,” 2021. [Online]. Available: https://arxiv.org/abs/ 2002.07650

    arXiv 2021

  26. [26]

    Listwise deletion in high dimensions,

    J. S. Wang and P. M. Aronow, “Listwise deletion in high dimensions,” Political Analysis, vol. 31, no. 1, p. 149–155, 2023

    2023

  27. [27]

    An introduction to variable and feature selection,

    I. Guyon and A. Elisseeff, “An introduction to variable and feature selection,”J. Mach. Learn. Res., vol. 3, no. null, p. 1157–1182, Mar. 2003

    2003

  28. [28]

    Efficient feature selection via analysis of relevance and redundancy,

    L. Yu and H. Liu, “Efficient feature selection via analysis of relevance and redundancy,”J. Mach. Learn. Res., vol. 5, p. 1205–1224, Dec. 2004

    2004

  29. [29]

    Hyperjoin: Llm-augmented hypergraph link prediction for joinable table discovery,

    S. Liu, J. Wang, X. Lin, L. Qin, W. Zhang, and Y . Zhang, “Hyperjoin: Llm-augmented hypergraph link prediction for joinable table discovery,”

  30. [30]

    Available: https://arxiv.org/abs/2601.01015

    [Online]. Available: https://arxiv.org/abs/2601.01015

    arXiv

  31. [31]

    Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift,

    Y . Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. V . Dillon, B. Lakshminarayanan, and J. Snoek, “Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift,” 2019. [Online]. Available: https://arxiv.org/abs/1906.02530

    arXiv 2019

  32. [32]

    Missing data im- putation with uncertainty-driven network,

    J. Wang, Y . Zhang, K. Wang, X. Lin, and W. Zhang, “Missing data im- putation with uncertainty-driven network,” vol. 2, no. 3, 2024. [Online]. Available: https://doi.org/10.1145/3654920

    work page doi:10.1145/3654920 2024

  33. [33]

    Underspecification presents challenges for credibility in modern machine learning,

    A. D’Amour, K. Heller, D. Moldovan, B. Adlam, B. Alipanahi, A. Beu- tel, C. Chen, J. Deaton, J. Eisenstein, M. D. Hoffman, F. Hormozdi- ari, N. Houlsby, S. Hou, G. Jerfel, A. Karthikesalingam, M. Lucic, Y . Ma, C. McLean, D. Mincu, A. Mitani, A. Montanari, Z. Nado, V . Natarajan, C. Nielson, T. F. Osborne, R. Raman, K. Ramasamy, R. Sayres, J. Schrouff, M....

    arXiv 2020

  34. [34]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging llm-as-a-judge with mt-bench and chatbot arena,”

  35. [35]

    Available: https://arxiv.org/abs/2306.05685

    [Online]. Available: https://arxiv.org/abs/2306.05685

    Pith/arXiv arXiv

  36. [36]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert- V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Am...

    Pith/arXiv arXiv 2020

  37. [37]

    Maximum likelihood estimation of observer error-rates using the em algorithm,

    A. P. Dawid and A. M. Skene, “Maximum likelihood estimation of observer error-rates using the em algorithm,”Journal of the Royal Statistical Society. Series C (Applied Statistics), vol. 28, no. 1, pp. 20– 28, 1979. [Online]. Available: http://www.jstor.org/stable/2346806

    arXiv 1979

  38. [38]

    R. M. Neal and G. E. Hinton,A View of the Em Algorithm that Justifies Incremental, Sparse, and other Variants. Dordrecht: Springer Netherlands, 1998, pp. 355–368. [Online]. Available: https://doi.org/10. 1007/978-94-011-5014-9 12

    1998

  39. [39]

    Maximum likelihood from incomplete data via the em algorithm,

    A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via the em algorithm,”Journal of the Royal Statistical Society. Series B (Methodological), vol. 39, no. 1, pp. 1–38,

  40. [40]

    Available: http://www.jstor.org/stable/2984875

    [Online]. Available: http://www.jstor.org/stable/2984875

    arXiv

  41. [41]

    Self-training with noisy student improves imagenet classification,

    Q. Xie, M.-T. Luong, E. Hovy, and Q. V . Le, “Self-training with noisy student improves imagenet classification,” 2020. [Online]. Available: https://arxiv.org/abs/1911.04252

    arXiv 2020

  42. [42]

    Learning to reweight examples for robust deep learning,

    M. Ren, W. Zeng, B. Yang, and R. Urtasun, “Learning to reweight examples for robust deep learning,” 2019. [Online]. Available: https: //arxiv.org/abs/1803.09050

    Pith/arXiv arXiv 2019

  43. [43]

    Neural attributed community search at billion scale,

    J. Wang, K. Wang, X. Lin, W. Zhang, and Y . Zhang, “Neural attributed community search at billion scale,” 2024. [Online]. Available: https: //arxiv.org/abs/2403.18874

    arXiv 2024

  44. [44]

    Efficient unsupervised community search with pre-trained graph transformer,

    ——, “Efficient unsupervised community search with pre-trained graph transformer,” 2024. [Online]. Available: https://arxiv.org/abs/2403.18869

    arXiv 2024

  45. [45]

    Tabllm: Few-shot classification of tabular data with large language models,

    S. Hegselmann, A. Buendia, H. Lang, M. Agrawal, X. Jiang, and D. Son- tag, “Tabllm: Few-shot classification of tabular data with large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2210.10723

    arXiv 2023

  46. [46]

    Lift: Language-interfaced fine-tuning for non-language machine learning tasks,

    T. Dinh, Y . Zeng, R. Zhang, Z. Lin, M. Gira, S. Rajput, J. yong Sohn, D. Papailiopoulos, and K. Lee, “Lift: Language-interfaced fine-tuning for non-language machine learning tasks,” 2022. [Online]. Available: https://arxiv.org/abs/2206.06565

    arXiv 2022

  47. [47]

    Mentornet: Learn- ing data-driven curriculum for very deep neural networks on corrupted labels,

    L. Jiang, Z. Zhou, T. Leung, L.-J. Li, and L. Fei-Fei, “Mentornet: Learn- ing data-driven curriculum for very deep neural networks on corrupted labels,” 2018. [Online]. Available: https://arxiv.org/abs/1712.05055

    Pith/arXiv arXiv 2018

  48. [48]

    Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,

    E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuinness, “Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,” 2020. [Online]. Available: https://arxiv.org/abs/1908.02983

    arXiv 2020

  49. [49]

    Learning with noisy labels,

    N. Natarajan, I. S. Dhillon, P. Ravikumar, and A. Tewari, “Learning with noisy labels,” inProceedings of the 27th International Conference on Neural Information Processing Systems - Volume 1, ser. NIPS’13. Red Hook, NY , USA: Curran Associates Inc., 2013, p. 1196–1204

    2013

  50. [50]

    Confident learning: Estimating uncertainty in dataset labels,

    C. Northcutt, L. Jiang, and I. Chuang, “Confident learning: Estimating uncertainty in dataset labels,”J. Artif. Int. Res., vol. 70, p. 1373–1411, May 2021. [Online]. Available: https://doi.org/10.1613/jair.1.12125

    work page doi:10.1613/jair.1.12125 2021

  51. [51]

    A baseline for detecting misclassified and out-of-distribution examples in neural networks,

    D. Hendrycks and K. Gimpel, “A baseline for detecting misclassified and out-of-distribution examples in neural networks,” 2018. [Online]. Available: https://arxiv.org/abs/1610.02136

    Pith/arXiv arXiv 2018

  52. [52]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” 2017. [Online]. Available: https://arxiv.org/ abs/1706.04599

    Pith/arXiv arXiv 2017

  53. [53]

    Lasrepair++: Full version,

    Q. C. Jianwei Wang, Wenjie Zhang, “Lasrepair++: Full version,” https: //github.com/OZCQC/LasRepair, 2026

    2026

  54. [54]

    Messing up with bart: error generation for evaluating data- cleaning algorithms,

    P. C. Arocena, B. Glavic, G. Mecca, R. J. Miller, P. Papotti, and D. Santoro, “Messing up with bart: error generation for evaluating data- cleaning algorithms,”Proc. VLDB Endow., vol. 9, no. 2, p. 36–47, Oct

  55. [55]

    Available: https://doi.org/10.14778/2850578.2850579

    [Online]. Available: https://doi.org/10.14778/2850578.2850579

    work page doi:10.14778/2850578.2850579

  56. [56]

    Inference and missing data,

    D. B. Rubin, “Inference and missing data,”Biometrika, vol. 63, no. 3, pp. 581–592, 1976. [Online]. Available: http://www.jstor.org/stable/2335739

    arXiv 1976

  57. [57]

    Boostclean: Automated error detection and repair for machine learning,

    S. Krishnan, M. J. Franklin, K. Goldberg, and E. Wu, “Boostclean: Automated error detection and repair for machine learning,”CoRR, vol. abs/1711.01299, 2017. [Online]. Available: http://arxiv.org/abs/1711. 01299

    Pith/arXiv arXiv 2017

  58. [58]

    The truth of the f-measure,

    Y . Sasakiet al., “The truth of the f-measure,”Teach tutor mater, vol. 1, no. 5, pp. 1–5, 2007

    2007

  59. [59]

    Binary codes capable of correcting deletions, inser- tions and reversals,

    V . I. Levenshtein, “Binary codes capable of correcting deletions, inser- tions and reversals,”Soviet Physics Doklady, vol. 10, no. 8, pp. 707–710, 1966

    1966

  60. [60]

    Exploring the limits of transfer learn- ing with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learn- ing with a unified text-to-text transformer,” 2023. [Online]. Available: https://arxiv.org/abs/1910.10683

    Pith/arXiv arXiv 2023

  61. [61]

    Openai gpt-5 system card,

    OpenAI, “Openai gpt-5 system card,” 2026. [Online]. Available: https: //arxiv.org/abs/2601.03267

    Pith/arXiv arXiv 2026

  62. [62]

    Gpt-4o system card,

    ——, “Gpt-4o system card,” 2024. [Online]. Available: https://arxiv.org/ abs/2410.21276

    Pith/arXiv arXiv 2024

  63. [63]

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehen- sion,

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehen- sion,” 2019. [Online]. Available: https://arxiv.org/abs/1910.13461

    Pith/arXiv arXiv 2019