Pith. sign in

REVIEW 4 major objections 5 minor 110 references

EdgeLM: Edge Demonstrations for Language Models' Table Understanding

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A retrieval rule that adds local label contrast and the model's own past mistakes improves LLM table understanding without retraining.

desk verdict EdgeLM's local-contrast plus model-error retrieval is a real, practically useful ICL recipe with strong results on imbalanced tabular tasks; missing similarity function and window w are the main fixable gaps. read the letter →

arxiv 2608.04390 v1 pith:ZOSKON7K submitted 2026-08-05 cs.CL cs.DB

Soroush Omidvartehrani , Mohammadamin Habibollah , Mohammadreza Daviran , Davood Rafiei This is my paper · ORCID
classification cs.CLcs.DB
keywords in-contextlearningdemonstrationselectionedgedemonstrationstableunderstandingdatawranglingdecisionboundaryretrieval-augmentedpredictionlargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that for in-context learning over tabular data, demonstrations chosen for similarity to the query are not enough: near a decision boundary, the nearest neighbors all carry the same label and reinforce the model's likely answer instead of showing what distinguishes it. EdgeLM therefore retrieves edge demonstrations, examples that stay close to the query but sit near the local decision boundary, using two complementary signals: nearby examples with different ground-truth labels (data edge) and nearby examples the deployed LLM previously got wrong in zero-shot (model edge). Across five data-wrangling tasks, fifteen datasets, and five open-weight and proprietary LLMs, EdgeLM matches or beats similarity, diversity, random, and zero-shot selection in nearly every configuration, with the largest gains on rare-class and highly imbalanced tasks. If the claim holds, changing only how demonstrations are selected improves table understanding without fine-tuning, task-specific engineering, or extra inference-time model calls.

What carries the argument

The central object is the edge demonstration, built from two retrieval components that share the same task-specific similarity ranking used by ordinary nearest-neighbor retrieval. The data-edge component restricts selection to the top-w neighborhood of the query, partitions that neighborhood by ground-truth label, and fills the budget round-robin across label groups so competing outcomes appear together; the model-edge component uses a precomputed error pool of instances the frozen LLM answered incorrectly zero-shot and retrieves the K nearest such errors for the query, rendering each as a wrong-answer-then-correct-answer example. Concatenating the two sets into a single fixed-budget prompt is the mechanism that carries the argument, with no model retraining and no task-specific engineering.

What would settle it

A concrete falsifier: on a table task where the decision boundary is known to be non-local (labels depend on combinations of distant columns rather than local cell similarity), run EdgeLM against plain similarity retrieval; if edge does not beat similarity, the claim that local edge evidence exposes the decision boundary fails. A second check: on a task where the model's zero-shot errors are not locally representative, the model-edge component should contribute nothing, which could be tested by isolating the model-edge-only ablation on such a task.

Watch

Extended reading notes

Core claim

EdgeLM's central discovery is that the position of a demonstration relative to the local decision boundary can matter more than its raw similarity to the query. The paper operationalizes the unseen boundary from two views of the labeled pool: data-edge retrieval partitions the query's neighborhood by ground-truth label and round-robin selects the most similar instance from each label group, while model-edge retrieval precomputes the frozen model's zero-shot errors on the pool and retrieves the nearest such errors for each query, presenting them as wrong-to-right corrective demonstrations. The authors report that this combined edge prompt achieves the best or near-best F1 on every task-model combination among the five selection strategies tested, that the two edge views are complementary, and that the advantage is insensitive to how the demonstration budget is split between them.

Load-bearing premise

The load-bearing premise is that the task-specific similarity function used to define the local neighborhood correctly ranks examples by task-relevant closeness; if it misorders instances with respect to the true decision boundary, both the data-edge and model-edge demonstrations lose their discriminative value.

Editorial extensions

If this is right

  • A practitioner can adopt EdgeLM by swapping the demonstration retrieval rule; the underlying model, prompt template, and inference procedure stay fixed.
  • The largest wins appear on tasks where the nearest neighbors are label-homogeneous, such as anomaly detection and error detection, where edge raises F1 over similarity by up to 0.212 on open-weight models.
  • The method is robust to the budget split: a default 2/3 data-to-model split performs within 0.010 F1 of the per-dataset best split on average.
  • EdgeLM's gains extend across model families and scales, compressing on stronger proprietary models but never reversing.
  • Since the framework only assumes a labeled pool and a frozen model, it transfers directly to any table-centric prediction task expressible as in-context learning.

Reading between the lines

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

  • Editorial extension: the same 'similar examples reinforce the likely answer' mechanism is not table-specific, so near-boundary contrast retrieval should also lift text and code in-context learning; this is a testable prediction the paper does not make.
  • Editorial extension: because the data-edge component acts only when the nearest-neighbor set is label-homogeneous, a cheaper variant that checks label diversity before doing the round-robin pass would likely recover most of the gain at lower retrieval cost.
  • Editorial extension: the paper's reliance on an unspecified task-specific similarity function leaves open whether a learned similarity would amplify the edge effect; if the neighborhood ranking is poor, both edge views degrade, which the paper's ablations do not directly test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes EdgeLM, a demonstration-selection method for in-context table understanding that augments similarity retrieval with two kinds of 'edge' evidence: data-edge demonstrations (nearby labeled instances with different gold labels) and model-edge demonstrations (nearby instances that the deployed model predicted incorrectly in a zero-shot pass over the labeled pool). The method is evaluated against zero-shot, random, diverse, and similarity-based selection across five data-wrangling tasks, fifteen datasets, and five LLMs, with additional ablations on the two edge components, the data/model budget split, and sensitivity to K and w. The paper reports that EdgeLM is best or tied in nearly all settings, along with cost measurements in Appendix C and a sensitivity grid in Appendix D.

Significance. The experimental design has genuine strengths: the five selection strategies share identical prompt templates, label spaces, demonstration pools, and test sets within each model, and the model-error pool is built only from the labeled pool, so test queries are not leaked. The idea is simple, requires no retraining or task-specific engineering, and could be a useful baseline for in-context demonstration selection on tabular data. However, the manuscript as written does not fully support its headline claims: a core component of the retrieval rule is never specified, the main results table lacks the standard deviations used to define ties, and the 'every setting' claim in the abstract is contradicted by the paper's own per-dataset results.

major comments (4)
  1. [Sections 4.1, 4.2, 5.1, and Appendix D] The core retrieval rule is not reproducible as written. Both edge components are defined as selections from instances ranked by a 'task-specific similarity function': N_w(x) is the top w ranked instances in Section 4.1, and the model error pool is ranked by similarity to the query in Section 4.2. That function is never named, defined, or pointed to in the text or appendices, and the default window w used for Tables 3-5 is never stated; Section 5.1 fixes K=5 but is silent on w. Appendix D sweeps w in {5,10,20,50,100} only for Llama-3.1-8B and with a single seed, and does not say which value produced the main tables. Because both edge components and the 'local relevance' premise depend entirely on this ranking, please specify the similarity function and the default w, or state where in the released code they can be found.
  2. [Abstract, Section 5.2, and Table 6] The abstract claims EdgeLM 'consistently achieves the best or near-best performance in every setting,' and Section 5.2 says the method is 'at least as effective as the best alternative on every task and every model.' Table 6 contradicts this even within the single model it reports: on DBLP-ACM EdgeLM scores 0.824 F1 versus 0.894 for the diverse baseline, and on Hospital it scores 0.171 versus 0.221 for diverse, while on Amazon-Google, BalanceScale, and Splice it merely ties similarity. Please either soften the claim to 'best or tied on 13 of 15 datasets' and explicitly discuss the two losses, or provide evidence from the other models that justifies the universal phrasing.
  3. [Table 3 and Section 5.2] The Table 3 caption defines bold as the best F1 and 'any within seed-to-seed standard deviation of it,' but no standard deviations are reported in Table 3 or referenced to an appendix. Consequently the repeated statements about ties and 'within one standard deviation' cannot be checked; for example, GPT-4o-mini entity matching shows EdgeLM at 0.788 versus similarity at 0.793, and the Mistral-Nemo data-imputation row is 0.502 for both. Please report the seed-to-seed variation (standard deviations or confidence intervals) for the main comparison so wins, ties, and losses are verifiable.
  4. [Table 5 and Section 5.4] The allocation-robustness analysis is missing setup information. Table 5 does not state which model produced the numbers, whether each cell is a single seed or an average over three seeds, or which window w is used. The claim that the method is insensitive to the data/model split rests on differences such as +0.043 (BalanceScale) and +0.032 (DiabetesPima), which have no error bars, so it is not possible to tell whether these are real effects or noise. Please specify the experimental configuration and add variance information to the table or its caption.
minor comments (5)
  1. [Tables 4 and 5] Neither table identifies the model used; the numbers appear to match the Llama-3.1-8B rows of Table 3, but this should be stated explicitly in each caption.
  2. [Table 5] The dataset label 'Fligh.' is truncated; rename it to 'Flights' for consistency with Table 2.
  3. [Appendix D] The sensitivity sweep uses a single seed; the text should explicitly label this as single-seed evidence, since the robustness conclusion is partly based on it.
  4. [Section 4.1] The neighborhood is written both as N_w(x) and NW(x); standardize the notation to avoid confusion.
  5. [Section 5.2] The sentence 'Zero-shot is the weakest on average' is accurate as an average but is slightly at odds with cells such as GPT-4o-mini schema matching, where zero-shot (0.850) exceeds random (0.843); a caveat such as 'on average' would prevent a misreading.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: EdgeLM's edge evidence is selected from the labeled pool and evaluated on held-out test data, so the self-referential model-edge component is an empirical mechanism rather than a definitional identity.

full rationale

EdgeLM is an empirical retrieval method rather than a derived theorem. Both edge components are constructed from the labeled reference pool: the data edge uses ground-truth label contrast in a similarity neighborhood, and the model edge uses the frozen model's zero-shot errors on that same pool. No parameter is fitted to the held-out test set; the main results in Tables 3–6 are measured on test instances that were never used for selection, and the per-dataset table shows genuine losses and ties (e.g., DBLP-ACM and Hospital), which rules out a definitional identity between the method and its measured outcome. The self-referential character of the model edge—using the deployed model's own mistakes—is a design mechanism, not a circular derivation, because the errors are computed on the labeled pool and the evaluation target is held-out accuracy. The only self-citations (Omidvartehrani and Rafiei 2026; Habibollah and Rafiei 2026; Daviran et al. 2026) appear in related-work background and are not load-bearing for the central claim. The unspecified similarity function and unstated default window w are reproducibility gaps, not circularity. I find no circular step requiring a score above 1.

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

The central claim rests on a labeled reference pool, a task-specific similarity function, and the assumption that zero-shot failures on the pool are locally representative of test-time failures. Numeric hyperparameters K, the data/model split, and the unstated window w are hand-chosen, though the robustness analysis suggests the method is insensitive to them.

free parameters (3)
  • demonstration budget K = 5
    Fixed at 5 in the main experiments; the sensitivity sweep in Appendix D covers K in {3,5,8,10} and shows the advantage over baselines is stable. Hand-chosen, not fitted to data.
  • data/model edge budget split = 2/3 (two data-edge, three model-edge demonstrations)
    Default allocation of the K=5 budget used in Table 3. Table 5 shows the oracle best split beats the 2/3 default by only +0.010 F1 on average, so the method is insensitive to this parameter.
  • selection window w = not stated in main text; swept over {5,10,20,50,100} in Appendix D
    N_w(x), the neighborhood from which data-edge candidates are drawn, is defined in Section 4.1, but the value used in the main experiments is never reported. This is a reproducibility gap.
assumptions (3)
  • domain assumption A labeled reference pool P_tau is available for each task, drawn from the same domain as the queries.
    Stated in Section 3 and in the Limitations: if no labels exist, neither edge view can be built.
  • domain assumption The task-specific similarity function correctly orders pool instances by task-relevant closeness.
    Invoked in Sections 4.1 and 4.2; the paper never specifies this function, so the neighborhoods are defined only abstractly.
  • domain assumption Zero-shot errors on the labeled pool are locally representative of the model's test-time failures.
    Section 4.2 builds the error pool by zero-shot evaluation on the labeled pool; the ablation shows empirical support, but it is an assumption about model error locality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EdgeLM: Edge Demonstrations for Language Models' Table Understanding." pith.science (2026). https://pith.science/paper/ZOSKON7K

@misc{pith2026260804390,
  author       = {Pith},
  title        = {Pith review of: EdgeLM: Edge Demonstrations for Language Models' Table Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZOSKON7K}},
  note         = {Machine review of arXiv:2608.04390}
}
read the original abstract

Large language models (LLMs) perform table-centric prediction through in-context learning, making demonstration selection critical to performance. Existing retrieval methods prioritize similarity to the query, but similar demonstrations often reinforce the model's likely prediction rather than reveal the distinctions needed for difficult decisions. We propose EdgeLM, a retrieval framework that instead selects edge evidence, demonstrations that are both relevant to the query and informative about the decision boundary. EdgeLM retrieves two complementary forms of edge evidence by selecting data edges, nearby examples with different ground-truth labels, and model edges, similar examples previously misclassified by the deployed model. EdgeLM requires neither model retraining nor task-specific engineering. Across five data wrangling tasks, fifteen datasets, and five open-weight and proprietary LLMs, EdgeLM consistently achieves the best or near-best performance in every setting, while ablations show that the two forms of edge evidence provide complementary benefits. Our code and datasets are publicly available at https://github.com/soroushomidvar/EdgeLM.

Figures

Figures reproduced from arXiv: 2608.04390 by the authors.

Figure 1
Figure 1. Two complementary sources of edge evidence in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of EDGELM. Before inference, the LLM is evaluated zero-shot on the labeled pool to identify model-specific errors. For each test query, pool instances are ranked by similarity. EDGELM selects locally contrasting examples from the query neighborhood (data edge) and retrieves similar model errors from the precomputed error pool (model edge). The two sets are combined into a single prompt and passed to the LLM… view at source ↗
Figure 3
Figure 3. F1 gain of EDGELM over random and similarity selection across the demonstration budget K (rows) and the selection window w (columns), for Llama-3.1-8B on all fifteen datasets. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

110 extracted references · 61 canonical work pages

  1. [1]

    Proceedings of the 2019 International Conference on Management of Data (SIGMOD) , pages =

    Mahdavi, Mohammad and Abedjan, Ziawasch and Castro Fernandez, Raul and Madden, Samuel and Ouzzani, Mourad and Stonebraker, Michael and Tang, Nan , title =. Proceedings of the 2019 International Conference on Management of Data (SIGMOD) , pages =

    2019

  2. [2]

    Proceedings of the VLDB Endowment , volume =

    Mahdavi, Mohammad and Abedjan, Ziawasch , title =. Proceedings of the VLDB Endowment , volume =

  3. [3]

    Proceedings of the VLDB Endowment , volume =

    Li, Xian and Dong, Xin Luna and Lyons, Kenneth and Meng, Weiyi and Srivastava, Divesh , title =. Proceedings of the VLDB Endowment , volume =

  4. [4]

    Hospital Compare , author =

  5. [5]

    Evaluation of Entity Resolution Approaches on Real-World Match Problems , journal =

    K\". Evaluation of Entity Resolution Approaches on Real-World Match Problems , journal =

  6. [6]

    C., Paul and Doan, AnHai and Ardalan, Adel and Ballard, Jeffrey R

    Konda, Pradap and Das, Sanjib and Suganthan G. C., Paul and Doan, AnHai and Ardalan, Adel and Ballard, Jeffrey R. and Li, Han and Panahi, Fatemah and Zhang, Haojun and Naughton, Jeff and Prasad, Shishir and Krishnan, Ganesh and Deep, Rohit and Raghavendra, Vijay , title =. Proceedings of the VLDB Endowment , volume =

  7. [7]

    Proceedings of the 2018 International Conference on Management of Data (SIGMOD) , pages =

    Mudgal, Sidharth and Li, Han and Rekatsinas, Theodoros and Doan, AnHai and Park, Youngchoon and Krishnan, Ganesh and Deep, Rohit and Arcaute, Esteban and Raghavendra, Vijay , title =. Proceedings of the 2018 International Conference on Management of Data (SIGMOD) , pages =

    2018

  8. [8]

    Proceedings of the IEEE 37th International Conference on Data Engineering (ICDE) , pages =

    Koutras, Christos and Siachamis, George and Ionescu, Andra and Psarakis, Kyriakos and Brons, Jerry and Fragkoulis, Marios and Lofi, Christoph and Bonifati, Angela and Katsifodimos, Asterios , title =. Proceedings of the IEEE 37th International Conference on Data Engineering (ICDE) , pages =

Show all 110 references
  1. [9]

    doi:10.5281/zenodo.5084605 , year =

    Koutras, Christos and Siachamis, George and Ionescu, Andra and Psarakis, Kyriakos and Brons, Jerry and Fragkoulis, Marios and Lofi, Christoph and Bonifati, Angela and Katsifodimos, Asterios , title =. doi:10.5281/zenodo.5084605 , year =

    doi

  2. [10]

    and Towell, Geoffrey G

    Noordewier, Michiel O. and Towell, Geoffrey G. and Shavlik, Jude W. , title =. Advances in Neural Information Processing Systems (NIPS) , pages =

  3. [11]

    and Everhart, James E

    Smith, Jack W. and Everhart, James E. and Dickson, W. C. and Knowler, William C. and Johannes, Robert S. , title =. Proceedings of the Annual Symposium on Computer Application in Medical Care , pages =

  4. [12]

    , title =

    Siegler, Robert S. , title =. Cognitive Psychology , volume =

  5. [13]

    and Fan, Wei and Lee, Wenke and Prodromidis, Andreas and Chan, Philip K

    Stolfo, Salvatore J. and Fan, Wei and Lee, Wenke and Prodromidis, Andreas and Chan, Philip K. , title =. Proceedings of the DARPA Information Survivability Conference and Exposition (DISCEX) , pages =

  6. [14]

    , title =

    Tavallaee, Mahbod and Bagheri, Ebrahim and Lu, Wei and Ghorbani, Ali A. , title =. Proceedings of the IEEE Symposium on Computational Intelligence for Security and Defense Applications (CISDA) , pages =

  7. [15]

    and Bischl, Bernd and Torgo, Luis , title =

    Vanschoren, Joaquin and van Rijn, Jan N. and Bischl, Bernd and Torgo, Luis , title =. SIGKDD Explorations , volume =

  8. [16]

    Dua, Dheeru and Graff, Casey , title =

  9. [17]

    Greenwade

    George D. Greenwade. The C omprehensive T ex A rchive N etwork ( CTAN ). TUGBoat. 1993

    1993

  10. [18]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    A Computational Approach to Edge Detection , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 1986 , publisher=

    1986

  11. [19]

    2018 , publisher=

    Digital Image Processing , author=. 2018 , publisher=

    2018

  12. [20]

    Pattern recognition , volume=

    A review on image segmentation techniques , author=. Pattern recognition , volume=. 1993 , publisher=

    1993

  13. [21]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Graph Neural Networks: A Review of Methods and Applications , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , publisher=

    2021

  14. [22]

    Physics Reports , volume=

    Community Detection in Graphs , author=. Physics Reports , volume=. 2010 , publisher=

    2010

  15. [23]

    Knowledge and Information Systems , volume=

    A Survey of Methods for Time Series Change Point Detection , author=. Knowledge and Information Systems , volume=. 2017 , publisher=

    2017

  16. [24]

    1993 , publisher=

    Detection of Abrupt Changes: Theory and Application , author=. 1993 , publisher=

    1993

  17. [25]

    2023 , howpublished=

    Achieving 10,000x Training Data Reduction with High-Fidelity Labels , author=. 2023 , howpublished=

    2023

  18. [26]

    University of Wisconsin--Madison , year=

    Active Learning Literature Survey , author=. University of Wisconsin--Madison , year=

  19. [27]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    A Survey on Active Learning and Sample Selection Strategies , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , publisher=

    2021

  20. [28]

    Proceedings of the ACM on Management of Data , volume=

    Table-gpt: Table fine-tuned gpt for diverse table tasks , author=. Proceedings of the ACM on Management of Data , volume=. 2024 , publisher=

    2024

  21. [29]

    Journal of computational and applied mathematics , volume=

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis , author=. Journal of computational and applied mathematics , volume=. 1987 , publisher=

    1987

  22. [30]

    Proceedings of the 2010 ACM SIGMOD International Conference on Management of data , pages=

    ERACER: a database approach for statistical inference and data cleaning , author=. Proceedings of the 2010 ACM SIGMOD International Conference on Management of data , pages=

    2010

  23. [31]

    Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Imputing various incomplete attributes via distance likelihood maximization , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

  24. [32]

    arXiv preprint arXiv:2207.05221 , year=

    Language models (mostly) know what they know , author=. arXiv preprint arXiv:2207.05221 , year=

    arXiv

  25. [33]

    Proceedings of the VLDB Endowment , volume=

    HoloClean: Holistic Data Repairs with Probabilistic Inference , author=. Proceedings of the VLDB Endowment , volume=

  26. [34]

    Communications of the ACM , volume=

    Exact matrix completion via convex optimization , author=. Communications of the ACM , volume=. 2012 , publisher=

    2012

  27. [35]

    Bioinformatics , volume=

    MissForest—non-parametric missing value imputation for mixed-type data , author=. Bioinformatics , volume=. 2012 , publisher=

    2012

  28. [36]

    Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages=

    Xgboost: A scalable tree boosting system , author=. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining , pages=

  29. [37]

    2012 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , pages=

    Missing categorical data imputation approach based on similarity , author=. 2012 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , pages=. 2012 , organization=

    2012

  30. [38]

    Proceedings of the 17th International Conference on Pattern Recognition, 2004

    Nearest neighbor ensemble , author=. Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. , volume=. 2004 , organization=

    2004

  31. [39]

    The American Statistician , volume=

    An introduction to kernel and nearest-neighbor nonparametric regression , author=. The American Statistician , volume=. 1992 , publisher=

    1992

  32. [40]

    Transactions on computational science I , pages=

    Missing value imputation based on data clustering , author=. Transactions on computational science I , pages=. 2008 , publisher=

    2008

  33. [41]

    2017 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) , pages=

    A new iterative fuzzy clustering algorithm for multiple imputation of missing data , author=. 2017 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) , pages=. 2017 , organization=

    2017

  34. [42]

    Mathematical Problems in Engineering , volume=

    Missing value imputation based on gaussian mixture model for the internet of things , author=. Mathematical Problems in Engineering , volume=. 2015 , publisher=

    2015

  35. [43]

    Advances in Knowledge Discovery and Data Mining: 22nd Pacific-Asia Conference, PAKDD 2018, Melbourne, VIC, Australia, June 3-6, 2018, Proceedings, Part III 22 , pages=

    Mida: Multiple imputation using denoising autoencoders , author=. Advances in Knowledge Discovery and Data Mining: 22nd Pacific-Asia Conference, PAKDD 2018, Melbourne, VIC, Australia, June 3-6, 2018, Proceedings, Part III 22 , pages=. 2018 , organization=

    2018

  36. [44]

    Pattern Recognition , volume=

    Handling incomplete heterogeneous data using vaes , author=. Pattern Recognition , volume=. 2020 , publisher=

    2020

  37. [45]

    2021 IEEE 37th International Conference on Data Engineering (ICDE) , pages=

    Capturing semantics for imputation with pre-trained language models , author=. 2021 IEEE 37th International Conference on Data Engineering (ICDE) , pages=. 2021 , organization=

    2021

  38. [46]

    Proceedings of the VLDB Endowment , volume=

    Can Foundation Models Wrangle Your Data? , author=. Proceedings of the VLDB Endowment , volume=. 2022 , publisher=

    2022

  39. [47]

    Clinical trials , volume=

    Fallacies of last observation carried forward analyses , author=. Clinical trials , volume=. 2016 , publisher=

    2016

  40. [48]

    Journal of clinical epidemiology , volume=

    Imputation of missing longitudinal data: a comparison of methods , author=. Journal of clinical epidemiology , volume=. 2003 , publisher=

    2003

  41. [49]

    The Lancet Neurology , volume=

    Three-dimensional, task-specific robot therapy of the arm after stroke: a multicentre, parallel-group randomised trial , author=. The Lancet Neurology , volume=. 2014 , publisher=

    2014

  42. [50]

    last observation carried forward

    Does analysis using “last observation carried forward” introduce bias in dementia research? , author=. Cmaj , volume=. 2008 , publisher=

    2008

  43. [51]

    International conference on machine learning , pages=

    Gain: Missing data imputation using generative adversarial nets , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  44. [52]

    Proceedings of Machine Learning and Systems , volume=

    Attention-based learning for missing data imputation in holoclean , author=. Proceedings of Machine Learning and Systems , volume=

  45. [53]

    arXiv preprint arXiv:2501.02191 , year=

    On llm-enhanced mixed-type data imputation with high-order message passing , author=. arXiv preprint arXiv:2501.02191 , year=

    arXiv

  46. [54]

    Proceedings of the VLDB Endowment , volume=

    Data Imputation with Limited Data Redundancy Using Data Lakes , author=. Proceedings of the VLDB Endowment , volume=. 2025 , publisher=

    2025

  47. [55]

    2025 IEEE 41st International Conference on Data Engineering (ICDE) , pages=

    A Zero-Training Error Correction System with Large Language Models , author=. 2025 IEEE 41st International Conference on Data Engineering (ICDE) , pages=. 2025 , organization=

    2025

  48. [56]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  49. [57]

    International Conference on Learning Representations (ICLR) , year =

    Contrastive Learning with Hard Negative Samples , author =. International Conference on Learning Representations (ICLR) , year =

  50. [58]

    Schroff, Florian and Kalenichenko, Dmitry and Philbin, James , booktitle =

  51. [59]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , pages =

    Deep Bayesian Active Learning with Image Data , author =. Proceedings of the 34th International Conference on Machine Learning (ICML) , pages =

  52. [60]

    International Conference on Learning Representations (ICLR) , year =

    Active Learning for Convolutional Neural Networks: A Core-Set Approach , author =. International Conference on Learning Representations (ICLR) , year =

  53. [61]

    Yin, Pengcheng and Neubig, Graham and Yih, Wen-tau and Riedel, Sebastian , booktitle =

  54. [62]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Herzig, Jonathan and Nowak, Pawe. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  55. [63]

    International Conference on Learning Representations (ICLR) , year =

    Language Models are Realistic Tabular Data Generators , author =. International Conference on Learning Representations (ICLR) , year =

  56. [64]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Language Models are Few-Shot Learners , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  57. [65]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  58. [66]

    Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , pages =

    Learning to Retrieve Prompts for In-Context Learning , author =. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , pages =

    2022

  59. [67]

    Zhang, Tianshu and Yue, Xiang and Li, Yifei and Sun, Huan , booktitle =

  60. [68]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , pages =

    Rethinking Tabular Data Understanding with Large Language Models , author =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT) , pages =

    2024

  61. [69]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Jellyfish: Instruction-Tuning Local Large Language Models for Data Preprocessing , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2024

  62. [70]

    Jiang, Jinhao and Zhou, Kun and Dong, Zican and Ye, Keming and Zhao, Wayne Xin and Wen, Ji-Rong , booktitle =

  63. [71]

    Yang, Jingfeng and Gupta, Aditya and Upadhyay, Shyam and He, Luheng and Goel, Rahul and Paul, Shachi , booktitle =

  64. [72]

    International Conference on Learning Representations (ICLR) , year =

    Chain-of-Table: Evolving Tables in the Reasoning Chain for Table Understanding , author =. International Conference on Learning Representations (ICLR) , year =

  65. [73]

    Chen, Si-An and Miculicich, Lesly and Eisenschlos, Julian Martin and Wang, Zifeng and Wang, Zilong and Chen, Yanfei and Fujii, Yasuhisa and Lin, Hsuan-Tien and Lee, Chen-Yu and Pfister, Tomas , booktitle =

  66. [74]

    Liu, Qian and Chen, Bei and Guo, Jiaqi and Ziyadi, Morteza and Lin, Zeqi and Chen, Weizhu and Lou, Jian-Guang , booktitle =

  67. [75]

    Yang, Yazheng and Wang, Yuqi and Liu, Guang and Wu, Ledell and Liu, Qi , booktitle =

  68. [76]

    Table Meets

    Sui, Yuan and Zhou, Mengyu and Zhou, Mingjie and Han, Shi and Zhang, Dongmei , booktitle =. Table Meets

  69. [77]

    Proceedings of the VLDB Endowment , volume=

    RetClean: Retrieval-Based Data Cleaning Using LLMs and Data Lakes , author=. Proceedings of the VLDB Endowment , volume=. 2024 , publisher=

    2024

  70. [78]

    Proceedings of the 2019 International Conference on Management of Data , pages=

    Raha: A configuration-free error detection system , author=. Proceedings of the 2019 International Conference on Management of Data , pages=

    2019

  71. [79]

    Enhancing Text-to- SQL Capabilities of Large Language Models: A Study on Prompt Design Strategies

    Nan, Linyong and Zhao, Yilun and Zou, Weijin and Ri, Narutatsu and Tae, Jaesung and Zhang, Ellen and Cohan, Arman and Radev, Dragomir. Enhancing Text-to- SQL Capabilities of Large Language Models: A Study on Prompt Design Strategies. Findings of the Association for Computation...

    2023 doi

  72. [80]

    LLM -Forest: Ensemble Learning of LLM s with Graph-Augmented Prompts for Data Imputation

    He, Xinrui and Ban, Yikun and Zou, Jiaru and Wei, Tianxin and Cook, Curtiss and He, Jingrui. LLM -Forest: Ensemble Learning of LLM s with Graph-Augmented Prompts for Data Imputation. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/202...

    2025 doi

  73. [81]

    Proceedings of the VLDB Endowment , volume=

    Baran: Effective error correction via a unified context representation and transfer learning , author=. Proceedings of the VLDB Endowment , volume=. 2020 , publisher=

    2020

  74. [82]

    Proceedings of the VLDB Endowment , volume=

    Deep entity matching with pre-trained language models , author=. Proceedings of the VLDB Endowment , volume=. 2020 , publisher=

    2020

  75. [83]

    Proceedings of the VLDB Endowment , volume=

    Evaluation of entity resolution approaches on real-world match problems , author=. Proceedings of the VLDB Endowment , volume=. 2010 , publisher=

    2010

  76. [84]

    2021 IEEE 37th International Conference on Data Engineering (ICDE) , pages=

    Valentine: Evaluating matching techniques for dataset discovery , author=. 2021 IEEE 37th International Conference on Data Engineering (ICDE) , pages=. 2021 , organization=

    2021

  77. [85]

    ACM SIGKDD Explorations Newsletter , volume=

    OpenML: networked science in machine learning , author=. ACM SIGKDD Explorations Newsletter , volume=. 2014 , publisher=

    2014

  78. [86]

    ACM computing surveys (CSUR) , volume=

    Anomaly detection: A survey , author=. ACM computing surveys (CSUR) , volume=. 2009 , publisher=

    2009

  79. [87]

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

    Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? , author =. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2022

  80. [88]

    What Makes Good In-Context Examples for

    Liu, Jiachang and Shen, Dinghan and Zhang, Yizhe and Dolan, Bill and Carin, Lawrence and Chen, Weizhu , booktitle =. What Makes Good In-Context Examples for

  81. [89]

    Proceedings of the 38th International Conference on Machine Learning (ICML) , pages =

    Calibrate Before Use: Improving Few-Shot Performance of Language Models , author =. Proceedings of the 38th International Conference on Machine Learning (ICML) , pages =

  82. [90]

    International Conference on Learning Representations (ICLR) , year =

    Selective Annotation Makes Language Models Better Few-Shot Learners , author =. International Conference on Learning Representations (ICLR) , year =

  83. [91]

    2026 , issue_date =

    Daviran, Mohammadreza and Lin, Brian and Rafiei, Davood , title =. 2026 , issue_date =. doi:10.14778/3797919.3797935 , journal =

    2026

  84. [92]

    arXiv preprint arXiv:2604.27261 , year=

    SynSQL: Synthesizing Relational Databases for Robust Evaluation of Text-to-SQL Systems , author=. arXiv preprint arXiv:2604.27261 , year=

    arXiv

  85. [93]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Diverse Demonstrations Improve In-context Compositional Generalization , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  86. [94]

    Wan, Zhen and Cheng, Fei and Mao, Zhuoyuan and Liu, Qianying and Song, Haiyue and Li, Jiwei and Kurohashi, Sadao , booktitle =

  87. [95]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Revisiting Demonstration Selection Strategies in In-Context Learning , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  88. [96]

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

    In-Context Learning with Iterative Demonstration Selection , author =. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =

    2024

  89. [97]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    Active Prompting with Chain-of-Thought for Large Language Models , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

  90. [98]

    Tong, Yongqi and Li, Dawei and Wang, Sizhe and Wang, Yujia and Teng, Fei and Shang, Jingbo , booktitle =. Can

  91. [99]

    Peeters, Ralph and Bizer, Christian , booktitle =. Using

  92. [100]

    arXiv preprint arXiv:2310.11244 , year =

    Entity Matching using Large Language Models , author =. arXiv preprint arXiv:2310.11244 , year =

    arXiv

  93. [101]

    arXiv preprint arXiv:2407.11852 , year =

    Schema Matching with Large Language Models: an Experimental Study , author =. arXiv preprint arXiv:2407.11852 , year =

    arXiv

  94. [102]

    arXiv preprint arXiv:2410.24105 , year =

    Matchmaker: Self-Improving Large Language Model Programs for Schema Matching , author =. arXiv preprint arXiv:2410.24105 , year =

    arXiv

  95. [103]

    arXiv preprint arXiv:2410.12480 , year=

    Kcmf: A knowledge-compliant framework for schema and entity matching with fine-tuning-free llms , author=. arXiv preprint arXiv:2410.12480 , year=

    arXiv

  96. [104]

    Anomaly Detection of Tabular Data Using

    Li, Aodong and Zhao, Yunhan and Qiu, Chen and Kloft, Marius and Smyth, Padhraic and Rudolph, Maja and Mandt, Stephan , journal =. Anomaly Detection of Tabular Data Using

  97. [105]

    arXiv preprint arXiv:2402.17944 , year=

    Large Language Models (LLMs) on Tabular Data: Prediction, Generation, and Understanding--A Survey , author=. arXiv preprint arXiv:2402.17944 , year=

    arXiv

  98. [106]

    Active Prompting with Chain-of-Thought for Large Language Models

    Diao, Shizhe and Wang, Pengcheng and Lin, Yong and Pan, Rui and Liu, Xiang and Zhang, Tong. Active Prompting with Chain-of-Thought for Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. ...

    2024 doi

  99. [107]

    C over ICL : Selective Annotation for In-Context Learning via Active Graph Coverage

    Mavromatis, Costas and Srinivasan, Balasubramaniam and Shen, Zhengyuan and Zhang, Jiani and Rangwala, Huzefa and Faloutsos, Christos and Karypis, George. C over ICL : Selective Annotation for In-Context Learning via Active Graph Coverage. Proceedings of the 2024 Conference on ...

    2024 doi

  100. [108]

    C - ICL : Contrastive In-context Learning for Information Extraction

    Mo, Ying and Liu, Jiahao and Yang, Jian and Wang, Qifan and Zhang, Shun and Wang, Jingang and Li, Zhoujun. C - ICL : Contrastive In-context Learning for Information Extraction. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.fi...

    2024 doi

  101. [109]

    Improving In-context Learning Example Retrieval for Classroom Discussion Assessment with Re-ranking and Label Ratio Regulation

    Tran, Nhat and Litman, Diane and Pierce, Benjamin and Correnti, Richard and Matsumura, Lindsay Clare. Improving In-context Learning Example Retrieval for Classroom Discussion Assessment with Re-ranking and Label Ratio Regulation. Proceedings of the 20th Workshop on Innovative ...

    2025 doi

  102. [110]

    2026 , eprint=

    LDI: Localized Data Imputation for Text-Rich Tables , author=. 2026 , eprint=

    2026

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.