Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Asking an LLM to induce executable decision trees detects tabular errors more accurately, explainably, and robustly than having the LLM label cells directly.

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

2026-08-03 18:00 UTC pith:LGFAV2P7

load-bearing objection Interesting hybrid LLM-as-inducer method, but the headline F1 gain is measured on different test sets for baselines vs. the proposed method; needs a matched re-run before the claim holds. the 4 major comments →

arxiv 2512.07246 v3 pith:LGFAV2P7 submitted 2025-12-08 cs.CL

Ensembling LLM-Induced Decision Trees for Explainable and Robust Error Detection

Mengqi Wang (1) , Jianwei Wang (1) , Qing Liu (2) , Xiwei Xu (2) , Zhenchang Xing (2) , Liming Zhu (2) , Michael Bain (1) , Wenjie Zhang (1) ((1) UNSW Sydney
show 2 more authors
(2) Data61 CSIRO)
This is my paper
classification cs.CL
keywords error detectiontabular dataLLM-as-an-inducerdecision treesgraph neural networksensemble consensusexpectation-maximizationdata cleaning
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.

The paper tries to establish that a large language model is more useful as a decision-tree inducer than as a direct cell labeler for tabular error detection. It proposes TreeED, which prompts an LLM to output an executable decision tree whose nodes are either Python rule checks or trainable graph-neural-network checks, plus labeled decision paths. ForestED then induces several such trees on different informative row samples and combines them through an expectation-maximization consensus that estimates each tree's reliability. The authors report an average F1 improvement of about 16.1% over the strongest existing LLM-based method, while every prediction carries a traceable root-to-leaf explanation. A sympathetic reader would care because error detection is the first step in data cleaning, and current LLM approaches are accurate but opaque and unstable.

Core claim

TreeED prompts an LLM with a data profile and up to 100 sampled rows to output a decision-tree skeleton plus labeled decision paths. Rule nodes are inserted as executable Python code without further processing; GNN nodes are trained on the LLM's own path labels over a bipartite tuple-attribute graph; leaf nodes emit error or clean. ForestED samples multiple informative row subsets via Gaussian-process uncertainty sampling, induces a tree per subset, and runs an expectation-maximization algorithm that jointly estimates each tree's confusion matrix and the latent true cell labels, taking the MAP label per cell as the final consensus.

What carries the argument

The central object is the LLM-induced executable decision tree with three node types: rule nodes that run simple validation checks, GNN nodes that learn complex relational patterns such as functional dependencies, and leaf nodes that output error or clean. The load-bearing mechanism is the decision path: each cell's verdict is produced by a root-to-leaf sequence of Boolean checks, and GNN nodes generalize relational constraints from the LLM-supplied path labels. The EM ensemble—modeling each tree's 2x2 confusion matrix and the prior error rate as latent variables—converts multiple stochastic trees into a stable, reliability-weighted consensus.

Load-bearing premise

The whole pipeline trusts the LLM's induced output without verification: generated rule code is executed as-is, and GNN nodes are trained on the LLM's own path labels, so a hallucinated or over-broad rule turns into an executable false-positive generator on every unsampled row.

What would settle it

Take a table where the LLM generates a rule with an inverted comparison and show that unsampled rows matching the true condition are flagged as errors at the rule's false-positive rate. Alternatively, compare each GNN node's learned branch decisions against its LLM-written description on held-out cells; if the two disagree on a substantial share, the claim that GNN nodes implement the described checks is falsified.

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

If this is right

  • A handful of sampled rows (at most 100) can yield a full-table detector, removing the need to call an LLM per cell.
  • Each cell's verdict is accompanied by a decision path naming the exact rule or relational check that flagged it, making predictions auditable.
  • The EM consensus exposes per-tree reliability scores that track each tree's true F1, so analysts can see which reasoning patterns contributed to a verdict.
  • The ensemble buffers LLM stochasticity and backbone choice, shrinking variance across runs and across different models.

Where Pith is reading between the lines

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

  • Because the rule nodes are ordinary Python functions, a user could in principle edit them directly and re-run detection, turning the detector into a human-in-the-loop artifact—an extension the paper does not explore.
  • The same LLM-as-inducer recipe could transfer to other tabular tasks that need explainable predictions, such as schema matching or entity resolution, wherever a tree skeleton can encode the decision logic.
  • The EM consensus only sees prediction matrices, so ForestED could be applied to any set of detectors, not just LLM-induced trees, making it a generic aggregation layer for heterogeneous error-detection signals.

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 / 5 minor

Summary. The paper proposes TreeED and ForestED, an LLM-as-an-inducer framework for tabular error detection. TreeED prompts an LLM to emit an executable decision-tree skeleton with rule nodes, GNN nodes, and leaf nodes; GNN nodes are trained on decision-path labels produced by the LLM over sampled rows. ForestED builds multiple such trees on uncertainty-sampled row partitions and combines their full-table predictions via an EM-based consensus procedure that estimates per-tree confusion matrices and a global error prior. The authors claim state-of-the-art accuracy, explainability, and robustness, reporting an average F1 improvement of 16.1% over the strongest baseline (ZeroED), with ablations showing contributions from both GNN nodes and the ensemble.

Significance. If the claims hold, the framework is a meaningful step toward more transparent and robust LLM-based data cleaning: it replaces black-box LLM labeling with a structured, executable detector, uses an EM ensemble to mitigate LLM stochasticity, and ships code and artifacts. The idea of inducing hybrid rule/GNN decision trees from LLM outputs is original and practically relevant. The paper also provides a formal Lemma 1 with proof for the EM updates and an ablation study supporting the design choices. However, the significance is contingent on the experimental comparison being fair and on the induced rules/GNN nodes actually implementing what the explanations claim.

major comments (4)
  1. [§6.1, Table 3] The headline 16.1% F1 gain is computed on non-comparable test sets. Section 6.1 states that LLM baselines are evaluated on randomly sampled 50 rows (FM_ED) or 500 rows (ZeroED), while ForestED's predictions are generated over the full table (Algorithm 2 loops over all N×M cells and no sampling cap is mentioned). F1 scores on different subsets are not comparable because error rate and class balance vary across rows; for Tax (200,000 rows, 4.05% errors), a 500-row sample may contain too few errors to give a stable ZeroED F1. The paper should re-run baselines on exactly the same rows used for ForestED (or, if this is infeasible, report ForestED on the same 500-row sample) and recompute the claimed improvement. This is load-bearing for the central state-of-the-art claim.
  2. [§4.2, Algorithm 1/2] The evaluation is in-sample for the sampled rows. TreeED induces the tree from the sampled set S (Algorithm 1, lines 1–10), trains each GNN node on the LLM's decision-path labels for those cells (lines 10–14), and then Algorithm 2 applies the tree to all rows, including the same sampled rows. For small datasets such as Rayyan (N=1,000, up to 100 sampled rows, i.e., 10% of the data), this can materially inflate F1 because the tree has effectively seen those cells during induction/training. Report results with the sampled rows excluded from the evaluation set, or at least quantify the difference.
  3. [§4.2, Algorithm 1 lines 3–4] Rule nodes are inserted 'without further processing' and GNN supervision comes directly from the LLM's path labels. There is no verification of rule correctness or label accuracy. A hallucinated or over-broad rule is then executed verbatim on every unsampled row, and a mislabeled sampled cell trains its GNN node to imitate that mistake. The paper presents no analysis of rule validity, no error analysis, and no sensitivity test to LLM label noise. Since the method's executable detector is only as trustworthy as the unverified LLM output, the accuracy and explainability claims need support from a rule-correctness audit or a robustness experiment with injected label noise.
  4. [Remark 1, §4.2 GNN training] The claim that GNN nodes 'remain explainable' is not established. Remark 1 argues that each GNN node has an LLM-written descriptive message and is dedicated to a specific function, but the training objective (binary cross-entropy on branch labels) does not constrain the learned GNN to implement the declared check. The GNN could learn any function consistent with the path labels, so the descriptive message is an attribution rather than a faithful explanation. The authors should either provide evidence that the trained GNN indeed implements the declared relational check (e.g., targeted probes or counterfactual tests) or soften the explainability claim.
minor comments (5)
  1. [Abstract vs. §6.2] The abstract says 'average F1 improvement of approximately 16.1%' while §6.2 says 'average absolute gains of 16.1%'. These are different: the relative improvement over ZeroED is about 27% (0.752 vs. 0.591). Please state clearly whether the reported number is an absolute percentage-point gain or a relative improvement.
  2. [§5.1] The Gaussian Process kernel uses a length-scale chi, but its value or tuning procedure is not reported. Since the uncertainty sampler directly affects which rows are selected and thus the induced trees, please provide the setting or a sensitivity analysis.
  3. [Table 5] Robustness variance is measured on only two datasets (Beers and Billionaire). The claim of a 7×/10× reduction in F1 standard deviation would be more convincing if reported on all seven datasets or a representative larger subset.
  4. [§6.1] The implementation details state that ZeroED uses the same sampling ratio without a cap, but Section 6.1 also says ZeroED is evaluated on 500 randomly sampled rows. Please clarify how the 500-row limit interacts with the ratio and whether the reported ZeroED F1 is based on those 500 rows only.
  5. [Table 7] The '–' entries for tree numbers 2 and 5 on Movies are explained by LLM timeout; consider reporting the timeout threshold or re-running with a lower temperature to avoid missing data in the ablation.

Circularity Check

0 steps flagged

No circularity in the derivation chain; the main quantitative claim is threatened by an evaluation mismatch, which is a correctness risk rather than a circular step.

full rationale

The paper's derivation chain is a pipeline, not a reduction. TreeED prompts an LLM to output a tree skeleton and path labels; rule nodes execute LLM-generated code, and GNN nodes are trained on branch labels extracted from the same LLM's decision paths. This is self-training/distillation on pseudo-labels, not a definitional equivalence: the reported F1 is computed on full tables, while the supervision labels cover only a small sampled subset, so the full-table predictions are not statistically forced to equal the input labels. The EM-based consensus is a standard Dawid-Skene-style procedure; the paper provides its own ELBO proof (Lemma 1, Section 5.3), so the citation to the authors' prior work [39] is not load-bearing. The 'explainability' claim for GNN nodes rests on LLM-written messages rather than verification of learned behavior, but that is an unsupported characterization, not a circular derivation. The most serious concern is that Table 3 compares baselines evaluated on 50/500-row samples with ForestED evaluated on full tables ('Due to the high token usage of LLM-based baselines, we randomly sample 50 rows for FM_ED and 500 rows for ZeroED for evaluation'), making the 16.1% improvement not a like-for-like comparison. That is an experimental confound and a correctness risk, not a circularity of the kind defined here. No load-bearing step reduces by construction to its own inputs, so the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

This is an ML-systems paper, so the ledger records hyperparameters and behavioral assumptions about the external LLM rather than physical constants. The most serious entries are the unverified-LLM axiom and the unspecified GP lengthscale. No new theoretical entities are introduced.

free parameters (6)
  • sampling_ratio_rho = 0.05 (cap 100 tuples)
    Hand-chosen; controls how many rows are LLM-labeled per dataset and indirectly caps the tree count.
  • partition_size = 10 tuples per tree
    Hand-chosen; determines the number of induced trees R and their diversity.
  • tree_depth = 4-8 levels
    Selected after Exp-9 showed 2-6 underfits and 6-10 overfits on the same benchmark datasets used for headline results.
  • number_of_trees_R = ~10
    Selected after Exp-8 (Table 7): 10 trees best, 20 degrades, 2/5 suboptimal, with LLM timeouts at higher counts.
  • LLM_temperature = 1.0
    Fixed to maximize stochasticity; the robustness claims depend on this choice.
  • GP_kernel_lengthscale_chi = unspecified
    The RBF kernel in Sec 5.1 requires a lengthscale; its value and fitting procedure are never given.
axioms (5)
  • standard math EM updates monotonically improve the marginal log-likelihood (Lemma 1)
    Standard EM/ELBO argument; the proof in Sec 5.3 is correct, but the phrasing 'improves consensus quality' goes beyond what the lemma strictly shows.
  • domain assumption Ground-truth error matrices D* of the seven benchmark datasets are accurate
    All F1 numbers assume the benchmark corruptions are the correct label source (Sec 6.1).
  • domain assumption Bipartite tuple-attribute graph suffices to represent relational error patterns
    Sec 4.2 adopts the GRAPE-style graph; whether one shared graph per node captures cross-column FDs is assumed, not validated.
  • ad hoc to paper GPT-5 reliably induces correct, executable rules and accurate per-cell labels with no verification
    Rule code is inserted 'without further processing' (Sec 4.2) and GNN labels come from LLM decision paths; this is the load-bearing behavioral assumption.
  • ad hoc to paper GNN nodes trained on up to 100 sampled rows of LLM labels generalize to the full table
    Sec 4.2: each GNN node is trained only on cells whose path passes through it from the sampled label set, then applied to all N x M cells.

pith-pipeline@v1.3.0-alltime-deepseek · 4607 in / 8216 out tokens · 240963 ms · 2026-08-03T18:00:37.402524+00:00 · methodology

0 comments
read the original abstract

Error detection (ED), which aims to identify incorrect or inconsistent cell values in tabular data, is important for ensuring data quality. Recent state-of-the-art ED methods leverage the pre-trained knowledge and semantic capability of large language models (LLMs) to directly label whether a cell is erroneous. However, this LLM-as-a-labeler pipeline produces predictions through an implicit black-box process with limited traceability and explicit justification, and relies on stochastic single-pass inference, resulting in inconsistent and insufficiently robust detections across contexts. To address these limitations, we propose an LLM-as-an-inducer framework that uses an LLM to induce a decision tree for ED, termed TreeED, and ensembles multiple such trees for consensus detection, termed ForestED. Based on prompts derived from data context, decision tree specifications, and output requirements, TreeED queries the LLM to induce a decision tree skeleton whose root-to-leaf paths specify the stepwise procedure for evaluating a sample. Each tree contains three types of nodes: (1) rule nodes that perform simple validation checks, such as format or range constraints; (2) Graph Neural Network (GNN) nodes that capture complex patterns, such as functional dependencies; and (3) leaf nodes that output the final decision as error or clean. ForestED employs uncertainty sampling to obtain multiple informative row subsets and constructs a decision tree for each subset using TreeED. It then applies an Expectation-Maximization-based algorithm to jointly estimate tree reliability and optimize the consensus ED prediction. Experiments demonstrate that our methods are accurate, explainable, and robust, achieving an average F1-score improvement of 16.1% over the best baseline.

Figures

Figures reproduced from arXiv: 2512.07246 by (2) Data61, CSIRO), Jianwei Wang (1), Liming Zhu (2), Mengqi Wang (1), Michael Bain (1), Qing Liu (2), Wenjie Zhang (1) ((1) UNSW Sydney, Xiwei Xu (2), Zhenchang Xing (2).

Figure 1
Figure 1. Figure 1: An illustrated example of common data errors. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework comparisons of LLM-based ED methods. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Framework overview of TreeED and ForestED. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Runtime and token cost across datasets [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of varying the number of labeled records on [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Correlation between F1 and ensemble reliability. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: An example of decision tree. 690 and column City, where the tuple contains the value “birminx￾ham” and the correct value is “birmingham”. One of the decision trees induced for this dataset is shown in [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Interpretable Column Annotation with LLM-Symbolized Decision Process Materialization

    cs.CL 2026-07 conditional novelty 6.0

    SymCA reaches state-of-the-art column type and property annotation by inducing a label-skeleton with LLMs and evolving per-node random-forest operators.

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

    cs.DB 2026-06 unverdicted novelty 6.0

    LasRepair++ pairs an LLM instructor with an SLM corrector, refines context via EM, and down-weights uncertain repairs using column-calibrated confidence, reporting 18.1% average F1 gain over baselines on data repair tasks.

Reference graph

Works this paper leans on

55 extracted references · 10 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Mohamed Abdelaal, Tim Ktitarev, Daniel Städtler, and Harald Schöning. 2024. SAGED: Few-Shot Meta Learning for Tabular Data Error Detection.. InEDBT. 386–398

    2024

  2. [2]

    Tommaso Bendinelli, Artur Dox, and Christian Holz. 2025. Exploring llm agents for cleaning tabular machine learning datasets.arXiv preprint arXiv:2503.06664 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    Tobias Bleifuß, Sebastian Kruse, and Felix Naumann. 2017. Efficient denial constraint discovery with hydra.Proceedings of the VLDB Endowment11, 3 (2017), 311–323

    2017

  4. [4]

    Lucas Carrasco, Felipe Urrutia, and AndrÊs Abeliuk. 2025. Zero-shot decision tree construction via large language models.arXiv preprint arXiv:2501.16247 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Jiahui Chen, Yu Sun, Shaoxu Song, Haiwei Zhang, and Xiaojie Yuan. 2025. Mini- mum Change≠ Best Cleaning: Parallel and Incremental Error Detection under Integrity Constraints.Proceedings of the ACM on Management of Data3, 4 (2025), 1–26

    2025

  6. [6]

    Xu Chu, Ihab F Ilyas, Sanjay Krishnan, and Jiannan Wang. 2016. Data cleaning: Overview and emerging challenges. InProceedings of the 2016 international conference on management of data. 2201–2206

    2016

  7. [7]

    Alexander Philip Dawid and Allan M Skene. 1979. Maximum likelihood esti- mation of observer error-rates using the EM algorithm.Journal of the Royal Statistical Society: Series C (Applied Statistics)28, 1 (1979), 20–28

    1979

  8. [8]

    Amr Ebaid, Ahmed Elmagarmid, Ihab F Ilyas, Mourad Ouzzani, Jorge-Arnulfo Quiane-Ruiz, Nan Tang, and Si Yin. 2013. NADEEF: A generalized data cleaning system.Proceedings of the VLDB Endowment6, 12 (2013), 1218–1221

    2013

  9. [9]

    Wenfei Fan, Floris Geerts, Xibei Jia, and Anastasios Kementsietsidis. 2008. Con- ditional functional dependencies for capturing data inconsistencies.ACM Trans- actions on Database Systems (TODS)33, 2 (2008), 1–48

    2008

  10. [10]

    Saeed Fathollahzadeh, Essam Mansour, and Matthias Boehm. 2025. CatDB: Data- catalog-guided, LLM-based Generation of Data-centric ML Pipelines.Proceedings of the VLDB Endowment18, 8 (2025), 2639–2652

    2025

  11. [11]

    Congcong Ge, Yunjun Gao, Xiaoye Miao, Bin Yao, and Haobo Wang. 2022. A Hybrid Data Cleaning Framework Using Markov Logic Networks.IEEE Trans- actions on Knowledge and Data Engineering34, 5 (2022), 2048–2062. https: //doi.org/10.1109/TKDE.2020.3012472

    arXiv 2022

  12. [12]

    Nitin Gupta, Shashank Mujumdar, Hima Patel, Satoshi Masuda, Naveen Panwar, Sambaran Bandyopadhyay, Sameep Mehta, Shanmukha Guttula, Shazia Afzal, Ruhi Sharma Mittal, et al . 2021. Data quality for machine learning tasks. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 4040–4041

    2021

  13. [13]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems30 (2017)

    2017

  14. [14]

    Alireza Heidari, Joshua McGrath, Ihab F Ilyas, and Theodoros Rekatsinas. 2019. Holodetect: Few-shot learning for error detection. InProceedings of the 2019 International Conference on Management of Data. 829–846

    2019

  15. [15]

    Tzu-Heng Huang, Catherine Cao, Vaishnavi Bhargava, and Frederic Sala. 2024. The alchemist: Automated labeling 500x cheaper than llm data annotators.Ad- vances in Neural Information Processing Systems37 (2024), 62648–62672

    2024

  16. [16]

    Zhipeng Huang and Yeye He. 2018. Auto-detect: Data-driven error detection in tables. InProceedings of the 2018 International Conference on Management of Data. 1377–1392

    2018

  17. [17]

    Dongjun Jang, Youngchae Ahn, and Hyopil Shin. 2025. RCScore: Quantifying Response Consistency in Large Language Models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 5701–5719

    2025

  18. [18]

    M Kopanja, M Savić, and L Longo. [n.d.]. CORTEX: A Cost-Sensitive Rule and Tree Extraction Method, 2025.URL: https://arxiv. org/abs/2502.03200([n. d.])

    Pith/arXiv arXiv 2025

  19. [19]

    Sanjay Krishnan, Jiannan Wang, Eugene Wu, Michael J Franklin, and Ken Gold- berg. 2016. Activeclean: Interactive data cleaning for statistical modeling.Pro- ceedings of the VLDB Endowment9, 12 (2016), 948–959

    2016

  20. [20]

    David D Lewis and Jason Catlett. 1994. Heterogeneous uncertainty sampling for supervised learning. InMachine learning proceedings 1994. Elsevier, 148–156

    1994

  21. [21]

    Guoliang Li, Xuanhe Zhou, and Xinyang Zhao. 2024. Llm for data management. Proceedings of the VLDB Endowment17, 12 (2024), 4213–4216

    2024

  22. [22]

    Peng Li, Xi Rao, Jennifer Blase, Yue Zhang, Xu Chu, and Ce Zhang. 2021. Cleanml: A study for evaluating the impact of data cleaning on ml classification tasks. In 2021 IEEE 37th International Conference on Data Engineering (ICDE). IEEE, 13–24

    2021

  23. [23]

    Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey.arXiv preprint arXiv:2406.15126(2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Mohammad Mahdavi and Ziawasch Abedjan. 2021. Semi-Supervised Data Clean- ing with Raha and Baran.. InCIDR

    2021

  25. [25]

    Mohammad Mahdavi, Ziawasch Abedjan, Raul Castro Fernandez, Samuel Mad- den, Mourad Ouzzani, Michael Stonebraker, and Nan Tang. 2019. Raha: A configuration-free error detection system. InProceedings of the 2019 Interna- tional Conference on Management of Data. 865–882

    2019

  26. [26]

    Katie Matton, Robert Osazuwa Ness, John Guttag, and Emre Kıcıman. 2025. Walk the talk? Measuring the faithfulness of large language model explanations.arXiv preprint arXiv:2504.14150(2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Avanika Narayan, Ines Chami, Laurel Orr, Simran Arora, and Christopher Ré

  28. [28]

    Mona Nashaat, Aindrila Ghosh, James Miller, and Shaikh Quader. 2021. TabRe- former: Unsupervised representation learning for erroneous data detection. ACM/IMS Transactions on Data Science2, 3 (2021), 1–29

    2021

  29. [29]

    Wei Ni, Kaihang Zhang, Xiaoye Miao, Xiangyu Zhao, Yangyang Wu, Yaoshu Wang, and Jianwei Yin. 2025. ZeroED: Hybrid Zero-shot Error Detection through Large Language Model Reasoning. In2025 IEEE 41st International Conference on Data Engineering (ICDE). IEEE Computer Society, 3126–3139

    2025

  30. [30]

    Marcel Parciak, Brecht Vandevoort, Frank Neven, Liesbet M Peeters, and Stijn Vansummeren. 2024. Schema matching with large language models: an experi- mental study.arXiv preprint arXiv:2407.11852(2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Eduardo HM Pena, Eduardo C de Almeida, and Felix Naumann. 2021. Fast detection of denial constraint violations.Proceedings of the VLDB Endowment15, 4 (2021), 859–871

    2021

  32. [32]

    Jeremy Petch, Shuang Di, and Walter Nelson. 2022. Opening the black box: the promise and limitations of explainable machine learning in cardiology.Canadian Journal of Cardiology38, 2 (2022), 204–213

    2022

  33. [33]

    Minh Pham, Craig A Knoblock, Muhao Chen, Binh Vu, and Jay Pujara. 2021. SPADE: A Semi-supervised Probabilistic Approach for Detecting Errors in Tables.. InIJCAI. 3543–3551

    2021

  34. [34]

    Clement Pit-Claudel, Zelda Mariet, Rachael Harding, and Sam Madden. 2016. Outlier detection in heterogeneous datasets using automatic tuple expansion. (2016)

    2016

  35. [35]

    Jiyuan Ren, Zhaocheng Du, Zhihao Wen, Qinglin Jia, Sunhao Dai, Chuhan Wu, and Zhenhua Dong. 2025. Few-shot LLM Synthetic Data with Distribution Matching. InCompanion Proceedings of the ACM on Web Conference 2025. 432– 441

    2025

  36. [36]

    Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, and Wenjie Zhang

  37. [37]

    Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, and Wenjie Zhang. 2025. HydraRAG: Structured Cross-Source Enhanced Large Language Model Reasoning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 14442–14470

    2025

  38. [38]

    Jianwei Wang, Kai Wang, Ying Zhang, Wenjie Zhang, Xiwei Xu, and Xuemin Lin. 2025. On LLM-Enhanced Mixed-Type Data Imputation with High-Order Message Passing.Proc. VLDB Endow.18, 10 (Sept. 2025), 3421–3434. https: //doi.org/10.14778/3748191.3748205

    arXiv 2025

  39. [39]

    Jianwei Wang, Yuehai Wang, Kai Wang, Xuemin Lin, Wenjie Zhang, and Ying Zhang. 2025. Ensemble-based Deep Multilayer Community Search.arXiv preprint arXiv:2501.02194(2025)

    Pith/arXiv arXiv 2025

  40. [40]

    Jianwei Wang, Ying Zhang, Kai Wang, Xuemin Lin, and Wenjie Zhang. 2024. Missing data imputation with uncertainty-driven network.Proceedings of the ACM on Management of Data2, 3 (2024), 1–25

    2024

  41. [41]

    Tianshu Wang, Xiaoyang Chen, Hongyu Lin, Xuanang Chen, Xianpei Han, Le Sun, Hao Wang, and Zhenyu Zeng. 2025. Match, compare, or select? an investigation of large language models for entity matching. InProceedings of the 31st International Conference on Computational Linguistics. 96–109

    2025

  42. [42]

    Xinru Wang, Hannah Kim, Sajjadur Rahman, Kushan Mitra, and Zhengjie Miao

  43. [43]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reason- ing in large language models.Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  44. [44]

    Yangyang Wu, Chen Yang, Mengying Zhu, Xiaoye Miao, Wei Ni, Meng Xi, Xinkui Zhao, and Jianwei Yin. 2025. A Zero-Training Error Correction System with Large Language Models. In2025 IEEE 41st International Conference on Data Engineering (ICDE). 2949–2962. https://doi.org/10.1109/ICDE65448.2025.00221

    arXiv 2025

  45. [45]

    Sichao Xiong, Yigit Ihlamur, Fuat Alican, and Aaron Ontoyin Yin. 2024. Gptree: Towards explainable decision-making via llm-powered decision trees.arXiv preprint arXiv:2411.08257(2024)

    Pith/arXiv arXiv 2024

  46. [46]

    Jiaxuan You, Xiaobai Ma, Yi Ding, Mykel J Kochenderfer, and Jure Leskovec

  47. [47]

    Xiao Yu, Zexian Zhang, Feifei Niu, Xing Hu, Xin Xia, and John Grundy. 2024. What Makes a High-Quality Training Dataset for Large Language Models: A Practitioners’ Perspective. InProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering. 656–668

    2024

  48. [48]

    Haochen Zhang, Yuyang Dong, Chuan Xiao, and Masafumi Oyamada. 2023. Jellyfish: A large language model for data preprocessing.arXiv preprint arXiv:2312.01678(2023). 13

    Pith/arXiv arXiv 2023

  49. [49]

    Shuo Zhang, Zezhou Huang, and Eugene Wu. 2025. Data cleaning using large language models. In2025 IEEE 41st International Conference on Data Engineering Workshops (ICDEW). IEEE, 28–32

    2025

  50. [50]

    Kangfei Zhao, Jeffrey Xu Yu, Hao Zhang, Qiyan Li, and Yu Rong. 2021. A learned sketch for subgraph counting. InProceedings of the 2021 International Conference on Management of Data. 2142–2155

    2021

  51. [51]

    Junhao Zhu, Tao Wang, Danlei Hu, Ziquan Fang, Lu Chen, Yunjun Gao, Tianyi Li, and Christian S. Jensen. 2024. T-Assess: An Efficient Data Quality Assessment System Tailored for Trajectory Data.Proc. VLDB Endow.18, 3 (Nov. 2024), 666–674. https://doi.org/10.14778/3712221.3712233 14

    arXiv 2024

  52. [2020]

    Handling missing data with graph representation learning.Advances in Neural Information Processing Systems33 (2020), 19075–19087

    2020

  53. [2022]

    Can foundation models wrangle your data?arXiv preprint arXiv:2205.09911 (2022)

    Pith/arXiv arXiv 2022

  54. [2024]

    InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems

    Human-llm collaborative annotation through effective verification of llm labels. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–21

    2024

  55. [2025]

    InProceedings of the ACM on Web Conference 2025

    Paths-over-graph: Knowledge graph empowered large language model reasoning. InProceedings of the ACM on Web Conference 2025. 3505–3522

    2025