Does Prompt Design Impact Quality of Data Imputation by LLMs?

Lydia Manikonda; Shreenidhi Srinivasan

REVIEW 4 major objections 6 minor 29 references

Does Prompt Design Impact Quality of Data Imputation by LLMs?

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that restricting LLM imputation prompts to features strongly correlated with the missing column reduces token use while maintaining or improving imputation quality, with the largest gains on small class-imbalanced…

desk verdict A credible but narrow empirical result: pruning features in LLM imputation prompts helps on small datasets, but the evaluation is too indirect to fully support the general claim. read the letter →

arxiv 2506.04172 v1 pith:MIGFJLZP submitted 2025-06-04 cs.LG cs.ET

Shreenidhi Srinivasan , Lydia Manikonda This is my paper

classification cs.LGcs.ET

keywords dataimputationlargelanguagemodelspromptengineeringclassimbalancetabularfeatureselectionin-contextlearningsyntheticgeneration

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 tries to establish that prompt design is a lever for LLM-based tabular data imputation: a group-wise CSV-style prompt that includes only the features most correlated with the column being imputed can match or beat a prompt using all features, while consuming far fewer tokens. If true, this makes LLM imputation cheaper and opens a practical route to filling missing values in small class-imbalanced datasets, where minority-class performance often suffers. The claim is supported by experiments on two Kaggle datasets (Adult Income and Travel) with classification-based metrics from XGBoost and Random Forest, showing that aggressive feature pruning (up to about 76% of columns) leaves minority-class F1 essentially unchanged on the large dataset and improves it on the small one.

What carries the argument

The central mechanism is the group-wise CSV-style prompt combined with correlation-threshold feature selection. The prompt template begins with an instruction, feature descriptions, a header, and groups of completed examples sampled separately for each target class, so the LLM sees majority and minority distributions in a balanced way; the model is told to output only the imputed column to save tokens. Around this template sits a feature-selection step that computes pairwise association strengths (Pearson for numeric-numeric, Cramer's V for categorical-categorical, eta ratio for mixed pairs), ranks them, and chooses a correlation threshold at the elbow of the ranked plot; only features above the threshold enter the prompt. The imputation order is chosen so that the feature relevant to the most other imputation features goes first, and each completed column is folded back into the prompt for the next feature. This mechanism is what lets the method shrink the prompt while directing the LLM's attention to the information that carries the missing value.

What would settle it

For the Travel dataset, retrieve the 120 artificially removed 'FrequentFlyer' entries and compare the LLM's imputed values to these ground-truth values; if the per-value accuracy is at chance level while classification metrics remain high, the claim that imputation quality is 'maintained or improved' would be shown to rest entirely on the classification proxy rather than on factually correct imputation.

Watch

Extended reading notes

Core claim

The central claim is that discarding dataset features with weak associations with the feature being imputed yields imputation quality at least as good as using the complete dataset, and sometimes better. The paper demonstrates this with a token-aware imputation method built on a group-wise CSV-style prompt: completed records are presented in class-separated groups, and the prompt instructs the model to output only the imputed column. Features are selected by computing Pearson correlation, Cramer's V, and eta correlation ratio between the imputation feature and all other columns, ranking the absolute correlations, and picking a threshold at the elbow of the ranked plot. The imputation proceeds one feature at a time, starting with the feature relevant to the most other imputation features, and the completed column replaces the missing column in later prompts. On the Adult Income dataset, the minority-class F1 stays within 0.04 while the feature space is cut by about 76%; on the Travel dataset, omitting weakly correlated features raises minority-class F1 by about 7.46%, balanced accuracy by 4.45%, and ROC AUC from about 0.91 to 0.95.

Load-bearing premise

The entire evaluation assumes that classification performance on the imputed rows is a valid measure of imputation quality, and the paper never checks the imputed values against the true values that are known for the Travel dataset's intentionally introduced missingness.

Editorial extensions

If this is right

LLM imputation costs drop substantially with no penalty in classification-based imputation quality when weakly correlated features are excluded, so users can fit more in-context examples within a fixed token budget.
Small class-imbalanced datasets benefit most: Travel-style datasets with few records saw minority-class F1, balanced accuracy, and ROC AUC all improve when the prompt was pruned.
For larger datasets such as Adult Income, aggressive pruning (around 76% of columns) preserves minority-class F1 within 0.04 and keeps balanced accuracy and ROC AUC nearly stable, suggesting the method trades tokens for little or no quality.
Because the completed column is re-inserted into later prompts, the approach chains imputations so that the information available for successive missing features grows, which is a direct corollary of the method's design.

Reading between the lines

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

The paper's evaluation never compares imputed values to ground truth, even though the Travel dataset's artificial missingness preserves the true values; a natural next step would be to measure token-level imputation accuracy directly, which could either corroborate or undercut the classification-based conclusion.
The correlation threshold is chosen per dataset by visual elbow inspection; our inference is that a fixed or automated threshold-selection rule (e.g., a permutation-based null) would make the method more reproducible and would clarify whether the Travel gains are sensitive to the exact cutoff.
The paper's explanation for why pruning helps small datasets—that a small sample cannot let the model filter noise by itself—suggests a testable extension: vary the number of in-context examples while holding the feature set fixed, and the benefit of pruning should shrink as sample size grows.
Since only categorical imputation features were tested, the method's applicability to numeric missing values is open; our inference is that the same correlation-based pruning should carry over, but numeric imputation errors are more likely to be visible to a numeric metric, so the classification proxy may overstate quality.

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.

Desk Editor's Note

read the letter

Colleague,

Quick take: the paper has a modest, credible empirical finding—dropping weakly correlated columns from a group-wise CSV prompt can improve LLM imputation quality on a small dataset (Travel) while cutting token costs. The same effect is largely absent on Adult, where minority-class F1 actually drops slightly (0.77→0.74 with XGBoost) even though they phrase it as "maintaining." So the headline claim is overstated.

What's genuinely good: they take Kim et al.'s group-wise prompting from synthetic data generation and adapt it to imputation, which is a sensible extension. The prompt template is clear and reproducible, and they are explicit about feature descriptions and correlation thresholds. The Travel result is a real positive: pruning 2–4 of 6 features improved F1 and ROC AUC consistently across two classifiers. That's enough to justify a follow-up.

Where it's soft: the main issue is the evaluation proxy. They train classifiers on original rows and test on imputed rows, but never compare imputed values against ground truth, even though Travel's missingness is artificially generated and the true values are known. Classification performance can shift for reasons unrelated to imputation accuracy—e.g., feature pruning changes the classifier's input space. So the stress-test note is fair: the "maintaining or improving imputation quality" claim is not fully established. The paper does disclose this methodology, so it's not hiding anything, but the abstract overreaches.

Other soft spots: no repeated runs or confidence intervals; a single GPT-4.1 pass means the differences could be noise. The correlation threshold is chosen from the same data, which is a mild selection bias. There are no standard imputation baselines (mode, median, MICE), so we don't know how LLM imputation compares to the obvious alternatives. All of these are fixable.

Who it's for: practitioners who want a quick, token-efficient way to impute categorical columns in small imbalanced datasets, and researchers working on prompt design for LLM-based tabular data. It deserves a serious referee, but it needs another round of experiments to support the general claim.

Referee Report

4 major / 6 minor

Summary. This paper proposes a token-aware LLM-based imputation method for class-imbalanced tabular data. The method uses group-wise CSV-style prompts (building on Kim et al.'s EPIC) and prunes features that are weakly correlated with the feature being imputed, based on an elbow of sorted absolute correlations (Pearson/Cramér's V/η). The authors evaluate on Adult Income and Travel datasets, training XGBoost and Random Forest on complete rows and testing on imputed rows. They report precision/recall/F1, balanced accuracy, and ROC AUC at correlation thresholds 0, 0.15/0.2. The central claim is that feature pruning maintains or improves imputation quality while reducing prompt tokens, with a stronger benefit on the smaller Travel dataset.

Significance. If the claims held, the work would be a practical contribution: cheaper LLM imputation for imbalanced datasets by omitting irrelevant features. The paper is transparent about its setup and includes a grouped-vs-ungrouped prompt comparison (Table 2) that supports the benefit of structured grouping. However, the evaluation is indirect (classifier scores rather than fidelity of imputed values), single-run, and the Adult results show a small decline rather than maintenance; the Limitations section acknowledges the categorical-only scope. The evidence is therefore suggestive rather than conclusive.

major comments (4)

[Data Imputation Technique / Experiments and Insights (Tables 3-4)] The central claim that pruning weakly associated features achieves 'an imputation quality at least as good as using the complete dataset information' is not directly tested. The Travel dataset includes artificially introduced missingness with known ground-truth values, yet the paper never compares the imputed FrequentFlyer values against those true values. Because pruning changes the feature space seen by the downstream classifiers, classification metrics can improve or decline even if the imputed values are equally correct. The improvements in Tables 3-4 therefore do not by themselves establish improved imputation quality. Please report direct accuracy metrics for the Travel dataset (e.g., imputation accuracy, per-class confusion between true and imputed values), and at least discuss the relationship between the proxy metrics and imputation fidelity.
[Tables 5-6] The description of the Adult Income results as 'maintained' is contradicted by the reported numbers. At threshold 0.2, XGBoost minority-class F1 drops from 0.77 (threshold 0) to 0.74, and balanced accuracy from 0.814 to 0.792; Random Forest shows the same pattern (F1 0.78 to 0.74, balanced accuracy 0.813 to 0.790). A drop of 0.03-0.04 F1 and about 2 percentage points in balanced accuracy is a decline, not maintenance. Please either revise the claim to 'minor decline' or provide statistical evidence (e.g., confidence intervals) that these differences are within noise.
[Experiments and Insights (Tables 3-6)] All reported metrics are point estimates from a single LLM run per condition, with no multiple seeds, random restarts, or confidence intervals. Since the prompt examples are obtained by random sampling and GPT-4.1 is stochastic, observed differences such as the Travel F1 increase from 0.67 to 0.72 (Table 3) could reflect run-to-run variance rather than a true effect. Please repat the imputation several times per condition and report means with standard deviations or confidence intervals, or justify why a single run is sufficient.
[Determining Correlation Threshold] The correlation threshold is chosen from an elbow plot of the same dataset on which the method is then evaluated, and the comparison is made between thresholds on that same test set (e.g., 0, 0.15, 0.2). This data-dependent selection makes it difficult to separate the benefit of pruning from selection bias. To support a general claim, either fix the threshold a priori (e.g., a default of 0.2) or select it on a held-out validation set and apply it to the test data. At minimum, state explicitly that the elbow heuristic may overfit the current datasets.

minor comments (6)

[Throughout] There are several typos and wording issues: 'irrelevent' in the Introduction, 'atleast' in the Data Imputation Technique section, and 'the the F1' in the Experiments section.
[Table 1] The Adult Income row lists 15 features while the text states the dataset contains 14 features; please clarify whether the target column is included in the feature count.
[Abstract / Data Imputation Technique] The paper claims a token-aware method and significant input prompt reduction, but reports only feature-space reduction (Tables 7-8) and never measures actual token counts. Reporting token usage per condition would substantiate the central token-efficiency claim.
[Discussion] The statement that the LLM 'encounters both the majority and minority class samples in a proportionate manner' is not fully supported, since the group-wise sampling randomly draws a fixed number of examples per class; clarify how proportionality is guaranteed and how the group sizes are chosen.
[Appendix Table 9] The example prompt contains inconsistent spellings ('Doesnt' vs 'Doesn’t', 'No record' vs 'No Record') and a duplicated 'Set Header' instruction; please standardize the template and ensure the header matches the feature list.
[References] The reference to Kim, Kim, and Choo (2024b) is duplicated, and the second entry cites 'Advances in Neural Information Processing Systems, 33: 1877–1901' which appears to be the volume for NeurIPS 2020; please verify the correct venue, volume, and page numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the feature-pruning claim is an empirical comparison, not a fitted-prediction reduction.

full rationale

Most of the paper is an empirical comparison. The claimed result that pruning weakly correlated features maintains or improves imputation quality follows from measured XGBoost and Random Forest scores (Tables 3–6), not from an equation that defines the outcome in terms of the pruning rule. The group-wise CSV prompting method is borrowed from an external prior work (Kim et al.), not from the authors' own earlier papers, so no self-citation chain is load-bearing. The correlation threshold is selected from an elbow plot on the same datasets used for evaluation; this is data-dependent threshold selection, which can bias the comparison, but it is not a circular derivation because the reported improvements are not forced by the selection rule—the LLM outputs could have been worse. The main validity concern is that 'imputation quality' is operationalized as downstream classifier performance rather than direct accuracy against known missing values (which exist for the Travel dataset), but this is a measurement-validity/proxy issue, not a case of the conclusion being equivalent to the input by definition. No circular step satisfying the quote-and-reduction standard was found.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central comparisons rest on a small set of modeling assumptions: classifier scores on imputed rows stand in for imputation quality, the elbow-point heuristic yields a valid feature-selection threshold, classifiers trained on complete rows transfer to imputed rows, and GPT-4.1 behavior represents LLMs generally. Two design choices, the correlation threshold and the prompt group size, are set by hand or left unspecified. No invented entities are introduced.

free parameters (2)

Correlation threshold for feature selection = 0.2 (with 0.1 and 0.15 tested for comparison)
Chosen from elbow plots of correlations computed on the same datasets; the threshold controls which columns enter the prompt and is central to the token-reduction claim.
Number of completed examples per group in the prompt
The template in Table 9 shows fixed group sizes, but the exact count used in the experiments is not reported; this design choice affects all imputation outputs.

assumptions (5)

domain assumption Classification metrics on imputed rows measure imputation quality.
All conclusions about quality are drawn from XGBoost and Random Forest scores on imputed rows; no direct comparison with ground-truth missing values is reported.
ad hoc to paper The elbow point in the sorted absolute correlation plot identifies a meaningful feature-selection threshold.
Defined in the Determining Correlation Threshold section; threshold 0.2 is selected from these plots without held-out validation.
domain assumption Classifiers trained on complete rows transfer to imputed rows.
The evaluation trains on original complete records and tests on imputed records, assuming the imputed rows are drawn from the same distribution.
domain assumption GPT-4.1 imputation results represent LLM behavior in general.
All experiments use a single model, so the conclusions about prompt design are not shown to generalize across models.
standard math Pearson correlation, Cramer's V, and eta ratio are appropriate measures of association for the feature types used.
Used to rank features in the Determining Correlation Threshold section; the paper does not test alternative association measures.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Does Prompt Design Impact Quality of Data Imputation by LLMs?." pith.science (2026). https://pith.science/paper/MIGFJLZP

@misc{pith2026250604172,
  author       = {Pith},
  title        = {Pith review of: Does Prompt Design Impact Quality of Data Imputation by LLMs?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MIGFJLZP}},
  note         = {Machine review of arXiv:2506.04172}
}

read the original abstract

Generating realistic synthetic tabular data presents a critical challenge in machine learning. It adds another layer of complexity when this data contain class imbalance problems. This paper presents a novel token-aware data imputation method that leverages the in-context learning capabilities of large language models. This is achieved through the combination of a structured group-wise CSV-style prompting technique and the elimination of irrelevant contextual information in the input prompt. We test this approach with two class-imbalanced binary classification datasets and evaluate the effectiveness of imputation using classification-based evaluation metrics. The experimental results demonstrate that our approach significantly reduces the input prompt size while maintaining or improving imputation quality compared to our baseline prompt, especially for datasets that are of relatively smaller in size. The contributions of this presented work is two-fold -- 1) it sheds light on the importance of prompt design when leveraging LLMs for synthetic data generation and 2) it addresses a critical gap in LLM-based data imputation for class-imbalanced datasets with missing data by providing a practical solution within computational constraints. We hope that our work will foster further research and discussions about leveraging the incredible potential of LLMs and prompt engineering techniques for synthetic data generation.

Figures

Figures reproduced from arXiv: 2506.04172 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 17 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Benjelloun, O.; Chen, S.; and Noy, N. 2020. Google dataset search by the numbers. In International semantic web conference, 667--682. Springer

work page 2020
[4]

Borisov, V.; Seßler, K.; Leemann, T.; Pawelczyk, M.; and Kasneci, G. 2022. Language Models are Realistic Tabular Data Generators. arXiv preprint arXiv:2210.06280

arXiv 2022
[5]

D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33: 1877--1901

work page 2020
[6]

V.; Bowyer, K

Chawla, N. V.; Bowyer, K. W.; Hall, L. O.; and Kegelmeyer, W. P. 2002. SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16: 321--357

work page 2002
[7]

Eigenschink, P.; Reutterer, T.; Vamosi, S.; Vamosi, R.; Sun, C.; and Kalcher, K. 2023. Deep generative models for synthetic data: A survey. IEEE Access, 11: 47304--47320

work page 2023
[8]

Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems, volume 27

work page 2014

Show all 29 references

[9]

Google . 2025. Google Dataset Search . https://datasetsearch.research.google.com/. Accessed: 2025-05-22

2025
[10]

Guo, J.; Li, J.; Li, D.; Tiong, A. M. H.; Li, B.; Tao, D.; and Hoi, S. 2023. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10867--10877

2023
[11]

Isaell. 2023. Travel Customer Churn Analysis Prediction [Dataset]. Kaggle. Accessed: 2025-04-20

2023
[12]

Jiang, M.; Huang, T.; Guo, B.; Lu, Y.; and Zhang, F. 2024 a . Enhancing robustness in large language models: Prompting for mitigating the impact of irrelevant information. arXiv preprint arXiv:2408.10615

2024 arXiv
[13]

B.; Girard, P.; and Terranova, N

Jiang, Y.; Garc \' a-Dur \'a n, A.; Losada, I. B.; Girard, P.; and Terranova, N. 2024 b . Generative models for synthetic data generation: application to pharmacokinetic/pharmacodynamic data. Journal of pharmacokinetics and pharmacodynamics, 1--9

2024
[14]

Jin, H.; Chen, P.; Yang, J.; Wang, Z.; Jiang, M.; Gao, Y.; Huang, B.; Zhang, X.; Li, Z.; Liu, T.; et al. 2025. END: Early Noise Dropping for Efficient and Effective Context Denoising. arXiv preprint arXiv:2502.18915

2025 arXiv
[15]

Kim, J.; Kim, T.; and Choo, J. 2024 a . EPIC: Effective Prompting for Imbalanced-Class Data Synthesis in Tabular Data Classification via Large Language Models. Advances in Neural Information Processing Systems, 37: 31504--31542

2024
[16]

Kim, J.; Kim, T.; and Choo, J. 2024 b . Group-wise Prompting for Synthetic Tabular Data Generation using Large Language Models. Advances in Neural Information Processing Systems, 33: 1877--1901

2024
[17]

P.; and Welling, M

Kingma, D. P.; and Welling, M. 2013. Auto-Encoding Variational Bayes. arXiv preprint arXiv:1312.6114

2013 arXiv
[18]

S.; Reid, M.; Matsuo, Y.; and Iwasawa, Y

Kojima, T.; Gu, S. S.; Reid, M.; Matsuo, Y.; and Iwasawa, Y. 2022. Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems, 35: 22199--22213

2022
[19]

Li, Y.; Bonatti, R.; Abdali, S.; Wagle, J.; and Koishida, K. 2024. Data generation using large language models for text classification: An empirical case study. arXiv preprint arXiv:2407.12813

2024 arXiv
[20]

Liu, R.; Wei, J.; Liu, F.; Si, C.; Zhang, Y.; Rao, J.; Zheng, S.; Peng, D.; Yang, D.; Zhou, D.; et al. 2024. Best practices and lessons learned on synthetic data. arXiv preprint arXiv:2404.07503

2024 arXiv
[21]

Long, L.; Wang, R.; Xiao, R.; Zhao, J.; Ding, X.; Chen, G.; and Wang, H. 2024. On llms-driven synthetic data generation, curation, and evaluation: A survey. arXiv preprint arXiv:2406.15126

2024 arXiv
[22]

Nadas, M.; Diosan, L.; and Tomescu, A. 2025. Synthetic Data Generation Using Large Language Models: Advances in Text and Code. arXiv preprint arXiv:2503.14023

2025
[23]

Park, N.; Mohammadi, M.; Gorde, K.; Jajodia, S.; Park, H.; and Kim, Y. 2018. Data synthesis based on generative adversarial networks. arXiv preprint arXiv:1806.03384

2018 arXiv
[24]

H.; Sch \"a rli, N.; and Zhou, D

Shi, F.; Chen, X.; Misra, K.; Scales, N.; Dohan, D.; Chi, E. H.; Sch \"a rli, N.; and Zhou, D. 2023. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, 31210--31227. PMLR

2023
[25]

V.; Zhou, D.; et al

Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35: 24824--24837

2022
[26]

Wenruliu; and Becker, B. 2024. Adult Income Dataset [Dataset]. Kaggle. Accessed: 2024-04-10

2024
[27]

J.; Krishna, R.; Shen, J.; and Zhang, C

Yu, Y.; Zhuang, Y.; Zhang, J.; Meng, Y.; Ratner, A. J.; Krishna, R.; Shen, J.; and Zhang, C. 2023 a . Large language model as attributed training data generator: A tale of diversity and bias. Advances in Neural Information Processing Systems, 36: 55734--55784

2023
[28]

Yu, Y.; Zhuang, Y.; Zhang, R.; Meng, Y.; Shen, J.; and Zhang, C. 2023 b . Regen: Zero-shot text classification via training data generation with progressive dense retrieval. arXiv preprint arXiv:2305.10703

2023 arXiv
[29]

I.; Han, Z.; Paster, K.; Pitis, S.; Chan, H.-C.; and Ba, J

Zhou, Y.; Muresanu, A. I.; Han, Z.; Paster, K.; Pitis, S.; Chan, H.-C.; and Ba, J. 2022. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910

2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Benjelloun, O.; Chen, S.; and Noy, N. 2020. Google dataset search by the numbers. In International semantic web conference, 667--682. Springer

work page 2020

[4] [4]

Borisov, V.; Seßler, K.; Leemann, T.; Pawelczyk, M.; and Kasneci, G. 2022. Language Models are Realistic Tabular Data Generators. arXiv preprint arXiv:2210.06280

arXiv 2022

[5] [5]

D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33: 1877--1901

work page 2020

[6] [6]

V.; Bowyer, K

Chawla, N. V.; Bowyer, K. W.; Hall, L. O.; and Kegelmeyer, W. P. 2002. SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16: 321--357

work page 2002

[7] [7]

Eigenschink, P.; Reutterer, T.; Vamosi, S.; Vamosi, R.; Sun, C.; and Kalcher, K. 2023. Deep generative models for synthetic data: A survey. IEEE Access, 11: 47304--47320

work page 2023

[8] [8]

Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. 2014. Generative Adversarial Nets. In Advances in Neural Information Processing Systems, volume 27

work page 2014