pith. machine review for the scientific record. sign in

arxiv: 2603.01444 · v2 · submitted 2026-03-02 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Autoregressive Synthesis of Sparse and Semi-Structured Mixed-Type Data

Thomas R\"uckstie{\ss} , Robin Vujanic
Authors on Pith no claims yet

Pith reviewed 2026-05-15 18:10 UTC · model grok-4.3

classification 💻 cs.LG
keywords synthetic data generationsemi-structured dataJSON synthesisautoregressive transformerprivacy-preserving datamixed-type datadata flatteninggrammar constraints
0
0 comments X

The pith

ORiGAMi synthesizes sparse semi-structured JSON data directly with an autoregressive transformer instead of flattening records into tables.

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

The paper presents ORiGAMi as a method to generate synthetic versions of mixed-type data that arrive as nested objects, arrays, and optional fields in JSON format. It serializes each record into a sequence of tokens that include keys, values, and structural markers, then uses path-based position encodings to keep track of nesting without turning the data into a wide flat table. Grammar and schema constraints guide the autoregressive generation so that outputs remain valid JSON and match the original dataset's structure. Evaluations across six datasets show this direct modeling beats multiple baselines that first flatten the data, winning on 17 of 18 fidelity, detection, and utility metrics while keeping privacy scores above 96 percent. A reader would care because real data systems increasingly store information in sparse, hierarchical forms rather than fixed-schema tables, so avoiding flattening artifacts could improve the usefulness of synthetic data for privacy sharing and testing.

Core claim

ORiGAMi is an autoregressive transformer architecture for modeling and synthesizing semi-structured records without flattening. It serializes JSON records into key, value, and structural tokens, encodes token positions by their path in the document tree, and applies grammar and schema constraints to enforce syntactically valid JSON and dataset-consistent structure. Across six datasets ranging from dense tabular benchmarks to large-scale semi-structured collections, ORiGAMi achieves the best score in 17 of 18 benchmark comparisons against VAE, GAN, diffusion, and autoregressive baselines that operate on flattened representations, while maintaining high privacy scores above 96 percent across所有

What carries the argument

An autoregressive transformer that processes sequences of key-value-structural tokens whose positions are encoded by their path through the JSON document tree, guided by grammar and schema constraints to produce valid nested records.

If this is right

  • Synthetic data can preserve nested objects, variable-length arrays, and optional keys without introducing flattening artifacts.
  • Performance gains appear on fidelity, detection, and utility metrics for both dense tabular and large semi-structured collections.
  • Privacy scores remain above 96 percent while fidelity improves, supporting privacy-preserving data sharing.
  • Native record modeling becomes a viable alternative to tabular synthesis pipelines for modern data systems.

Where Pith is reading between the lines

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

  • The same tokenization and path-encoding approach could be tested on other hierarchical formats such as XML or protocol buffers without major redesign.
  • Models trained on ORiGAMi outputs might show improved accuracy on queries that traverse nested structures compared with flattened alternatives.
  • For extremely deep or wide schemas the fixed token vocabulary and path encoding may require scaling adjustments to stay efficient.
  • The grammar-constraint mechanism could be combined with existing autoregressive code-generation techniques for structured output tasks.

Load-bearing premise

Serializing JSON into key-value-structural tokens with path positions plus grammar constraints is sufficient to capture all semantically important relationships in the original semi-structured records without additional domain-specific modeling.

What would settle it

A downstream task whose performance drops when trained on ORiGAMi synthetic data but not on data from a flattening baseline, specifically because certain nested relationships or array-length dependencies are missing.

Figures

Figures reproduced from arXiv: 2603.01444 by Robin Vujanic, Thomas R\"uckstie{\ss}.

Figure 1
Figure 1. Figure 1: Tokenization of an example movies record. and value tokens and maintain the structure of arrays and nested objects with special grammatical tokens. For a dataset D, each JSON record𝑑 ∈ D is serialized into a token sequence x = (𝑥1, . . . , 𝑥𝑇 ) by a depth-first traversal. The vocabulary V consists of three disjoint token classes, V = V𝑠 ∪ V𝑘 ∪ V𝑣 as follows: • Structural tokens V𝑠 = {start, end, obj_start,… view at source ↗
Figure 2
Figure 2. Figure 2: origami dual-head model architecture with gram￾mar and schema constraints imposed on the discrete head. The continuous loss is the negative log-likelihood over positions where the target is a num token. The total training loss is: L = LCE + 𝜆 LNLL, (8) where 𝜆 is set proportionally to the fraction of num tokens to total tokens in each batch. 3.6 Key-Order Shuffling Since JSON object keys are unordered by s… view at source ↗
Figure 3
Figure 3. Figure 3: Flattening and type separation of two movie records. Nested objects and arrays are mapped to dot-separated columns; [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: KDE visualizations of sparse numeric columns on [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Wasserstein distance of length distributions be [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: KVPE vs. sequential position encoding on a syn [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
read the original abstract

Synthetic data generation is an important capability for privacy-preserving data sharing, system benchmarking and test data provisioning. For mixed-type data, existing synthesizers largely target dense, fixed-schema tables, but many modern data systems store and exchange sparse, semi-structured JSON with nested objects, variable-length arrays and optional keys. Applying tabular synthesizers to such data requires flattening records into wide, sparse tables, turning nested structure and arrays into column-layout artifacts. We present ORiGAMi, an autoregressive transformer architecture for modeling and synthesizing semi-structured records without flattening. ORiGAMi serializes JSON records into key, value, and structural tokens, and encodes token positions by their path in the document tree. Grammar and schema constraints enforce syntactically valid JSON and dataset-consistent structure. We evaluate ORiGAMi against VAE, GAN, diffusion, and autoregressive baselines that operate on flattened representations across six datasets ranging from dense tabular benchmarks to large-scale semi-structured collections. Across fidelity, detection, and utility metrics, ORiGAMi achieves the best score in 17 of 18 benchmark comparisons, while maintaining high privacy scores above 96% across all settings. These results establish native record modeling as a strong alternative to tabular synthesis pipelines, preserving structure while achieving state-of-the-art benchmark performance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper introduces ORiGAMi, an autoregressive transformer for synthesizing sparse semi-structured mixed-type data in JSON format. Records are serialized into key-value-structural tokens with path-based position encodings in the document tree; grammar and schema constraints enforce syntactic validity and dataset consistency. Evaluated against VAE, GAN, diffusion, and autoregressive baselines on flattened representations across six datasets, ORiGAMi reports the best score in 17 of 18 comparisons on fidelity, detection, and utility metrics while maintaining privacy scores above 96%.

Significance. If the empirical superiority holds under proper statistical controls, the work establishes native path-encoded autoregressive modeling as a viable and potentially superior alternative to flattening-based tabular synthesizers for nested and sparse JSON data, with direct relevance to privacy-preserving data sharing in modern systems.

major comments (2)
  1. [§5 (Experimental Evaluation)] §5 (Experimental Evaluation): the central claim of best performance in 17 of 18 benchmark comparisons reports only point estimates with no standard deviations across seeds, no p-values, and no mention of multiple independent runs. Autoregressive transformers on variable-length sequences are known to exhibit high training variance; without these controls the reliability of the reported margins cannot be assessed.
  2. [§4 (Baselines and Adaptation)] §4 (Baselines and Adaptation): the manuscript provides insufficient detail on how the tabular baselines (VAE, GAN, diffusion, autoregressive) were adapted to semi-structured inputs after flattening, including any specific preprocessing for nested objects, variable-length arrays, and optional keys.
minor comments (2)
  1. [§3 (Model Architecture)] The path-based position encoding is described at a high level but would benefit from an explicit equation or pseudocode definition to clarify how tree paths are mapped to token positions.
  2. A summary table listing the six datasets with key characteristics (size, sparsity, nesting depth) would improve readability of the experimental setup.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below and will revise the manuscript accordingly to improve the statistical rigor of the evaluation and the clarity of the baseline adaptations.

read point-by-point responses

  1. Referee: §5 (Experimental Evaluation): the central claim of best performance in 17 of 18 benchmark comparisons reports only point estimates with no standard deviations across seeds, no p-values, and no mention of multiple independent runs. Autoregressive transformers on variable-length sequences are known to exhibit high training variance; without these controls the reliability of the reported margins cannot be assessed.

    Authors: We agree that point estimates alone are insufficient to establish the reliability of the reported performance margins, particularly for autoregressive models that can exhibit training variance. In the revised manuscript we will rerun all experiments across at least five independent random seeds, report mean and standard deviation for every metric, and include appropriate statistical tests (paired t-tests or Wilcoxon signed-rank tests with Bonferroni correction) to assess whether the observed differences are significant. These additions will be placed in §5 and the corresponding tables/figures will be updated. revision: yes

  2. Referee: §4 (Baselines and Adaptation): the manuscript provides insufficient detail on how the tabular baselines (VAE, GAN, diffusion, autoregressive) were adapted to semi-structured inputs after flattening, including any specific preprocessing for nested objects, variable-length arrays, and optional keys.

    Authors: We acknowledge that the current description of baseline adaptation is too brief. In the revised §4 we will add a dedicated subsection that explicitly describes the flattening procedure: nested objects are flattened using dot-path column names, variable-length arrays are expanded into multiple columns with a fixed maximum length (with padding and a length indicator column), and optional keys are represented as nullable columns with an explicit missing-value indicator. We will also document any additional preprocessing steps (e.g., type casting, normalization) applied uniformly to all methods to ensure a fair comparison. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmarks on held-out data with independent baselines

full rationale

The paper introduces ORiGAMi as an autoregressive transformer that serializes JSON into path-encoded tokens with grammar constraints, then reports direct empirical comparisons on six datasets against VAE/GAN/diffusion/autoregressive baselines. No equations, fitted parameters, or self-citations are invoked to derive the 17/18 best-score claim; metrics are computed on held-out test records. The derivation chain consists of model definition followed by standard train/test evaluation, with no reduction of outputs to inputs by construction. This matches the reader's assessment of non-circular empirical evaluation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that path-encoded token sequences plus grammar constraints are expressive enough for the target data distributions; no new physical entities are postulated.

free parameters (1)
  • transformer hyperparameters
    Standard model size, learning rate, and sampling temperature choices that are fitted or tuned during training.
axioms (1)
  • domain assumption JSON records can be losslessly serialized into a linear sequence of key, value, and structural tokens whose positions are fully determined by their tree path.
    Invoked in the description of the serialization and position-encoding step.
invented entities (1)
  • path-based position encoding for JSON tokens no independent evidence
    purpose: To inject document-tree structure into the autoregressive sequence model.
    New encoding mechanism introduced to avoid flattening.

pith-pipeline@v0.9.0 · 5529 in / 1346 out tokens · 35460 ms · 2026-05-15T18:10:54.098694+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 3 internal anchors

  1. [1]

    Alcorn and Anh Nguyen

    Michael A. Alcorn and Anh Nguyen. 2021. The DEformer: An Order-Agnostic Distribution Estimating Transformer. http://arxiv.org/abs/2106.06989

    work page arXiv 2021

  2. [2]

    Vadim Borisov, Kathrin Seßler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. 2023. Language Models are Realistic Tabular Data Generators. In International Conference on Learning Representations (ICLR)

    work page 2023

  3. [3]

    Reutter, Fernando Suárez, and Domagoj Vrgoč

    Pierre Bourhis, Juan L. Reutter, Fernando Suárez, and Domagoj Vrgoč. 2017. JSON: Data Model, Query Languages and Schema Specification. InProceedings of the 36th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS). 123–135. https://doi.org/10.1145/3034786.3056120

    work page doi:10.1145/3034786.3056120 2017

  4. [4]

    Kuntai Cai, Xiaokui Xiao, and Graham Cormode. 2023. PrivLava: Synthesizing Relational Data with Foreign Keys under Differential Privacy. InProceedings of the 2023 International Conference on Management of Data. ACM, 1–25

    work page 2023

  5. [5]

    Michael Carey, Wail Alkowaileet, Nick DiGeronimo, Peeyush Gupta, Sachin Smotra, and Till Westmann. 2025. Towards Principled, Practical Document Database Design.Proceedings of the VLDB Endowment18, 12 (2025), 4804–4816. https://doi.org/10.14778/3750601.3750606

    work page doi:10.14778/3750601.3750606 2025

  6. [6]

    Sonia Cromp, Satya Sai Srinath Namburi GNVV, Mohammed Alkhudhayri, Catherine Cao, Samuel Guo, Nicholas Roberts, and Frederic Sala. 2026. Tabby: A Language Model Architecture for Tabular and Structured Data Synthesis.Trans- actions on Machine Learning Research(2026). https://openreview.net/forum?id= b9FPVnb0Bn

    work page 2026

  7. [7]

    2026.Synthetic Data Metrics

    DataCebo, Inc. 2026.Synthetic Data Metrics. DataCebo, Inc. https://docs.sdv. dev/sdmetrics/ Version 0.12.0

    work page 2026

  8. [8]

    Arsene Fansi Tchango, Rishab Goel, Zhi Wen, Julien Martel, and Joumana Ghosn

    work page

  9. [9]

    InAdvances in Neural Information Processing Systems, Vol

    DDXPlus: A New Dataset for Automatic Medical Diagnosis. InAdvances in Neural Information Processing Systems, Vol. 35. 31306–31318

    work page

  10. [10]

    Philip Gage. 1994. A new algorithm for data compression.C Users J.12, 2 (Feb. 1994), 23–38

    work page 1994

  11. [11]

    Yunqing Ge et al. 2025. Privacy-Enhanced Database Synthesis for Benchmark Publishing.Proceedings of the VLDB Endowment18, 2 (2025)

    work page 2025

  12. [12]

    Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, Bruno Régaldo-Saint Blancard, Tiberiu Tesileanu, Kyunghyun Cho, and Shirley Ho. 2023. xVal: A Continuous Number Encoding for Large Language Models. http://arxiv.org/abs/2310.02989

    work page arXiv 2023

  13. [13]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial Networks. InAdvances in Neural Information Processing Systems, Vol. 27

    work page 2014

  14. [14]

    Dan Hendrycks and Kevin Gimpel. 2016. Gaussian Error Linear Units (GELUs). http://arxiv.org/abs/1606.08415

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  15. [15]

    Cohen, and Adrian Weller

    James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N. Cohen, and Adrian Weller. 2022. Synthetic Data – What, Why and How? http://arxiv.org/abs/2205.03257

    work page arXiv 2022

  16. [16]

    Markelle Kelly, Rachel Longjohn, and Kolby Nottingham. [n.d.]. The UCI Machine Learning Repository. https://archive.ics.uci.edu

    work page

  17. [17]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. 2014. Auto-Encoding Variational Bayes. InInternational Conference on Learning Representations (ICLR)

    work page 2014

  18. [18]

    Terry Koo, Frederick Liu, and Luheng He. 2024. Automata-Based Constraints for Language Model Decoding. InConference on Language Modeling. https: //openreview.net/forum?id=BDBdblmyzY

    work page 2024

  19. [19]

    Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. 2023. TabDDPM: Modelling Tabular Data with Diffusion Models. InInternational Conference on Machine Learning (ICML). 17564–17579

    work page 2023

  20. [20]

    Qinyi Liu, Mohammad Khalil, Jelena Jovanovic, and Ronas Shakya. 2024. Scaling While Privacy Preserving: A Comprehensive Synthetic Tabular Data Generation and Evaluation in Learning Analytics. InProceedings of the 14th Learning Analyt- ics and Knowledge Conference (LAK). 620–631. https://doi.org/10.1145/3636555. 3636921

    work page doi:10.1145/3636555 2024

  21. [21]

    Tennison Liu, Zhaozhi Qian, Jeroen Berrevoets, and Mihaela van der Schaar

    work page

  22. [22]

    InInternational Conference on Learning Representations (ICLR)

    GOGGLE: Generative Modelling for Tabular Data by Learning Relational Structure. InInternational Conference on Learning Representations (ICLR)

    work page

  23. [23]

    David Lopez-Paz and Maxime Oquab. 2017. Revisiting Classifier Two-Sample Tests. InInternational Conference on Learning Representations (ICLR)

    work page 2017

  24. [24]

    Shubhankar Mohapatra, Jianqiao Zong, Florian Kerschbaum, and Xi He. 2024. Differentially Private Data Generation with Missing Data.Proceedings of the VLDB Endowment17, 8 (2024), 2022–2035

    work page 2024

  25. [25]

    Phil Ostheimer, Mayank Nagda, Andriy Balinskyy, Jean Radig, Carl Herrmann, Stephan Mandt, Marius Kloft, and Sophie Fellenz. 2025. Sparse Data Diffusion for Scientific Simulations in Biology and Physics.arXiv preprint arXiv:2502.02448 (2025)

    work page arXiv 2025

  26. [26]

    Neha Patki, Roy Wedge, and Kalyan Veeramachaneni. 2016. The Synthetic Data Vault. In2016 IEEE International Conference on Data Science and Advanced Analytics (DSAA). 399–410. https://doi.org/10.1109/DSAA.2016.49

    work page doi:10.1109/dsaa.2016.49 2016

  27. [27]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour- napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python.Journal of Machine Learning Research12 (2011), 2825–2830

    work page 2011

  28. [28]

    Reutter, Fernando Suarez, Martín Ugarte, and Domagoj Vrgoč

    Felipe Pezoa, Juan L. Reutter, Fernando Suarez, Martín Ugarte, and Domagoj Vrgoč. 2016. Foundations of JSON schema. InProceedings of the 25th international conference on world wide web (WWW). 263–273

    work page 2016

  29. [29]

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever

    work page

  30. [30]

    https://cdn.openai.com/research-covers/language-unsupervised/language_ understanding_paper.pdf

    Improving Language Understanding by Generative Pre-Training. https://cdn.openai.com/research-covers/language-unsupervised/language_ understanding_paper.pdf

    work page

  31. [31]

    Tobias Schmidt, Viktor Leis, Peter Boncz, and Thomas Neumann. 2025. SQLStorm: Taking Database Benchmarking into the LLM Era.Proceedings of the VLDB Endowment18, 11 (2025), 4144–4157. https://doi.org/10.14778/3749646.3749683

    work page doi:10.14778/3749646.3749683 2025

  32. [32]

    Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec. 2025. TabDiff: a Mixed-type Diffusion Model for Tabular Data Gen- eration. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=swvURjrt8z

    work page 2025

  33. [33]

    Ruxue Shi, Yili Wang, Mengnan Du, Xu Shen, Yi Chang, and Xin Wang. 2025. A Comprehensive Survey of Synthetic Tabular Data Generation. http://arxiv.org/ abs/2504.16506

    work page arXiv 2025

  34. [34]

    Solatorio and Olivier Dupriez

    Aivin V. Solatorio and Olivier Dupriez. 2023. REaLTabFormer: Generating Real- istic Relational and Tabular Data using Transformers. http://arxiv.org/abs/2302. 02041

    work page 2023

  35. [35]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A Simple Way to Prevent Neural Networks from Overfitting.Journal of Machine Learning Research15, 56 (2014), 1929–1958

    work page 2014

  36. [36]

    Michael Stonebraker and Andrew Pavlo. 2024. What Goes Around Comes Around... And Around...ACM Sigmod Record53, 2 (2024), 21–37

    work page 2024

  37. [37]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu

    work page

  38. [38]

    https://doi.org/10.1016/j.neucom.2023.127063

    RoFormer: Enhanced Transformer with Rotary Position Embedding.Neu- rocomputing568 (2024), 127063. https://doi.org/10.1016/j.neucom.2023.127063

    work page doi:10.1016/j.neucom.2023.127063 2024

  39. [39]

    Paul Tiwald, Ivona Krchova, Andrey Sidorenko, Mariana Vargas Vieyra, Mario Scriminaci, and Michael Platzer. 2025. TabularARGN: A Flexible and Efficient Auto-Regressive Framework for Generating High-Fidelity Synthetic Data. http: //arxiv.org/abs/2501.12012

    work page arXiv 2025

  40. [40]

    Benigno Uria, Iain Murray, and Hugo Larochelle. 2014. A Deep and Tractable Density Estimator. InInternational Conference on Machine Learning (ICML). 467– 475

    work page 2014

  41. [41]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All You Need. InAdvances in Neural Information Processing Systems, Vol. 30. 5998– 6008

    work page 2017

  42. [42]

    2009.Optimal transport: old and new

    Cédric Villani et al. 2009.Optimal transport: old and new. Vol. 338. Springer

    work page 2009

  43. [43]

    Efficient Guided Generation for Large Language Models

    Brandon T. Willard and Rémi Louf. 2023. Efficient Guided Generation for Large Language Models. http://arxiv.org/abs/2307.09702 arXiv:2307.09702 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason R...

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  45. [45]

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni

    work page

  46. [46]

    InAdvances in Neural Information Processing Systems, Vol

    Modeling Tabular Data using Conditional GAN. InAdvances in Neural Information Processing Systems, Vol. 32

    work page

  47. [47]

    Jingyi Yang, Peizhi Wu, Gao Cong, Tong Yang, and Jianfei Ruan. 2022. SAM: Database Generation from Query Workloads with Supervised Autoregressive Models. InProceedings of the 2022 International Conference on Management of Data. ACM, 1542–1555

    work page 2022

  48. [48]

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. XLNet: Generalized Autoregressive Pretraining for Lan- guage Understanding. InAdvances in Neural Information Processing Systems, Vol. 32

    work page 2019

  49. [49]

    Hellerstein, Sanjay Krishnan, and Ion Stoica

    Zongheng Yang, Eric Liang, Amog Kamsetty, Chenggang Wu, Yan Duan, Xi Chen, Pieter Abbeel, Joseph M. Hellerstein, Sanjay Krishnan, and Ion Stoica. 2019. Deep Unsupervised Cardinality Estimation.Proceedings of the VLDB Endowment13, 3 (2019), 279–292. https://doi.org/10.14778/3368289.3368294

    work page doi:10.14778/3368289.3368294 2019

  50. [50]

    Yefeng Yuan, Yuhong Liu, and Liang Cheng. 2025. A Multi-Faceted Evaluation Framework for Assessing Synthetic Data Generated by Large Language Models. http://arxiv.org/abs/2404.14445

    work page arXiv 2025

  51. [51]

    Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2024. Mixed- Type Tabular Data Synthesis with Score-based Diffusion in Latent Space. In International Conference on Learning Representations (ICLR). 13

    work page 2024