Pith. sign in

REVIEW 4 major objections 6 minor 22 references

Seeding API-based private text generation with DP tabular metadata plus in-context examples removes the initialization bottleneck, giving strictly better privacy-utility trade-offs, faster convergence, and far lower API cost than prior Priv

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 ↗

MAPLE boosts Private Evolution's initial prompt with DP tabular metadata and donated in-context examples, improving synthetic-text utility, convergence, and API cost on specialized corpora.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection MAPLE is a promising initialization fix for Private Evolution, but the headline privacy-utility trade-off was measured with an un-private metadata extractor, so the central claim is not yet supported as stated. the 4 major comments →

arxiv 2603.19258 v2 pith:DIDJKWMJ submitted 2026-02-26 cs.CL cs.AIcs.CRcs.LG

MAPLE: Metadata Augmented Private Language Evolution

Eli Chien , Yuzheng Hu , Ryan McKenna , Shanshan Wu , Zheng Xu , Peter Kairouz This is my paper
classification cs.CL cs.AIcs.CRcs.LG
keywords differential privacysynthetic text generationPrivate Evolutionmetadata augmentationin-context learningtabular DP synthesisAPI-only LLMsspecialized domain text
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 reading

The paper claims that the main obstacle to API-based differentially private synthetic text generation is a bad starting point: when private text is far from an LLM's pretraining distribution, the Private Evolution algorithm wastes many rounds and privacy budget just to reach the right region. MAPLE supplies a data-dependent starting point by extracting a tabular metadata schema from raw private text, synthesizing that table under differential privacy with a CPU-only tabular method, and then prompting the foundation model to generate initial synthetic texts conditioned on the DP metadata plus a handful of donated (metadata, text) examples. The paper reports that this initialization makes MAPLE consistently beat the prior Private Evolution baseline on every dataset, metric, and privacy budget tested, converge in two refinement rounds where the baseline needs eight or nine, and use far fewer API calls—while also outperforming a DP-finetuned generator. If right, it means high-utility DP synthetic text for specialized domains is achievable with only black-box API access and no gradient-level DP training.

Core claim

On the paper's own terms, the discovery is that metadata conditions the random-draw phase of Private Evolution without breaking end-to-end DP. A lightweight LLM annotator converts each private text into a row of structured attributes; a CPU-only DP tabular synthesizer (AIM) turns those rows into a DP synthetic table at a small fraction of the total privacy budget; and the prompt for the initial random draw then contains a row sampled from that table plus ten donated (metadata, text) pairs chosen by Hamming-distance similarity. This grounds the initial synthetic distribution in the target domain, so the subsequent PE voting-and-paraphrase loop starts close and converges in one or two iteratio

What carries the argument

The load-bearing mechanism is the metadata-grounded RANDOM_API prompt. It has three parts: a DP synthetic row over a rich categorical schema (e.g., research area, experimental approach, sample size, tone); a set of up to ten donated (metadata, text) pairs, selected per-row by smallest Hamming distance, serving as in-context demonstrations; and the foundation model's in-context learning ability, which converts a synthetic row plus examples into a full text sample. This initial sample set then feeds the unchanged AugPE refinement loop. The two-part prompt is what carries the argument: the ablation shows each part alone is marginal, but the combination re-anchors the distribution and cuts refin

Load-bearing premise

The end-to-end DP guarantee assumes the metadata annotator reads the raw private text on a trusted local host; if that annotation step is outsourced to an external API, the private text is exposed before any DP noise is added.

What would settle it

Run MAPLE with the DP metadata rows randomly permuted across attributes while keeping the in-context examples unchanged; if MAUVE and downstream accuracy stay at MAPLE levels, the gains are not from metadata content. Run it again with a weak keyword-based annotator that must be hosted locally; if quality collapses to baseline, the local-annotator assumption is the binding constraint.

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

If this is right

  • Specialized-domain DP synthetic text is practical with only black-box API access and a CPU-only DP tabular synthesizer, removing the need for DP fine-tuning or gradient-level APIs.
  • Because the metadata takes a small slice of the privacy budget (a 1/9 split in the experiments) yet drives most of the gain, adding a structured prior is not a net privacy cost.
  • Fewer refinement rounds mean less composed DP noise and proportionally lower API spend, so the same nominal epsilon buys more utility.
  • Because the final synthetic text is a post-processing of a DP mechanism, it can be reused across downstream tasks without additional privacy cost.
  • Richer metadata schemas converge faster and peak higher, so schema design is a direct control knob for practitioners.

Where Pith is reading between the lines

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

  • A direct testable prediction follows from the mechanism: MAPLE's advantage over the vanilla PE baseline should grow as the private domain moves further from the LLM's pretraining priors; testing on an even more specialized corpus (e.g., clinical notes or legal text) would check that.
  • The end-to-end guarantee stands on the assumption that the metadata annotator runs locally and is trusted; the paper's private-inference fallback is untested, and the two-attribute ablation suggests weaker annotation narrows but likely does not erase the gain.
  • Because the fix is purely in the initialization, it should stack with any improvements to the refinement loop itself (better paraphrase operators, multi-model fusion), yielding compounding gains.
  • The same metadata-anchored initialization should transfer to other modalities—images described by scene attributes or tabular data by column summaries—since the mechanism only needs a DP-synthesizable structured description of the private samples.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper proposes MAPLE, a modification of the Private Evolution (PE) framework for differentially private synthetic text generation. MAPLE extracts tabular metadata from raw private text (Step 0), trains a DP tabular generator (AIM) on that metadata, and uses the DP synthetic metadata plus a small set of donated (metadata, text) pairs as in-context examples in the RANDOM_API prompt, before running AugPE refinement. Experiments on bioRxiv and OpenReview are reported, with claims that MAPLE outperforms AugPE and DPFT-Gen across metrics and privacy budgets, converges in fewer iterations, and reduces API calls.

Significance. If the claims hold, the contribution is practically important: it would enable high-utility DP synthetic text for specialized domains using only API access and no DP fine-tuning, with faster convergence and lower cost. The paper addresses a real initialization bottleneck in PE, and the ablation isolating metadata vs. in-context examples is informative. The re-implementation of AugPE with vLLM is a useful methodological contribution. However, the headline privacy-utility claim is not currently supported because the experimental pipeline uses an external LLM API (gemini-2.5-flash-lite) on raw private text before any DP noise is added, while the fallback assumption of a local trusted extractor is not evaluated. The strong empirical claim also lacks variance estimates. If the privacy gap is closed and uncertainty is quantified, the result would be a meaningful advance; as it stands, the central claim is not yet established.

major comments (4)
  1. [Appendix A.1, Section 5 (Implementation details)] The end-to-end DP guarantee is not realized by the experiments. Step 0 sends raw private text to gemini-2.5-flash-lite (an external API) to extract metadata, and this occurs before any DP noise is added. Appendix A.1 explicitly assumes 'the data holder can locally host a metadata extractor that does not lead to a privacy breach,' but no such local extractor is used; the reported results are for the non-private configuration. The fallback options (private inference or manual annotation) are not evaluated. Consequently, the utility numbers in Figure 3 and Section 5.1 correspond to a pipeline that is not DP, while the utility of the private configuration is unknown. This is load-bearing for the abstract's claim of 'end-to-end differential privacy guarantees' and a 'strictly better privacy-utility trade-off.'
  2. [Section 5.1, Figure 3] The claim that 'MAPLE consistently outperforms AugPE across all datasets, metrics, and privacy constraints' is not supported with uncertainty quantification. No error bars, repeated seeds, or statistical tests are reported for the main results. The comparison also rests on only two datasets, which is thin for a 'strictly better' trade-off claim. This is load-bearing because the headline contribution is an empirical privacy-utility improvement.
  3. [Section 5 (Implementation details), Section 3 Step 2] The 'donated' in-context examples are described in Section 5 as 'sampled from the validation set.' If the validation set is derived from the same private corpus, then these examples are private data used without DP protection, contradicting the paper's description of them as non-private donated data (Section 3, Step 2). This is a load-bearing clarity issue: in-context examples are a core component of MAPLE, and their privacy status must be unambiguous. If they are private, their use must be accounted for in the DP guarantee; if they are public, the source needs to be stated explicitly.
  4. [Section 5.2, Figure 5] The 'converges faster' claim is supported only by a single dataset (bioRxiv) and the privacy budget is not labeled on the figure (the axis reads only 'ε =' and appears truncated). The 'weak M' ablation varies schema richness but does not address the effect of extractor privacy or quality, so it cannot support the Appendix A.1 statement that weaker metadata from a private extractor would still improve over vanilla AugPE. This is a limitation for extrapolating the convergence and cost conclusions to the private configuration.
minor comments (6)
  1. [Abstract and Introduction] The abstract and introduction state that experiments use 'three domain-specific datasets,' but Section 5 describes only two (bioRxiv and OpenReview). Please reconcile.
  2. [Figure 5] The epsilon value in the title is incomplete ('( = )'). Specify the privacy budget used for this convergence study and clarify that the convergence claim is dataset-specific.
  3. [Section 5 (Evaluation metrics)] The 'average JSD on metadata' metric uses the same LLM-annotated metadata schema that MAPLE conditions on, which introduces a mild circularity. Please discuss this and, if possible, include an independent metric.
  4. [Section 5.1, Figure 3] The comparison with DPFT-Gen uses GPT2-small (bioRxiv) and GPT2-large (OpenReview), while MAPLE uses Qwen2.5-7B-Instruct. The model-size difference is a confound; it is acknowledged in the OpenReview caption but should be stated in the text.
  5. [Appendix B.1] Typo: 'quntile' should be 'quantile' in the bioRxiv dataset description.
  6. [General] The paper does not mention code availability. If the authors intend to release code, a link would improve reproducibility.

Circularity Check

1 steps flagged

Mild evaluative circularity in the metadata-JSD metric; central MAUVE/downstream comparisons are independent.

specific steps
  1. self definitional [Section 5, 'Evaluation metrics' and 'Implementation details'; Figures 3-4]
    "We also examine how the distribution of the annotated metadata from synthetic text aligns with the private one via the Jensen-Shannon distance between them. ... For the metadata extraction, we utilize gemini-2.5-flash-lite as the metadata annotator."

    MAPLE's RANDOM_API prompt is constructed from DP synthetic metadata generated under the same tabular schema and by the same LLM annotator (gemini-2.5-flash-lite) that is then used to annotate both synthetic and private texts for the reported 'average JSD on metadata'. The metric therefore measures how well the generator follows its own conditioning signal rather than an independent property of the synthetic corpus. Low JSD is in part forced by construction: the prompt already contains the target metadata distribution, so re-annotating the generated text tends to reproduce it. This makes the metadata-alignment result favor MAPLE by design. The central utility claims, however, also rest on MAUVE and downstream NTP/classification accuracy, which are external to the metadata signal, so this ci

full rationale

The paper's central contribution is an empirical method, not a formal derivation, and its headline utility claims are measured with external benchmarks: MAUVE (embedding-based distribution similarity), next-token prediction accuracy on a downstream GPT-2 model, and OpenReview classification accuracy with RoBERTa. These metrics do not reduce to MAPLE's inputs. The one mild circularity is the 'average JSD on metadata' metric, which uses the same metadata schema and the same gemini-2.5-flash-lite annotator that MAPLE conditions on, so it partly rewards prompt-following by construction. This is non-central because MAUVE and downstream tasks carry the main argument. The Appendix A.1 assumption about locally hosting the metadata extractor is a real privacy-correctness caveat—the experiments used the Gemini API on raw private text, so the end-to-end privacy guarantee is not demonstrated in the measured configuration—but this is a validity concern, not a circularity. Self-citations to Hu et al. (2025) for the metadata schema are transparent prior art and are not used to forbid alternatives or to substitute for empirical evidence. Overall, no load-bearing derivation reduces to its own input; the score reflects only the minor evaluative circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No invented physical or formal entities; MAPLE is a software framework that combines existing components. The main uncharged premises are the local metadata extractor, the availability of donated public examples, and the hand-designed metadata schema.

free parameters (4)
  • epsilon_1/epsilon_2 budget split = 1/9
    Fixed split of total privacy budget between AIM metadata generation and AugPE; no sensitivity analysis is reported, so results could shift under other splits.
  • number of in-context examples (k) = 10
    Chosen after observing that more than 10 gives 'diminished return'; this is tuned on the evaluation data and affects the prompt design.
  • donated examples pool size = 50
    Chosen as a practical setting (<3% of the private set); the in-context examples are sampled from the validation set, which influences generation quality.
  • metadata schema attributes = 9 attributes for bioRxiv; 10+ for OpenReview
    The hand-designed schema determines what DP information is available to the LLM; Figure 5 shows richer metadata converges faster, so the schema is a consequential design choice.
axioms (4)
  • domain assumption Metadata extraction on raw private text can be performed locally/trusted without privacy breach
    Stated in Appendix A.1; if violated, raw private text leaves the data holder before DP noise is added, breaking the privacy guarantee.
  • domain assumption A small set of non-private donated (metadata, text) examples is available
    MAPLE requires 50 donated examples; the paper cites prior work on data donation, but such data is not universally available in private or adversarial settings.
  • standard math DP composition and post-processing (zCDP, sequential composition)
    Used to add the privacy budgets of AIM and AugPE, and to reuse synthetic data post-hoc without extra privacy cost.
  • domain assumption LLM in-context learning and paraphrasing behave as prior work assumes
    The entire PE loop depends on VARIATION_API preserving similarity and RANDOM_API following the metadata-conditioned prompt; MAPLE changes only the initialization, not these underlying behaviors.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MAPLE: Metadata Augmented Private Language Evolution." pith.science (2026). https://pith.science/paper/DIDJKWMJ

@misc{pith2026260319258,
  author       = {Pith},
  title        = {Pith review of: MAPLE: Metadata Augmented Private Language Evolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIDJKWMJ}},
  note         = {Machine review of arXiv:2603.19258}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Differentially private (DP) fine-tuning of large language models (LLMs) requires massive compute and full model access, which rules out state-of-the-art proprietary APIs for general users. Generating DP synthetic data offers a practical workaround. This approach also allows for transparent exploratory data analysis and arbitrary reuse across downstream tasks, sidestepping the rigid constraints of a model's parameter space. Private Evolution (PE) provides a promising API-based framework for generating this data, but its success relies heavily on initialization. If the private data distribution falls too far outside the foundation model's pre-training priors -- a common issue in highly specialized domain -- PE struggles to align with the target data. This misalignment causes poor convergence, degraded utility, and wasted API calls. To solve this initialization bottleneck, we introduce Metadata Augmented Private Language Evolution (MAPLE). MAPLE extracts DP tabular metadata and uses in-context learning to firmly ground the initial synthetic distribution in the target domain. Our evaluations on domain-specific text generation tasks show that MAPLE yields a strictly better privacy-utility trade-off, converges significantly faster, and sharply reduces API costs compared to baseline PE methods.

Figures

Figures reproduced from arXiv: 2603.19258 by Eli Chien, Peter Kairouz, Ryan McKenna, Shanshan Wu, Yuzheng Hu, Zheng Xu.

Figure 1
Figure 1. Figure 1: Illustration of AugPE and its limitations. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MAPLE. We first extract metadata in tabular format, either based on a designed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: repeatedly paraphrasing a review about “pizza” is unlikely to produce content related to [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Main results. Top: Biorxiv datasets. The utility metrics are MAUVE score, average JSD [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on Biorxiv dataset. +M: leveraging only metadata in the prompt. +E: [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: How metadata richness affects PE conver [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

22 extracted references · 9 linked inside Pith

  1. [1]

    B., Mironov, I., Talwar, K., and Zhang, L

    Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., and Zhang, L. Deep learning with differential privacy. InProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pp. 308–318,

    2016

  2. [3]

    In practice, if this assumption does not hold, the alternative solution is to leverage privacy-preserving inference methods (Duan et al., 2023; Hong et al.,

    Note that we assume the data holder can locally host a metadata extractor that does not lead to a privacy breach in our experiment. In practice, if this assumption does not hold, the alternative solution is to leverage privacy-preserving inference methods (Duan et al., 2023; Hong et al.,

    2023

  3. [7]

    Liu, Y ., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V

    URLhttps://openreview.net/forum?id=YEhQs8POIo. Liu, Y ., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V . Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692,

    Pith/arXiv arXiv 1907

  4. [10]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al

    See also: GPT-5 System Card (August 2025). Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744,

    2025

  5. [11]

    B., Kairouz, P., Rosenblatt, L., Cohen-Addad, V ., Guzmán, C., McKenna, R., Andrew, G., Bie, A., et al

    Ponomareva, N., Xu, Z., McMahan, H. B., Kairouz, P., Rosenblatt, L., Cohen-Addad, V ., Guzmán, C., McKenna, R., Andrew, G., Bie, A., et al. How to dp-fy your data: A practical guide to generating synthetic data with differential privacy.arXiv preprint arXiv:2512.03238,

    Pith/arXiv arXiv

  6. [13]

    Rumshisky, A., Ghassemi, M., Naumann, T., Szolovits, P., Castro, V ., McCoy, T., and Perlis, R

    URL http://arxiv.org/abs/1908.10084. Rumshisky, A., Ghassemi, M., Naumann, T., Szolovits, P., Castro, V ., McCoy, T., and Perlis, R. Predicting early psychiatric readmission with natural language processing of narrative discharge summaries.Translational psychiatry, 6(10):e921–e921,

    Pith/arXiv arXiv 1908

  7. [15]

    Dpga-textsyn: Differentially private genetic algorithm for synthetic text generation

    Sun, Z., Tian, Z., Song, Y ., Si, Y ., Zhang, J., Huang, M., Lu, K., Xiong, Z., Liu, X., and Li, D. Dpga-textsyn: Differentially private genetic algorithm for synthetic text generation. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 16159–16179,

    2025

  8. [16]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y ., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    Pith/arXiv arXiv

  9. [17]

    Xu, Z., Zhang, Y ., Andrew, G., Choquette, C., Kairouz, P., Mcmahan, B., Rosenstock, J., and Zhang, Y

    URL https: //openreview.net/forum?id=LWD7upg1ob. Xu, Z., Zhang, Y ., Andrew, G., Choquette, C., Kairouz, P., Mcmahan, B., Rosenstock, J., and Zhang, Y . Federated learning of gboard language models with differential privacy. In Sitaram, S., Beigman Klebanov, B., and Williams, J. D. (eds.),Proceedings of the 61st Annual Meeting of the As- sociation for Com...

    2023

  10. [18]

    Synthesizing and adapting error correction data for mobile large language model applications.arXiv preprint arXiv:2505.18488,

    Zhang, Y ., Xu, Z., Wu, S., Zhang, Y ., and Ramage, D. Synthesizing and adapting error correction data for mobile large language model applications.arXiv preprint arXiv:2505.18488,

    Pith/arXiv arXiv

  11. [19]

    Contrastive private data synthesis via weighted multi-plm fusion.arXiv preprint arXiv:2502.00245,

    Zou, T., Liu, Y ., Li, P., Xiong, Y ., Zhang, J., Liu, J., Ye, X., Ouyang, Y ., and Zhang, Y .-Q. Contrastive private data synthesis via weighted multi-plm fusion.arXiv preprint arXiv:2502.00245,

    Pith/arXiv arXiv

  12. [20]

    area" and

    A Additional Details of MAPLE A.1 Metadata extraction We follow the strategy of Hu et al. (2025) to extract metadata in the form of a structured tabular schema. We include the details here for self-completeness. Schema design.For the biorxiv dataset, we follow the same schema as in Hu et al. (2025) with an additional word count field recording the word co...

    2025

  13. [50]

    " " return t ex twr ap . dedent (

    ) , } Listing 2: Metadata schema for OpenReview dataset def a n n o t a t e _ f e a t u r e s _ p r o m p t ( text : str , ) -> str : " " " Returns a prompt to be used for feature e x t r a c t i o n from text . This fu nc tio n g e n e r a t e s a system prompt for an LLM to an no ta te a given text based on a p r e d e f i n e d set of fe at ur es . The...

    2025

  14. [2015]

    A., Huang, Y ., Yu, D., Kaissis, G., Charles, Z., Liu, R., et al

    Sinha, A., Mesnard, T., McKenna, R., Liu, D., Choquette-Choo, C. A., Huang, Y ., Yu, D., Kaissis, G., Charles, Z., Liu, R., et al. Vaultgemma: A differentially private gemma model.arXiv preprint arXiv:2510.15001,

    arXiv

  15. [2017]

    Harnessing large-language models to generate private synthetic text.arXiv preprint arXiv:2306.01684,

    Kurakin, A., Ponomareva, N., Syed, U., MacDermed, L., and Terzis, A. Harnessing large-language models to generate private synthetic text.arXiv preprint arXiv:2306.01684,

    Pith/arXiv arXiv

  16. [2018]

    Private prediction for large-scale synthetic text generation

    Amin, K., Bie, A., Kong, W., Kurakin, A., Ponomareva, N., Syed, U., Terzis, A., and Vassilvitskii, S. Private prediction for large-scale synthetic text generation. InFindings of the Association for Computational Linguistics: EMNLP 2024, pp. 7244–7262,

    2024

  17. [2019]

    and Gurevych, I

    Reimers, N. and Gurevych, I. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. As- sociation for Computational Linguistics, 11

    2019

  18. [2021]

    Synthetic query generation for privacy-preserving deep retrieval systems using differentially private language models

    Carranza, A., Farahani, R., Ponomareva, N., Kurakin, A., Jagielski, M., and Nasr, M. Synthetic query generation for privacy-preserving deep retrieval systems using differentially private language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume ...

    2024

  19. [2022]

    and Habernal, I

    10 Ochs, S. and Habernal, I. Private synthetic text generation with diffusion models. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 10612–10626,

    2025

  20. [2023]

    H., Constant, N., Ma, J., Hall, K., Cer, D., and Yang, Y

    Ni, J., Abrego, G. H., Constant, N., Ma, J., Hall, K., Cer, D., and Yang, Y . Sentence-t5: Scal- able sentence encoders from pre-trained text-to-text models. InFindings of the association for computational linguistics: ACL 2022, pp. 1864–1874,

    2022

  21. [2024]

    Actg-arl: Differentially private conditional text generation with rl-boosted control.arXiv preprint arXiv:2510.18232,

    Hu, Y ., McKenna, R., Yu, D., Wu, S., Zhao, H., Xu, Z., and Kairouz, P. Actg-arl: Differentially private conditional text generation with rl-boosted control.arXiv preprint arXiv:2510.18232,

    Pith/arXiv arXiv

  22. [2025]

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al

    URL https://storage.googleapis.com/deepmind-media/Model-Cards/ Gemini-3-Pro-Model-Card.pdf. Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.