mlsanityData to 14 Oct 2026

On the Price of Privacy for Language Identification and Generation

arXiv 2604.07238 ↗First posted 8 Apr 2026cs.LG cs.CL cs.CRcited by 3

Not on the map. Why?

Abstract (arXiv)

As large language models (LLMs) are increasingly trained on sensitive user data, understanding the fundamental cost of privacy in language learning becomes essential. We initiate the study of differentially private (DP) language identification and generation in the agnostic statistical setting, establishing algorithms and matching lower bounds that precisely quantify the cost of privacy. For both tasks, approximate (varepsilon, δ)-DP with constant varepsilon > 0 recovers the non-private error rates: exp(-r(n)) for identification (for any r(n) = o(n)) and exp(-Omega(n)) for generation. Under pure varepsilon-DP, the exponents degrade by a multiplicative factor of min1, varepsilon, which we show is tight up to constants. Notably, for generation under pure DP with mild assumptions, the upper bound exp(-min1,varepsilon · Omega(n)) matches the lower bound up to some constants, establishing an optimal rate. Our results show that the cost of privacy in language learning is surprisingly mild: absent entirely under approximate DP, and exactly a min1,varepsilon factor in the exponent under pure DP.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Builds on 49 of its 68 references that are in the map. First posted on arXiv 8 Apr 2026.

Builds on

49 of 68 references.

References · 49

  1. 2016Deep Learning with Differential Privacycited by 2,728
  2. 2017Renyi Differential Privacycited by 650
  3. 2021Large Language Models Can Be Strong Differentially Private Learnerscited by 412
  4. 2021Differentially Private Fine-tuning of Language Modelscited by 400
  5. 2018Improving the Gaussian Mechanism for Differential Privacy: Analytical Calibration and Optimal Denoisingcited by 218
  6. 2021Large-Scale Differentially Private BERTcited by 118
  7. 2022Synthetic Text Generation with Differential Privacy: A Simple and Practical Recipecited by 108
  8. 2013Fingerprinting Codes and the Price of Approximate Differential Privacycited by 82
  9. 2018Private PAC learning implies finite Littlestone dimensioncited by 74
  10. 2021Learning with User-Level Privacycited by 72
  11. 2020An Equivalence Between Private Classification and Online Predictioncited by 64
  12. 2014Private Learning and Sanitization: Pure vs. Approximate Differential Privacycited by 63
37 more
  1. 2015Differentially Private Release and Learning of Threshold Functionscited by 60
  2. 2015Between Pure and Approximate Differential Privacycited by 56
  3. 2020Differentially Private Assouad, Fano, and Le Camcited by 55
  4. 2022TAN Without a Burn: Scaling Laws of DP-SGDcited by 52
  5. 2020Learning discrete distributions: user vs item-level privacycited by 41
  6. 2020On the Sample Complexity of Privately Learning Unbounded High-Dimensional Gaussianscited by 41
  7. 2024Language Generation in the Limitcited by 40
  8. 2022Private Estimation with Public Datacited by 32
  9. 2014Sample Complexity Bounds on Differentially Private Learning via Communication Complexitycited by 31
  10. 2024Generation through the lens of learning theorycited by 31
  11. 2024Exploring Facets of Language Generation in the Limitcited by 28
  12. 2025Scaling Laws for Differentially Private Language Modelscited by 22
  13. 2025Density Measures for Language Generationcited by 21
  14. 2025Generation from Noisy Examplescited by 20
  15. 2025Representative Language Generationcited by 19
  16. 2020Locally Private Hypothesis Selectioncited by 17
  17. 2025On Union-Closedness of Language Generationcited by 17
  18. 2025Language Generation in the Limit: Noise, Loss, and Feedbackcited by 16
  19. 2025(Im)possibility of Automated Hallucination Detection in Large Language Modelscited by 15
  20. 2024On Characterizations for Language Generation: Interplay of Hallucinations, Breadth, and Stabilitycited by 14
  21. 2025Language Generation and Identification From Partial Enumeration: Tight Density Bounds and Topological Characterizationscited by 13
  22. 2025How to DP-fy Your Data: A Practical Guide to Generating Synthetic Data With Differential Privacycited by 13
  23. 2023User-Level Differential Privacy With Few Examples Per Usercited by 12
  24. 2024Data-adaptive Differentially Private Prompt Synthesis for In-Context Learningcited by 10
  25. 2025Pareto-optimal Non-uniform Language Generationcited by 10
  26. 2025Language Generation with Infinite Contaminationcited by 10
  27. 2025Learning Algorithms in the Limitcited by 9
  28. 2025A Characterization of List Language Identification in the Limitcited by 9
  29. 2025Language Generation: Complexity Barriers and Implications for Learningcited by 9
  30. 2026Safe Language Generation in the Limitcited by 9
  31. 2020A Computational Separation between Private Learning and Online Learningcited by 8
  32. 2025Towards hyperparameter-free optimization with differential privacycited by 7
  33. 2025DP-Fusion: Token-Level Differentially Private Inference for Large Language Modelscited by 6
  34. 2026Language Generation with Replay: A Learning-Theoretic View of Model Collapsecited by 5
  35. 2026Characterizing the Effect of Noise in Language Generation in the Limitcited by 4
  36. 2026Agnostic Language Identification and Generationcited by 4
  37. 2026On Generation in Metric Spacescited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Contrastive Identification and Generation in the Limitcited by 4
  2. 2026Flood and Harvest: The Provable Necessity of Trivia for Generating Valuable Mathematics via the Lens of Language Generation in the Limitcited by 2
  3. 2026Hallucination Rates in Language Generationcited by 0