Preprint · 2026 · Open access

DP-OPD: Differentially Private On-Policy Distillation for Language Models

Fatemeh Khadem, Sajad Mousavi, Yi Fang, Yuhong Liu

arXiv (Cornell University) · Published 6 April 2026 · DOI 10.48550/arxiv.2604.04461

Open PDFView at the publisherCite this paper

In brief

“DP-OPD: Differentially Private On-Policy Distillation for Language Models” is a 2026 preprint by Fatemeh Khadem and 3 co-authors, published in arXiv (Cornell University). OpenAlex counts 4 citations as of 19 November 2026. It is open access: the full text is free to read at arXiv (Cornell University).

Citations
4
Most cited in
2026 (4)
References
15
Authors
4

Abstract

Large language models (LLMs) are increasingly adapted to proprietary and domain-specific corpora that contain sensitive information, creating a tension between formal privacy guarantees and efficient deployment through model compression. Differential privacy (DP), typically enforced via DP-SGD, provides record-level protection but often incurs substantial utility loss in autoregressive generation, where optimization noise can amplify exposure bias and compounding errors along long rollouts. Existing approaches to private distillation either apply DP-SGD to both teacher and student, worsening computation and the privacy--utility tradeoff, or rely on DP synthetic text generation from a DP-trained teacher, avoiding DP on the student at the cost of DP-optimizing a large teacher and introducing an offline generation pipeline. We propose \textbf{Differentially Private On-Policy Distillation (DP-OPD)}, a synthesis-free framework that enforces privacy solely through DP-SGD on the student while leveraging a frozen teacher to provide dense token-level targets on \emph{student-generated} trajectories. DP-OPD instantiates this idea via \emph{private generalized knowledge distillation} on continuation tokens. Under a strict privacy budget ($\varepsilon=2.0$), DP-OPD improves perplexity over DP fine-tuning and off-policy DP distillation, and outperforms synthesis-based DP distillation (Yelp: 44.15$\rightarrow$41.68; BigPatent: 32.43$\rightarrow$30.63), while substantially simplifying the training pipeline. In particular, \textbf{DP-OPD collapses private compression into a single DP student-training loop} by eliminating DP teacher training and offline synthetic text generation. Code will be released upon publication at https://github.com/khademfatemeh/dp_opd.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Khadem, F., Mousavi, S., Fang, Y., & Liu, Y. (2026). DP-OPD: Differentially Private On-Policy Distillation for Language Models. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2604.04461

MLA 9

Khadem, Fatemeh, et al. “DP-OPD: Differentially Private On-Policy Distillation for Language Models.” arXiv (Cornell University), 2026. https://doi.org/10.48550/arxiv.2604.04461.

Chicago

Khadem, Fatemeh, Sajad Mousavi, Yi Fang, and Yuhong Liu. 2026. “DP-OPD: Differentially Private On-Policy Distillation for Language Models.” arXiv (Cornell University). https://doi.org/10.48550/arxiv.2604.04461.

Harvard

Khadem, F. et al. (2026) ‘DP-OPD: Differentially Private On-Policy Distillation for Language Models’, arXiv (Cornell University). doi:10.48550/arxiv.2604.04461.

IEEE

F. Khadem, S. Mousavi, Y. Fang, and Y. Liu, “DP-OPD: Differentially Private On-Policy Distillation for Language Models,” arXiv (Cornell University), 2026, doi: 10.48550/arxiv.2604.04461.

BibTeX

@misc{khadem2026differentially,
  title     = {{DP-OPD: Differentially Private On-Policy Distillation for Language Models}},
  author    = {Khadem, Fatemeh and Mousavi, Sajad and Fang, Yi and Liu, Yuhong},
  howpublished = {arXiv (Cornell University)},
  year      = {2026},
  doi       = {10.48550/arxiv.2604.04461},
  url       = {https://doi.org/10.48550/arxiv.2604.04461}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data

  • differential privacy
  • knowledge distillation
  • on-policy distillation
  • large language models
  • DP-SGD
  • model compression
  • privacy-utility trade-off
  • exposure bias

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Khadem, F., Mousavi, S., Fang, Y., & Liu, Y. (2026). DP-OPD: Differentially Private On-Policy Distillation for Language Models. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2604.04461

How many times has this paper been cited?

4 times, according to OpenAlex on 19 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at arXiv (Cornell University): sign in and open the PDF from this page.

Who wrote this paper?

Fatemeh Khadem, Sajad Mousavi, Yi Fang, Yuhong Liu.

What is the DOI of this paper?

10.48550/arxiv.2604.04461, permanently at https://doi.org/10.48550/arxiv.2604.04461. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 19 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.