Conference paper · 2026 · Open access

Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models

David Ilić, David Stanojević, Kostadin Cvejoski

Proceedings of the conference - Association for Computational Linguistics. Meeting, 14077–14093 · Published 1 January 2026 · DOI 10.18653/v1/2026.acl-long.640

Open PDFView at the publisherCite this paper

In brief

“Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models” is a 2026 conference paper by David Ilić and 2 co-authors, published in Proceedings of the conference - Association for Computational Linguistics. Meeting. OpenAlex counts 5 citations as of 19 November 2026, which puts it in the top 3% of works published in 2026 by citations. It is open access under a CC BY licence: the full text is free to read at Proceedings of the conference - Association for Computational Linguistics. Meeting.

Citations
5
Most cited in
2026 (5)
Field-weighted impact
4.45
References
22
Authors
3

Abstract

Fine-tuned language models pose significant privacy risks, as they may memorize and expose sensitive information from their training data.Membership inference attacks (MIAs) provide a principled framework for auditing these risks, yet existing methods achieve limited detection rates, particularly at the low falsepositive thresholds required for practical privacy auditing.We present EZ-MIA, a membership inference attack that exploits a key observation: memorization manifests most strongly at error positions, specifically tokens where the model predicts incorrectly yet still shows elevated probability for training examples.We introduce the Error Zone (EZ) score, which measures the directional imbalance of probability shifts at error positions relative to a pretrained reference model.This principled statistic requires only two forward passes per query and no model training of any kind.On WikiText with GPT-2, EZ-MIA achieves 3.8× higher detection than the previous state-of-the-art under identical conditions (66.3% versus 17.5% true positive rate at 1% false positive rate), with near-perfect discrimination (AUC 0.98).At the stringent 0.1% FPR threshold critical for realworld auditing, we achieve 8× higher detection than prior work (14.0%versus 1.8%), requiring no reference model training.These gains extend to larger architectures: on AG News with Llama-2-7B, we achieve 3× higher detection (46.7% versus 15.8% TPR at 1% FPR).These results establish that privacy risks of finetuned language models are substantially greater than previously understood, with implications for both privacy auditing and deployment decisions.

Cite this paper

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

APA 7

Ilić, D., Stanojević, D., & Cvejoski, K. (2026). Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models. Proceedings of the conference - Association for Computational Linguistics. Meeting, 14077–14093. https://doi.org/10.18653/v1/2026.acl-long.640

MLA 9

Ilić, David, et al. “Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models.” Proceedings of the conference - Association for Computational Linguistics. Meeting, 2026, pp. 14077–14093. https://doi.org/10.18653/v1/2026.acl-long.640.

Chicago

Ilić, David, David Stanojević, and Kostadin Cvejoski. 2026. “Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models.” Proceedings of the conference - Association for Computational Linguistics. Meeting: 14077–14093. https://doi.org/10.18653/v1/2026.acl-long.640.

Harvard

Ilić, D., Stanojević, D. and Cvejoski, K. (2026) ‘Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models’, Proceedings of the conference - Association for Computational Linguistics. Meeting, pp. 14077–14093. doi:10.18653/v1/2026.acl-long.640.

IEEE

D. Ilić, D. Stanojević, and K. Cvejoski, “Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models,” Proceedings of the conference - Association for Computational Linguistics. Meeting, pp. 14077–14093, 2026, doi: 10.18653/v1/2026.acl-long.640.

BibTeX

@inproceedings{ilic2026powerful,
  title     = {{Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models}},
  author    = {Ilić, David and Stanojević, David and Cvejoski, Kostadin},
  booktitle = {Proceedings of the conference - Association for Computational Linguistics. Meeting},
  year      = {2026},
  pages     = {14077--14093},
  doi       = {10.18653/v1/2026.acl-long.640},
  url       = {https://doi.org/10.18653/v1/2026.acl-long.640}
}

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

  • membership inference attacks
  • training data memorization
  • fine-tuned language models
  • privacy auditing
  • autoregressive language models
  • GPT-2
  • Llama 2

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?

Ilić, D., Stanojević, D., & Cvejoski, K. (2026). Powerful Training-Free Membership Inference Against Fine-Tuned Autoregressive Language Models. Proceedings of the conference - Association for Computational Linguistics. Meeting, 14077–14093. https://doi.org/10.18653/v1/2026.acl-long.640

How many times has this paper been cited?

5 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 Proceedings of the conference - Association for Computational Linguistics. Meeting: sign in and open the PDF from this page.

Who wrote this paper?

David Ilić, David Stanojević, Kostadin Cvejoski.

What is the DOI of this paper?

10.18653/v1/2026.acl-long.640, permanently at https://doi.org/10.18653/v1/2026.acl-long.640. 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.