Elamur Research⌕
☰
home / papers / 2026 / Persuasion Tokens for Editing Factual Knowledge in LLMs

Persuasion Tokens for Editing Factual Knowledge in LLMs

Paper · 2026 · Paul Youssef, Christin Seifert and Jörg Schlötterer

Summary

In-context knowledge editing (IKE) updates a fact in an LLM by placing the new fact in the prompt alongside 32 demonstrations. These demonstrations must be built separately for every edit, take up much of the context window, and slow down inference. The authors replace them with persuasion tokens (P-Tokens), which are learnable BEGIN_EDIT/END_EDIT special tokens wrapped around the new fact. Their embeddings are optimized with a KL-divergence loss so that the model's output distribution imitates what the full IKE prompt would produce. Training also covers paraphrased queries, neighboring facts that should stay unchanged, intervening distractor edits, and prompts with no edit at all. The method was tested on CounterFact and zsRE with GPT-J-6B, Qwen2.5-7B, Qwen2.5-14B and Llama3-8B. The abstract mentions three LLMs, but the setup lists four. P-Tokens matched or exceeded IKE, mostly because they disturbed neighboring facts less. Prompts were about 16 times shorter and inference ran roughly five times faster. This makes demonstration-free in-context editing a cheaper, more scalable option, although the upfront training cost only pays off at high inference volumes.

Key points

  • Data splits: CounterFact uses the 2,000-example IKE subset (800 train / 200 validation / 1,000 test). zsRE uses 800/200 sampled for training and validation and its full 19,086-example test set. zsRE has no IKE demonstrations, so it is compared to a plain-edit baseline instead.
  • CounterFact results: P-Tokens beat IKE on the overall score S for every model. On Llama3 this was 95.82 vs 93.97, and on Qwen-7B the gap was about 5 p.p. The NS (neighborhood) score improved by about 4 p.p., or about 10 p.p. on Qwen-7B.
  • Number of tokens: Using more P-Tokens (m ∈ {1, 3, 5, 7, 10} per side) generally helps. Moving from 1 to 3 tokens raised Llama3's zsRE Paraphrase score by about 16 p.p.
  • Distractors: With 100 distractor edits, ES/PS stay near 100, but NS falls by about 10–15 p.p. on Qwen-7B, Qwen-14B and Llama3. Training without distractors lowers Llama3's S at 100 distractors from 89.55 to 82.94.
  • Efficiency (Qwen2.5-7B, 10×2 tokens):
    • Prompt length is about 58 tokens vs about 959 for IKE.
    • Time per edit is 0.03 s vs 0.17 s.
    • Training takes about 15.5 hours, which is recovered after roughly 398k inferences.
    • Code is available at github.com/paulyoussef/p-tokens.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Philipps University of MarburgMarburg · Germany8 works · 226 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Paul Youssef, Christin Seifert and Jörg Schlötterer. 2026. Persuasion Tokens for Editing Factual Knowledge in LLMs. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.35

@inproceedings{youssef2026persuasion,
  title = {{Persuasion Tokens for Editing Factual Knowledge in LLMs}},
  author = {Paul Youssef, Christin Seifert and Jörg Schlötterer},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.35},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.35}
}

References

22 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Canyu Chen, Baixiang Huang, Zekun Li, Zhaorun Chen, Shiyang Lai, Xiongxiao Xu, Jia-Chen Gu, Jindong Gu, Huaxiu Yao, Chaowei Xiao, Xifeng Yan, William Yang Wang, Philip Torr, Dawn Song, and Kai Shu. 2024. Can Editing LLMs Inject Harm? In ICML 2024 Next Generation of AI Safety Workshop. find
  2. Sizhe Chen, Yizhu Wang, Nicholas Carlini, Chawin Sitawarin, and David Wagner. 2025. Defending Against Prompt Injection with a Few DefensiveTo-kens. In ICML 2025 Workshop on Reliable and Responsible Foundation Models. find
  3. Yinjie Cheng, Paul Youssef, Christin Seifert, Jörg Schlötterer, and Zhixue Zhao. 2025. Can Fine-Tuning Erase Your Edits? On the Fragile Coexistence of Knowledge Editing and Adaptation. Preprint, arXiv:2511.05852. — [arXiv:2511.05852](https://arxiv.org/abs/2511.05852) find
  4. Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-Shot Relation Extraction via Reading Comprehension. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 333–342, Vancouver, Canada. Association for Computational Linguistics.
  5. Xiaopeng Li, Shasha Li, Shangwen Wang, Shezheng Song, Bin Ji, Huijun Liu, Jun Ma, and Jie Yu. 2025. Identifying Knowledge Editing Types in Large Language Models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, KDD ’25, page 1553–1564, New York, NY, USA. A… find
  6. Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. find
  7. Vittorio Mazzia, Alessandro Pedrani, Andrea Caciolai, Kay Rottmann, and Davide Bernardi. 2024. A Survey on Knowledge Editing of Neural Networks. Preprint, arXiv:2310.19704. — [arXiv:2310.19704](https://arxiv.org/abs/2310.19704) find
  8. Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in GPT. Advances in neural information processing systems, 35:17359–17372. find
  9. Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. 2023. Mass-Editing Memory in a Transformer. In The Eleventh International Conference on Learning Representations. find
  10. Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022a. Fast Model Editing at Scale. In International Conference on Learning Representations. find
  11. Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2024. Evaluating the Ripple Effects of Knowledge Editing in Language Models. Transactions of the Association for Computational Linguistics, 12:283–298. find
  12. Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022b. Memory-Based Model Editing at Scale. In International Conference on Machine Learning. find
  13. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  14. Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language Models as Knowledge Bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural…
  15. Dongliang Guo, Mengxuan Hu, Zihan Guan, Thomas Hartvigsen, and Sheng Li. 2025. BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing. In Forty-second International Conference on Machine Learning. find
  16. Thomas Hartvigsen, Swami Sankaranarayanan, Hamid Palangi, Yoon Kim, and Marzyeh Ghassemi. 2023. Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors. In Thirty-seventh Conference on Neural Information Processing Systems. find
  17. Chenmien Tan, Ge Zhang, and Jie Fu. 2023. Massive Editng for Large Language Models via Meta Learning. arXiv preprint arXiv:2311.04661. — [arXiv:2311.04661](https://arxiv.org/abs/2311.04661) find
  18. Qwen Team. 2024. Qwen2.5: A Party of Foundation Models. find
  19. Ben Wang and Aran Komatsuzaki. 2021. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. https://github.com/kingoflolz/ mesh-transformer-jax. ↗
  20. Peng Wang, Zexi Li, Ningyu Zhang, Ziwen Xu, Yunzhi Yao, Yong Jiang, Pengjun Xie, Fei Huang, and Huajun Chen. 2024a. WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. find
  21. Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2024b. Knowledge Editing for Large Language Models: A Survey. ACM Comput. Surv., 57(3). pages 4862–4876, Singapore. Association for Computational Linguistics.
  22. A Evaluation. We follow previous work (Meng et al., 2022, 2023) in using dataset-specific metrics. For CounterFact, we use the following metrics: find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/youssef-2026-persuasion-tokens-editing-factual-knowledge/