Elamur Research⌕
☰
home / papers / 2026 / Beyond Local vs. External: A Game-Theoretic Framework for Trustworthy Knowledge Acquisition

Beyond Local vs. External: A Game-Theoretic Framework for Trustworthy Knowledge Acquisition

Paper · 2026 · Rujing Yao, Yufei Shi, Yang Wu et al.

Summary

The paper addresses a dilemma in using LLMs for sensitive work. Cloud-hosted models such as GPT-5 give strong answers, but sending them raw queries can reveal confidential research or legal intent. Purely local models keep data private but answer less well. The authors propose GTKA, a client-side, semantic-level framework that frames this trade-off as an adversarial game. A local generator splits a sensitive query into generalized sub-queries that hide specific entities. These go to an external LLM, while a local attacker tries to reconstruct the original query from the fragments. A trusted local integrator, the only component that sees the original question, merges the external answers into a final response. The generator is rewarded for answer quality and penalized for leakage, and it is trained alternately against the attacker. For evaluation, the authors built two benchmarks: BioQA, from PubMed, and LawQA, from U.S. court decisions. GTKA reduced intent leakage compared with baselines while keeping answer quality high. It needs no cryptography and no changes to existing LLM services.

Key points

  • Reward and training: The reward combines answer quality and leakage as α·Quality − β·Leakage. The generator samples K candidate sub-query groups per query, and DPO-based optimization is part of training.
  • Human evaluation on BioQA: Three domain experts rated answers on a 5-point scale, with GTKA compared against Soft-PO:
    • accuracy: 4.65 vs 4.17
    • completeness: 4.54 vs 4.03
    • adoptability: 4.72 vs 3.96
  • Lower intent leakage than Soft-PO:
    • ASR@1: 21.67 vs 24.25
    • ASR@3: 40.13 vs 42.84
    • MRR: 31.26 vs 33.65
  • Ablation on BioQA:
    • Full GTKA scored R-L 16.42, BertScore 82.85, ASR@3 36.14 and MRR 28.43.
    • Dropping the leakage term (β = 0) raised answer quality slightly to R-L 17.08, but leakage rose to ASR@3 40.32.
    • Dropping the quality term (α = 0) lowered leakage but hurt answer quality (R-L 14.47).
  • Limitations:
    • Testing covered only the biomedical and legal domains.
    • Experiments used a single external LLM provider; spreading sub-queries across several providers is left for future work.
    • All attacks were run only on local models.

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

🎓Nanyang Technological UniversitySingapore · Singapore184 works · 27,484 citations🎓Zhejiang UniversityHangzhou · China224 works · 13,587 citations🎓Hong Kong Polytechnic UniversityHong Kong · China106 works · 5,190 citations🎓Indiana University BloomingtonBloomington · United States40 works · 4,003 citations🎓Nankai UniversityTianjin · China36 works · 762 citations🎓Worcester Polytechnic InstituteWorcester · United States12 works · 380 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

Cite

Cite as Rujing Yao, Yufei Shi, Yang Wu, Ang Li, Zhuoren Jiang, XiaoFeng Wang, Haixu Tang and Xiaozhong Liu. 2026. Beyond Local vs. External: A Game-Theoretic Framework for Trustworthy Knowledge Acquisition. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1299

@inproceedings{yao2026beyond,
  title = {{Beyond Local vs. External: A Game-Theoretic Framework for Trustworthy Knowledge Acquisition}},
  author = {Rujing Yao, Yufei Shi, Yang Wu, Ang Li, Zhuoren Jiang, XiaoFeng Wang, Haixu Tang and Xiaozhong Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1299},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1299}
}

References

35 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Naafey Aamer, Muhammad Nabeel Asim, Shan Munir, and Andreas Dengel. 2025. Automating ai discovery for biomedicine through knowledge graphs and llm agents. bioRxiv, pages 2025–05. find
  2. Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318. find
  3. Tamer Başar and Geert Jan Olsder. 1998. Dynamic noncooperative game theory. SIAM. find
  4. Amos Beimel. 2011. Secret-sharing schemes: A survey. In International conference on coding and cryptology, pages 11–46. Springer. find
  5. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel. 2021. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21)…
  6. Arup Kumar Chattopadhyay, Sanchita Saha, Amitava Nag, and Sukumar Nandi. 2024. Secret sharing: A comprehensive survey, taxonomy and applications. Computer Science Review, 51:100608. find
  7. Kang Chen, Xiuze Zhou, Yuanguo Lin, Shibo Feng, Li Shen, and Pengcheng Wu. 2025. A survey on privacy risks and protection in large language models. Journal of King Saud University Computer and Information Sciences, 37(7):163. find
  8. Xu Chu, Zhixin Zhang, Tianyu Jia, and Yujie Jin. 2025. Stackelberg game preference optimization for data-efficient alignment of language models. arXiv preprint arXiv:2502.18099. — [arXiv:2502.18099](https://arxiv.org/abs/2502.18099) find
  9. Badhan Chandra Das, M Hadi Amini, and Yanzhao Wu. 2025. Security and privacy challenges of large language models: A survey. ACM Computing Surveys, 57(6):1–39.
  10. Giuseppe De Nittis and Francesco Trovo. 2016. Machine learning techniques for stackelberg security games: a survey. arXiv preprint arXiv:1609.09341. — [arXiv:1609.09341](https://arxiv.org/abs/1609.09341) find
  11. Tanner Fiez, Benjamin Chasnov, and Lillian Ratliff. 2020. Implicit learning dynamics in stackelberg games: Equilibria characterization, convergence analysis, and empirical study. In International conference on machine learning, pages 3133–3144. PMLR. find
  12. Craig Gentry. 2009. Fully homomorphic encryption using ideal lattices. In Proceedings of the forty-first annual ACM symposium on Theory of computing, pages 169–178. find
  13. Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572.
  14. Zhengye Han and Quanyan Zhu. 2025. A dynamic stackelberg game framework for agentic ai defense against llm jailbreaking. arXiv preprint arXiv:2507.08207. — [arXiv:2507.08207](https://arxiv.org/abs/2507.08207) find
  15. Zhigang Kan, Linbo Qiao, Hao Yu, Liwen Peng, Yifu Gao, and Dongsheng Li. 2023. Protecting user privacy in remote conversational systems: A privacy-preserving framework based on text sanitization. arXiv preprint arXiv:2306.08223. — [arXiv:2306.08223](https://arxiv.org/abs/2306.08223) find
  16. Jiang Lai, Zheng Conghui, Zhang Xiaohan, Sun Fuhui, Wang Xiaoyan, and Pan Li. 2024. Leveraging llm based retrieval-augmented generation for legal knowledge graph completion. In 2024 IEEE 9th International Conference on Data Science in Cyberspace (DSC), pages 196–203. IEEE. find
  17. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Proc…
  18. Jiayang Li, Jing Yu, Yu Nie, and Zhaoran Wang. 2020. End-to-end learning and intervention in games. Advances in Neural Information Processing Systems, 33:16653–16665. find
  19. Tao Li, Henger Li, Yunian Pan, Tianyi Xu, Zizhan Zheng, and Quanyan Zhu. 2024. Meta stackelberg game: Robust federated learning against adaptive and mixed poisoning attacks. arXiv preprint arXiv:2410.17431. — [arXiv:2410.17431](https://arxiv.org/abs/2410.17431) find
  20. Lingjuan Lyu, Han Yu, and Qiang Yang. 2020. Threats to federated learning: A survey. arXiv preprint arXiv:2003.02133. — [arXiv:2003.02133](https://arxiv.org/abs/2003.02133) find
  21. Fan Mo, Hamed Haddadi, Kleomenis Katevas, Eduard Marin, Diego Perino, and Nicolas Kourtellis. 2021. Ppfl: Privacy-preserving federated learning with trusted execution environments. In Proceedings of the 19th annual international conference on mobile systems, applications, and services, pages 94–108. find
  22. Olga Ohrimenko, Felix Schuster, Cédric Fournet, Aastha Mehta, Sebastian Nowozin, Kapil Vaswani, and Manuel Costa. 2016. Oblivious {Multi-Party} machine learning on trusted processors. In 25th USENIX Security Symposium (USENIX Security 16), pages 619–636. find
  23. Sanjay Oruganti, Sergei Nirenburg, Jesse English, and Marjorie McShane. 2023. Automating knowledge acquisition for content-centric cognitive agents using llms. In Proceedings of the AAAI Symposium Series, volume 2, pages 379–385. find
  24. Shuchao Pang, Zhigang Lu, Haichen Wang, Peng Fu, Yongbin Zhou, and Minhui Xue. 2024. Reconstruction of differentially private text sanitization via large language models. arXiv preprint arXiv:2410.12443. — [arXiv:2410.12443](https://arxiv.org/abs/2410.12443) find
  25. David Pape, Sina Mavali, Thorsten Eisenhofer, and Lea Schönherr. 2025. Prompt obfuscation for large language models. In 34th USENIX Security Symposium (USENIX Security 25), pages 2323–2342. find
  26. Sanjaikanth E Vadakkethil Somanathan Pillai and Kiran Polimetla. 2024. Enhancing network privacy through secure multi-party computation in cloud environments. In 2024 International Conference on Integrated Circuits and Communication Systems (ICI-CACS), pages 1–6. IEEE. find
  27. Maciej Pondel, Iwona Chomiak-Orsa, Małgorzata Sobińska, Wojciech Grzelak, Artur Kotwica, Andrzej Małowiecki, Kamila Łuczak, Andrzej Greńczuk, Peter Busch, David Chudán, and Petr Berka. 2024. Ai tools for knowledge management–knowledge base creation via llm and rag for ai assistant. In European Con… find
  28. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741.
  29. Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. 2017. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083.
  30. Sofia Eleni Spatharioti, David Rothschild, Daniel G Goldstein, and Jake M Hofman. 2025. Effects of llm-based search on decision making: Speed, accuracy, and overreliance. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pages 1–15. find
  31. Jacob Makar-Limanov, Arjun Prakash, Denizalp Goktas, Nora Ayanian, and Amy Greenwald. 2024. Sta-rlhf: Stackelberg aligned reinforcement learning with human feedback. In Coordination and Cooperation for Multi-Agent Reinforcement Learning Methods Workshop. find
  32. Tan Su, Bingbing Zhang, Chi Zhang, and Lingbo Wei. 2024. Privacy leak detection in llm interactions with a user-centric approach. In 2024 IEEE 23rd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom), pages 1647–1652. IEEE. find
  33. Latanya Sweeney. 2002. k-anonymity: A model for protecting privacy. International journal of uncertainty, fuzziness and knowledge-based systems, 10(05):557– 570. find
  34. Saiteja Utpala, Sara Hooker, and Pin-Yu Chen. 2023. Locally differentially private document generation using zero shot prompting. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8442–8457. find
  35. Chong Wang, Mengyao Li, Junjun He, Zhongruo Wang, Erfan Darzi, Zan Chen, Jin Ye, Tianbin Li, Yanzhou Su, Jing Ke, Kaili Qu, Shuxin Li, Yi Yu, Pietro Liò, Tianyun Wang, Yu Guang Wang, and Yiqing Shen. 2025. A survey for large language models in biomedicine. Artificial Intelligence in Medicine, page 1… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yao-2026-beyond-local-vs-external-game/