![[2508.07667] 1-2-3 Check: Enhancing Contextual Privacy in LLM via Multi-Agent Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2508.07667] 1-2-3 Check: Enhancing Contextual Privacy in LLM via Multi-Agent Reasoning
Summary
The paper presents a multi-agent framework to enhance contextual privacy in large language models (LLMs), demonstrating a significant reduction in private information leakage during information processing.
Why It Matters
As LLMs increasingly handle sensitive information, ensuring privacy is crucial. This research addresses privacy concerns by introducing a systematic approach that enhances the reliability of privacy adherence, making it relevant for developers and researchers in AI safety and ethics.
Key Takeaways
- Introduces a multi-agent framework to improve contextual privacy in LLMs.
- Demonstrates an 18% and 19% reduction in private information leakage on benchmark tests.
- Highlights the importance of information-flow design in multi-agent systems for privacy.
- Conducts systematic ablation studies to understand privacy error propagation.
- Outperforms single-agent baselines in maintaining public content fidelity.
Computer Science > Artificial Intelligence arXiv:2508.07667 (cs) [Submitted on 11 Aug 2025 (v1), last revised 25 Feb 2026 (this version, v3)] Title:1-2-3 Check: Enhancing Contextual Privacy in LLM via Multi-Agent Reasoning Authors:Wenkai Li, Liwen Sun, Zhenxiang Guan, Xuhui Zhou, Maarten Sap View a PDF of the paper titled 1-2-3 Check: Enhancing Contextual Privacy in LLM via Multi-Agent Reasoning, by Wenkai Li and 4 other authors View PDF HTML (experimental) Abstract:Addressing contextual privacy concerns remains challenging in interactive settings where large language models (LLMs) process information from multiple sources (e.g., summarizing meetings with private and public information). We introduce a multi-agent framework that decomposes privacy reasoning into specialized subtasks (extraction, classification), reducing the information load on any single agent while enabling iterative validation and more reliable adherence to contextual privacy norms. To understand how privacy errors emerge and propagate, we conduct a systematic ablation over information-flow topologies, revealing when and why upstream detection mistakes cascade into downstream leakage. Experiments on the ConfAIde and PrivacyLens benchmark with several open-source and closed-sourced LLMs demonstrate that our best multi-agent configuration substantially reduces private information leakage (\textbf{18\%} on ConfAIde and \textbf{19\%} on PrivacyLens with GPT-4o) while preserving the fidelity of public conten...
