← 研踪 PaperTrailarXiv:2611.16603 · v1
Regime-Uncertain Data Transactions: A Game-Theoretic Framework for Self-Adapting and Self-Diagnosing Multi-ML-Agent Systems
首次提交 2026-11-12 18:38
版本更新 2026-11-12 18:38
雷达观测 2026-11-13 14:15
摘要
Autonomous ML agents in large organizations exchange data and model outputs across organizational and legal boundaries. Which rules bind a given exchange is not fixed. It depends on where the data subjects live, where the agents run, who owns them, whom they target, and which version of each rule is in force. Tags that describe the data can be noisy, and a counterparty can overclaim its safeguards. We model one exchange as a \emph{regime-uncertain transfer game} between two ML agents, with an enforcer as a third player. Rules enter as hard constraints with a strictest-wins structure. Agents hold a Bayesian belief over which rules apply, can pay to unlock safeguards or verify claims, and then bargain over the remaining legal bundles.Our method, Legally-Constrained Nash Bargaining (\LCNB), implements this with a Monotonic Concession Protocol using Zeuthen's rule. We prove four properties of the model and test it in a synthetic simulator with GDPR, DPDP, CCPA/CPRA and a stylised stricter regime. In two of three stress cases where residence and citizenship differ, a citizenship-based rule violates in 21--37\% of shared records, and \LCNB{} in under 1\%. The results also show limits. With our default economics, \LCNB{} shares almost only aggregates, and its low violation rate follows partly from that. Violation tolerance, tag noise and enforcement strength matter more than the choice of bargaining solution.
作者备注：Working paper. Developing for Conference Submission
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-12 18:38 · 最新观测版本
Regime-Uncertain Data Transactions: A Game-Theoretic Framework for Self-Adapting and Self-Diagnosing Multi-ML-Agent Systems
Autonomous ML agents in large organizations exchange data and model outputs across organizational and legal boundaries. Which rules bind a given exchange is not fixed. It depends on where the data subjects live, where the agents run, who owns them, whom they target, and which version of each rule is in force. Tags that describe the data can be noisy, and a counterparty can overclaim its safeguards. We model one exchange as a \emph{regime-uncertain transfer game} between two ML agents, with an enforcer as a third player. Rules enter as hard constraints with a strictest-wins structure. Agents hold a Bayesian belief over which rules apply, can pay to unlock safeguards or verify claims, and then bargain over the remaining legal bundles.Our method, Legally-Constrained Nash Bargaining (\LCNB), implements this with a Monotonic Concession Protocol using Zeuthen's rule. We prove four properties of the model and test it in a synthetic simulator with GDPR, DPDP, CCPA/CPRA and a stylised stricter regime. In two of three stress cases where residence and citizenship differ, a citizenship-based rule violates in 21--37\% of shared records, and \LCNB{} in under 1\%. The results also show limits. With our default economics, \LCNB{} shares almost only aggregates, and its low violation rate follows partly from that. Violation tolerance, tag noise and enforcement strength matter more than the choice of bargaining solution.
查看此版本解读 →查看此版本原文 ↗