跳到正文
arXiv cs.MA· Zhuoning Xu, Xiucheng Zhang, Hanjun Luo, Yingbin Jin, Yinpeng Dong, Hanan Salam·· 2 天前AI 评分34

MasDrift:跨多智能体架构的授权保持基准测试

MasDrift: Benchmarking Authorization Preservation Across Multi-Agent Architectures

AI 导读

研究者推出 MasDrift 基准,包含 8 个领域的 600 个良性生产力任务,用于评测多智能体系统在委派任务时能否保留原始授权边界。测试显示,集中式层级架构任务完成率达 93.9–98.6%,高于对等网络的 85.7–87.0%,但未授权操作占比 2.7–19.8%,远高于后者的 0.6–0.8%,且差距随层级加深而扩大。

正文 · 原文

View PDF HTML (experimental)

Abstract:Multi-agent systems (MAS) decompose long-horizon tasks across supervisors and subagents, but delegated goals do not necessarily carry their original authorization boundaries. Existing safety benchmarks mainly study adversarial compromise, while work on constraint drift lacks controlled architecture-level evaluation. We introduce MasDrift, a benchmark of 600 benign productivity tasks across eight domains. Each task pairs required work with reserved actions. MasDrift compares single-agent, centralized, and decentralized coordination while varying hierarchy depth and peer width, measuring task completion and authorization preservation. Across generic multi-agent conditions, centralized hierarchies achieve 93.9--98.6% task completion versus 85.7--87.0% for peer networks, while unauthorized actions occur in 2.7--19.8% of tasks versus 0.6--0.8%, a gap that widens with hierarchy depth. We further compare two defenses that differ in where authorization evidence resides. One re-anchors every pending call to the original user request. The other carries an attenuated policy along the delegation chain. Re-anchoring reduces unauthorized actions in every model configuration we evaluate, at a cost of 1.6 points of pooled completion. Chain propagation blocks required work instead, forfeiting up to 36.3 points. A heterogeneous case study confirms that the failure follows from coordination rather than model strength. MasDrift exposes a centralization tradeoff and makes authorization preservation a measurable property of MAS design.
Comments: preprint
Subjects: Multiagent Systems (cs.MA); Artificial Intelligence (cs.AI)
Cite as: arXiv:2608.07556 [cs.MA]
  (or arXiv:2608.07556v3 [cs.MA] for this version)
  https://doi.org/10.48550/arXiv.2608.07556

arXiv-issued DOI via DataCite

Submission history

From: Hanjun Luo [view email]
[v1] Sun, 2 Aug 2026 04:46:15 UTC (686 KB)
[v2] Tue, 11 Aug 2026 07:14:50 UTC (690 KB)
[v3] Thu, 15 Oct 2026 07:27:12 UTC (665 KB)

来源:arXiv cs.MA · arxiv.org