PAC-BENCH: Privacy Constraints Substantially Degrade Multi-Agent Collaboration Performance

arXiv·medium signalXBlueskyLinkedInCopy link

Park et al. present PAC-BENCH, a benchmark for evaluating multi-agent collaboration under privacy constraints — the scenario where dedicated AI agents from different organizations must collaborate without sharing sensitive data. Experiments show that privacy constraints substantially degrade collaboration performance, quantifying a tension that every enterprise deploying cross-organizational agent systems will face. The benchmark provides the first systematic measurement of this tradeoff.