PAC-BENCH: Evaluating Multi-Agent Collaboration under Privacy Constraints

TL;DR

PAC-BENCH is a benchmark designed to systematically evaluate how privacy constraints impact multi-agent collaboration, revealing significant performance degradation and coordination challenges.

Contribution

This work introduces PAC-BENCH, the first benchmark for assessing multi-agent collaboration under privacy constraints, highlighting key challenges and areas for future research.

Findings

Privacy constraints significantly reduce collaboration effectiveness.

Outcomes depend more on the initiating agent under privacy restrictions.

Coordination breakdowns are driven by privacy violations and hallucinations.

Abstract

We are entering an era in which individuals and organizations increasingly deploy dedicated AI agents that interact and collaborate with other agents. However, the dynamics of multi-agent collaboration under privacy constraints remain poorly understood. In this work, we present $PAC\text{-}Bench$, a benchmark for systematic evaluation of multi-agent collaboration under privacy constraints. Experiments on $PAC\text{-}Bench$ show that privacy constraints substantially degrade collaboration performance and make outcomes depend more on the initiating agent than the partner. Further analysis reveals that this degradation is driven by recurring coordination breakdowns, including early-stage privacy violations, overly conservative abstraction, and privacy-induced hallucinations. Together, our findings identify privacy-aware multi-agent collaboration as a distinct and unresolved challenge that requires new coordination mechanisms beyond existing agent capabilities.