Blocking subagents from seeing each other's partial findings raises deep-research accuracy to 82.6%

arXiv·medium signalXBlueskyLinkedInCopy link

ArcticSwarm separates evidence gathering from evidence integration: subagents publish to a shared bulletin board, but gated isolation lets selected search tasks keep their own prior so parallel agents stop converging on an early candidate before alternatives are tested, and structured review at three commitment boundaries only propagates confident candidates. On full BrowseComp-Plus with open-weight Qwen 3.5-27B it reaches 82.6%, against 78.8% without gated isolation and 74.5% with structured review also disabled, beating aligned MiroFlow runs at 70.6%. On live-web BrowseComp with GPT-5 it hits 73.6% versus a reported 54.9% provider system and MiroFlow's 63.4%. The counterintuitive lesson is that restricting peer visibility improves a research swarm.