Premature Consensus Hurts Multi-Agent Research: Isolation Boosts Accuracy to 82.6%

rohanpaul_ai · x · 2026-09-08

A paper identifies 'premature consensus' in multi-agent research: once one agent finds a plausible answer, others search around the same idea instead of exploring alternatives. ArcticSwarm blocks selected agents from reading peers during search, then reviews findings together. On BrowseComp-Plus with Qwen 3.5-27B it hits 82.6% accuracy; removing isolation drops it to 78.8%, removing review to 74.5%, while 40 independent single-agent runs with majority voting reach only 63.5%. Lesson: let agents explore different paths first, then challenge the leading answer before consensus.

Related event: Delayed Consensus Boosts Multi-Agent Research Performance(2 posts)→

Original post →

More from coding & agent

coding & agent channel →