ArcticSwarm improves research accuracy by delaying agent collaboration, a post reports

A user discussing the ArcticSwarm paper says early collaboration can steer AI research agents toward the same wrong answer—a problem the paper calls “premature consensus.” The post reports 82.6% accuracy on BrowseComp-Plus with Qwen 3.5-27B, falling to 78.8% without search isolation and 74.5% when the review system was removed too. It says 40 independent single-agent runs with majority voting reached 63.5%. The author's takeaway: let agents explore different paths independently, then challenge and verify the leading answer before agreeing.