Metric Freedom: First A Priori Predictor of When Multi-Agent to Single-Agent Skill Distillation Is Beneficial — Lift Ranges from +28% to -2% Across Same Task

arXiv·low signalXBlueskyLinkedInCopy link

arXiv:2604.01608 (April 2, 2026) empirically shows that distilling multi-agent systems into single-agent skills produces wildly inconsistent results — from 28% improvement to 2% degradation on the exact same task depending on the metric. The paper introduces Metric Freedom, which measures topological rigidity of a scoring landscape via a Mantel test, as the first predictor of whether distillation will help before you try it. Practical takeaway: before investing in collapsing your multi-agent pipeline into a single agent, test Metric Freedom on your evaluation metrics first.