Research Introduces Metric Freedom to Predict AI Skill Distillation Success

A new research paper introduces Metric Freedom (F), a predictive measure for determining when to distill multi-agent AI systems into single-agent skills. Published on arXiv under identifier 2604.01608v2, the work addresses the inconsistent outcomes of skill distillation, where performance can vary from a 28% improvement to a 2% degradation across identical tasks. The researchers found that skill utility depends not on the task itself but on the evaluation metric used. Metric Freedom quantifies the topological rigidity of a metric's scoring landscape by measuring how output diversity correlates with score variance through a Mantel test. Based on this predictor, the team developed AdaSkill, a two-stage adaptive distillation framework. Stage 1 of AdaSkill functions as a selective extraction process, aiming to bypass the coordination overhead, context fragmentation, and brittle phase ordering often associated with multi-agent systems. The study reveals that the empirical benefits of converting multi-agent expertise into consolidated skills have lacked a principled approach until now.