模型互联网中基于输出相似度分簇的模型匹配方法OA

To address the challenge of balancing efficiency and gain in collaborative model matching within the AI-model network,a method named output similarity clustering based on model matching(OSC-MM)was proposed.Token-level behavioral differences between models were characterized via a similarity measurement approach that decouples the leader and synchronized models.The resulting similarity measurements were then used for clustering,and a represen-tative model mechanism was introduced to reduce matching complexity.Small-scale validation was employed to com-plete collaborative matching when a new model joined.Compared with traditional empirical measurement methods,the model matching overhead was reduced by 42％～68％using OSC-MM.Moreover,the method prioritizes complementary model combinations and thereby mitigates the risk of homogeneous parallel deployment.