模型互联网中基于自我效能的Token级多模型协作OA

To address the trade-off between inference performance and cost in Token-level collaboration within the AI-model network,a self-efficacy-based Token-level multi-model collaboration method named ConfiPara was proposed.Firstly,a Token-level collaborative method with an exit mechanism was designed to mitigate the high overhead of exis-ting approaches.Secondly,a self-efficacy assessment algorithm integrating the base model's confidence and reliability was introduced to determine the optimal exit timing.By leveraging self-efficacy to guide the base model in switching to independent inference at appropriate moments,redundant collaboration was skipped,thereby maintaining accuracy while reducing Token overhead.Experimental results demonstrate that the proposed ConfiPara method achieves a substantial reduction in Token consumption and inference latency with only a minor accuracy loss.In a single collaborative model scenario,the method reduces Token cost by approximately 21%and cuts per-Token generation latency by up to 75%,at the cost of only a 2.5%drop in accuracy.