面向移动端智能体的端到端自动化评测框架OA

With the growing deployment of mobile agents on smartphones,there remains a lack of a unified automated framework for evaluat-ing their functionality,performance,and safety.To address this,this paper proposes MESA:a unified automated evaluation framework for mo-bile agents.First,a multi-dimensional metric system is constructed to define evaluation standards.Then,based on this system,an end-to-end automated mechanism is built to generate user-level tasks,drive agent operations,collect interaction data,and automatically assess task completion and safety risks.Finally,a unified interface and abstraction model are designed to enable cross-platform adaptation.Comparative experiments with existing baseline evaluation methods show that MESA provides finer-grained characterization of task completion,a higher de-gree of automation in the evaluation pipeline,and broader coverage of functional evaluation metrics.Using MESA,we evaluate ten mainstream domestic and international mobile agents;the average task completion rate is 30.74％and the average latency is 13.49 s.In terms of safety,the incidence of prompt-injection attacks is 15.72％,the average data leakage and privacy violation rate is 23.28％,the average unauthorized tool use and permission abuse rate is 19.64％,and the average instruction drift and task interruption rate is 37.10％.These results validate the effectiveness of the proposed framework and provide a unified tool and methodology for evaluating mobile agents.