Evaluation

We conduct human preference evaluation to evaluate XAgent’s performance. We prepare over 50 real-world complex tasks for assessment, which can be categorized into 5 classes: Search and Report, Coding and Developing, Data Analysis, Math, and Life Assistant. We compare the results of XAgent with AutoGPT, which shows a total win of XAgent over AutoGPT. All running records can refer to here.

We report a significant improvement of XAgent over AutoGPT in terms of human preference.