แนวคิดหลัก
AutoDroid introduces a mobile task automation system powered by large language models (LLMs) to handle arbitrary tasks on Android applications efficiently.
บทคัดย่อ
AutoDroid is a novel system that leverages LLMs to automate tasks on Android apps. It combines UI representation, app-specific knowledge injection, and query optimization to enhance task completion rates and reduce computational costs.
The system explores app UIs, synthesizes simulated tasks, augments prompts with app memory, and fine-tunes local LLMs for improved performance. AutoDroid's multi-granularity query optimization reduces the frequency of LLM queries by merging functionally equivalent elements and utilizing shortcuts.
Overall, AutoDroid offers a comprehensive solution for efficient mobile task automation using cutting-edge technologies like large language models.
สถิติ
The accuracy of AutoDroid is 90.9%.
The success rate of completing tasks is 71.3%.
AutoDroid outperforms GPT-4-powered baselines by 36.4% and 39.7%.