LLM-Explorer: Towards Efficient and Affordable LLM-based Exploration for Mobile Apps

TL;DR

The paper tackles the high cost and inefficiency of LLM-based mobile app exploration by proposing LLM-Explorer, a two-module system that emphasizes knowledge maintenance over action generation. It introduces abstract UI states, abstract actions, and an Abstract Interaction Graph to compress exploration knowledge and guide efficient navigation, using LLMs mainly for knowledge organization and content-aware input generation. Empirical results across 20 apps show higher activity coverage and dramatically lower LLM costs (up to ~148x less) than strong baselines, with the approach achieving human-level performance on some apps. The work demonstrates that a knowledge-centric, LLM-aware exploration framework can enable affordable, scalable automated app analysis, and it provides open-source code for broader adoption.

Abstract

Large language models (LLMs) have opened new opportunities for automated mobile app exploration, an important and challenging problem that used to suffer from the difficulty of generating meaningful UI interactions. However, existing LLM-based exploration approaches rely heavily on LLMs to generate actions in almost every step, leading to a huge cost of token fees and computational resources. We argue that such extensive usage of LLMs is neither necessary nor effective, since many actions during exploration do not require, or may even be biased by the abilities of LLMs. Further, based on the insight that a precise and compact knowledge plays the central role for effective exploration, we introduce LLM-Explorer, a new exploration agent designed for efficiency and affordability. LLM-Explorer uses LLMs primarily for maintaining the knowledge instead of generating actions, and knowledge is used to guide action generation in a LLM-less manner. Based on a comparison with 5 strong baselines on 20 typical apps, LLM-Explorer was able to achieve the fastest and highest coverage among all automated app explorers, with over 148x lower cost than the state-of-the-art LLM-based approach.

Figures (13)

• Figure 1: The workflow of LLM-Explorer.
• Figure 2: The prompt for knowledge organization. From top to bottom: general instructions, UI representation, chain-of-thought module, output format, and response, respectively.
• Figure 3: The prompt for content-aware input generation. From top to bottom, the prompt comprises: overall guidance, page representation, input request, response format, and response, respectively.
• Figure 4: Progressive activity coverage of LLM-Explorer and baselines over time. The brown dotted line is the reference human performance.
• Figure 5: Progressive activity coverage of LLM-Explorer and baselines over steps within 2 hours. The brown dotted line is the reference human performance. The maximum step differs for each method due to the different per-step time. Note that Monkey produced over 20,000 steps in 2 hours of exploration, but the activity coverage nearly converged after 10,000 steps, so this figure only shows up to 10,000 steps.