Software Engineering
New Approach Reproduces Mobile App Performance Problems Using User Reviews
Researchers developed a system that synthesizes information from app store reviews to automatically detect and recreate hard-to-catch performance issues on Android devices.
Mobile app performance bugs often escape detection during development but emerge once users deploy the software on varied devices. A research team has unveiled RevPerf, an automated framework designed to bridge this gap by mining and analyzing user reviews to recreate these elusive issues.
The Challenge of Hidden Performance Problems#
Performance degradation in mobile applications remains one of the hardest problems for developers to diagnose. Unlike crashes or freezes that produce clear error signals, slowdowns and resource bottlenecks frequently depend on specific hardware configurations, network conditions, and usage patterns found only in real-world environments. Traditional testing cannot easily replicate these circumstances, leaving teams unable to pinpoint or fix the root causes.
User reviews, meanwhile, contain rich contextual details about when and how issues occur, yet they lack the structured format of formal bug reports. A single review may describe only a fragment of a problem, with critical details scattered across multiple reviews from different users.
How RevPerf Works#
The system tackles this ambiguity through several coordinated steps:
- Review Aggregation: RevPerf uses semantic search to locate multiple reviews describing the same or related performance issue, retrieving complementary accounts of the problem
- Information Synthesis: Prompt engineering merges data across reviews to construct a fuller picture of the issue, including device context and user actions
- Automated Reproduction: An execution agent translates this synthesized information into commands that recreate the scenario on an Android test environment
- Multi-Method Detection: During replay, the system monitors Android logs, visual interface changes, and resource metrics (CPU, memory, battery) to identify performance anomalies
Results and Performance#
Testing on a constructed dataset, RevPerf successfully reproduced performance issues at a 72.73% success rate, substantially outpacing the strongest baseline by more than 27 percentage points. This margin suggests the semantic retrieval and synthesis approach meaningfully improves reproducibility compared to treating reviews in isolation.
The work opens a path for development teams to leverage user feedback as an early warning system, turning unstructured app store commentary into actionable intelligence for performance optimization.