V-Droid is a mobile GUI task automation agent. Unlike traditional LLMs that directly generate actions for each step, V-Droid utilizes LLMs as a verifier to evaluate candidate actions. To achieve this, we present a comprehensive framework that includes a dedicated workflow for building and pre-populating a discretized action space, interactive progress preference learning, and a scalable human-agent joint annotation scheme. On multiple mobile task automation benchmarks, including AndroidWorld, AndroidLab, and MobileAgentBench, V-Droid achieves higher success rates (59.5%, 38.3%, and 49%, respectively) and significantly faster processing speed (4.3 seconds per step, 6.1x faster than existing agents). The source code is available on GitHub.
