A mobile artificial intelligence (AI) agent developed by a Korean startup has achieved 100% accuracy in the mobile AI benchmark “AndroidWorld,” outperforming AI agents from big tech research organizations such as Google, Microsoft, and Alibaba.

Fluiz, a faculty-founded startup led by Shin In-sik, professor at KAIST’s School of Computing, announced on the 13th that its mobile AI agent “FluidGPT” achieved 100% accuracy for the first time in the AndroidWorld benchmark evaluation and ranked first on the official leaderboard.

A mobile AI agent is a program installed on smartphones that autonomously carries out complex tasks on behalf of users, such as operating apps, scheduling appointments, sending messages, and making financial payments. The mobile environment features highly diverse screen structures, frequent transitions as apps are opened and closed, and difficulties in constructing user interfaces (UI). For these reasons, building AI agents that operate precisely in mobile environments is regarded as a top-priority challenge for global big tech companies and academia.

AndroidWorld is a representative Android OS–based benchmark developed by the Google DeepMind research team to evaluate the performance of mobile AI agents. It verifies capabilities for handling a wide variety of complex real-world tasks that can arise in actual mobile environments.

FluidGPT competed against mobile AI agents developed by big tech research organizations and by academia and industry, including Google’s agent research team “Artemis,” Microsoft Research, and Alibaba. FluidGPT solved all 137 tasks presented by AndroidWorld with 100% accuracy, placing first among the mobile AI agents that successfully completed the AndroidWorld tasks.

The tasks presented by AndroidWorld go beyond simply finding and pressing buttons, requiring screen recognition, contextual understanding, decision-making, and long-horizon execution over dozens of steps, such as moving across multiple apps and remembering previously viewed content. For this reason, performance in AndroidWorld is considered an indicator of whether a mobile AI agent can move beyond demonstration-level behavior and fully complete real work tasks.

Shin In-sik, CEO of Fluiz, stated, “Achieving 100% in AndroidWorld proves that Fluiz’s FluidGPT technology possesses the world’s highest level of accuracy and practicality in the global mobile AI agent market,” adding, “Fluiz will grow into a global AI leader that spearheads the on-device AI and mobile operating system (OS) ecosystem.”

Popular News