AndroidWorld Leaderboard Tracks Multimodal Agents on Real Android Tasks

LLM Stats has published an AndroidWorld leaderboard summarizing reported results for multimodal AI agents that operate in Android graphical user interfaces.

What AndroidWorld measures

AndroidWorld is a benchmark environment from Google Research and Google DeepMind for testing autonomous agents on Android devices. The project page describes a suite of 116 programmatic tasks across 20 real-world Android apps, with links to the paper, code, data, and leaderboard.

Unlike benchmarks that focus only on text responses, AndroidWorld evaluates agents that must perceive a phone screen and take actions such as touching UI elements and typing. The LLM Stats benchmark page describes AndroidWorld as measuring multimodal agents in Android GUI environments through multi-step screen perception and touch/type actions.

That makes the benchmark relevant to a growing area of AI research: systems that can use software interfaces rather than only answer questions. In AndroidWorld, the agent is expected to navigate apps, interpret visual state, and complete tasks through the same kinds of interactions a human user would perform on a phone.

Current leaderboard entries

According to LLM Stats, its AndroidWorld leaderboard results were last updated on June 7, 2026. The linked Google Sheets community leaderboard says the results are community-submitted and self-reported. That disclosure matters: self-reported scores can be useful for tracking research activity, but they are not the same as independently audited evaluations.

The community leaderboard excerpt lists top AndroidWorld pass@1 scores including AGI-0, MobileUseAgent, and Finalrun at 97.4%. Pass@1 generally refers to success on the first attempt, although readers should consult the benchmark documentation and submission notes for the exact evaluation setup used for each listed result.

Because the leaderboard is community-maintained, comparisons should be interpreted carefully. Differences in environment setup, task selection, model access, prompting, tooling, or allowed retries can affect outcomes unless all submissions follow identical rules. The AndroidWorld project page and the Google Sheets leaderboard provide the primary context for what is being measured and how results are reported.

Why it matters

AndroidWorld sits in a practical testing category for AI agents: real application control. Many proposed AI assistants depend on the ability to use existing apps, not just generate text or code. A benchmark based on Android apps can help researchers examine whether agents can carry out multi-step procedures in dynamic interfaces, recover from visual changes, and interact reliably with mobile software.

The benchmark’s scope also highlights the difficulty of evaluating this class of systems. Mobile UI tasks combine visual understanding, planning, memory, and action execution. A high score on a leaderboard may indicate strong performance under a particular test configuration, but it does not automatically prove general reliability across all apps, devices, languages, account states, or unexpected UI changes.

For that reason, the most useful reading of the AndroidWorld leaderboard is as a snapshot of reported progress, not a final ranking of real-world capability. LLM Stats provides an accessible benchmark page, the AndroidWorld project page documents the environment, and the community spreadsheet offers current submitted results with the important caveat that they are self-reported.

Sources

• LLM Stats, “AndroidWorld Leaderboard”
• Google Sheets / AndroidWorld community leaderboard, “AndroidWorld Leaderboard”
• Google Research / Google DeepMind, “AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents”