MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models
2026-06-03 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors present MIRAGE, a system that helps software agents control apps more efficiently by learning to think in a compact hidden space instead of writing out long step-by-step explanations. This hidden thinking is trained to predict what the app screen will look like next, helping the agent plan better before acting. MIRAGE reduces the amount of text the agent needs to generate, speeding up interaction while maintaining or improving task performance. They tested it on Android app tasks and found it uses fewer resources but works as well or better than previous methods.
mobile agentslatent representationschain-of-thought reasoningworld modelscreen affordancesmulti-step navigationinference efficiencyAndroid automationgenerative modeling
Authors
Zhichao Yang, Yuanze Hu, Haojie Hao, Longkun Hao, Dongshuo Huang, Hongyu Lin, Gen Li, Lanqing Hong, Yihang Lou, Yan Bai
Abstract
Mobile agents are increasingly expected to operate everyday applications from screenshots and language goals, where reliable control requires reasoning over screen affordances, multi-step navigation, and future state changes. However, many agents externalize this computation as long textual chains of thought, which slows interaction, increases supervision cost, and complicates deployment. We introduce MIRAGE, a framework that learns continuous latent reasoning representations from visible textual reasoning traces. MIRAGE transfers explicit reasoning into compact hidden states, enabling the agent to reason internally without decoding long rationales. It also incorporates a generative world-model objective: latent reasoning vectors are aligned with future screenshots, encouraging the agent to anticipate upcoming interface states before acting. This turns hidden computation into both a compressed thought representation and a forward-looking model of environment dynamics. At inference time, MIRAGE reasons in continuous latent space, reducing token generation while improving execution efficiency. On AndroidWorld, MIRAGE matches explicit chain-of-thought supervised fine-tuning in the 4B ablation with a 3-5x lower decoded-token budget and improves a comparable instruction-tuned baseline by 10.2 points; on AndroidControl, it improves action grounding while generating over 75% fewer tokens.