MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models
Zhichao Yang, Yuanze Hu, Haojie Hao, Longkun Hao, Dongshuo Huang, Hongyu Lin, Gen Li, Lanqing Hong, Yihang Lou, Yan Bai
Why It Matters
What makes this one worth your time
This work addresses the inefficiencies in mobile agent interactions, potentially streamlining applications that rely on visual and language inputs, which is crucial for real-world deployment.
MIRAGE enhances mobile agent reasoning by compressing thought processes into latent representations.
Summary
The paper presents MIRAGE, a framework that learns continuous latent reasoning representations to enable mobile agents to reason internally without relying on long textual chains of thought, thereby improving efficiency and reducing token generation.
Key contributions
- Introduction of a framework that learns continuous latent reasoning representations.
- Demonstration of improved execution efficiency and reduced token generation in mobile agents.
- Validation of the approach through performance metrics on AndroidWorld and AndroidControl benchmarks.
Notable insights
- The alignment of latent reasoning vectors with future screenshots suggests a novel approach to anticipating interface changes, which could enhance user experience.
- The reduction in token generation while maintaining performance indicates a significant advancement in resource efficiency for mobile agents.
Possible limitations
- Not stated in the abstract.
Abstract
arXiv:2606.04627v1 Announce Type: new 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.