MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models

2026-06-03

AI Summary

Existing mobile agents rely on explicit textual chains of thought, resulting in high interaction latency, substantial supervision costs, and complex deployment. This work proposes an implicit reasoning framework that compresses explicit chains of thought into compact representations within a continuous latent space, jointly trained with a generative world model to enable the latent states to support both efficient internal reasoning and prediction of future interface states. The approach uniquely unifies reasoning compression with environmental dynamics modeling, achieving on AndroidWorld the performance of fine-tuned explicit chain-of-thought methods using 3–5× fewer decoding tokens and surpassing the baseline by 10.2 points. On AndroidControl, it reduces token generation by over 75% while significantly improving action grounding.

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.

Problem

• mobile agents
• reasoning
• generative world models
• chain-of-thought
• latent reasoning

Innovation

• implicit reasoning
• generative world model
• latent representation
• mobile agents
• token efficiency