Title: MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models

Original: 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.

Rewritten: Abstract: As mobile agents become more prevalent in handling daily tasks via screenshots and natural language instructions, ensuring robust control demands sophisticated reasoning regarding screen affordances, sequential navigation, and anticipated state transitions. Traditionally, many systems externalize this cognitive load through extensive textual chain-of-thought processes. This approach not only hampers interaction speed but also escalates supervision expenses and creates deployment hurdles. To address these challenges, we present MIRAGE, a novel framework that derives continuous latent reasoning representations directly from observable textual reasoning traces. By converting explicit reasoning into compact hidden states, MIRAGE allows agents to perform internal reasoning without the need to decode lengthy rationales. Furthermore, the system integrates a generative world-model objective that aligns latent reasoning vectors with subsequent screenshots. This alignment prompts the agent to predict future interface states prior to taking action, effectively transforming hidden computation into both a compressed thought structure and a predictive model of environmental dynamics. During inference, MIRAGE operates within a continuous latent space, which significantly curtails token generation and enhances execution efficiency. In evaluations on AndroidWorld, the 4B ablation of MIRAGE achieves performance parity with explicit chain-of-thought supervised fine-tuning while utilizing a decoded-token budget that is 3 to 5 times smaller, and it boosts a comparable instruction-tuned baseline by 10.2 points. Additionally, on the AndroidControl benchmark, the model demonstrates improved action grounding while producing more than 75% fewer tokens.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC