MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models
For developers of mobile agents, MIRAGE reduces inference cost and latency while maintaining or improving task performance, addressing the bottleneck of verbose chain-of-thought reasoning.
MIRAGE learns continuous latent reasoning representations from explicit textual reasoning traces, enabling mobile agents to reason internally without decoding long rationales. It matches explicit chain-of-thought performance on AndroidWorld with 3-5x fewer tokens and improves an instruction-tuned baseline by 10.2 points, while on AndroidControl it improves action grounding with over 75% fewer tokens.
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.