Huawei's MIRAGE Teaches Phone AI to Think Silently, 75% Fewer Tokens
MIRAGE teaches mobile AI agents to reason silently in compressed hidden states, cutting token output by 75% while matching or beating chain-of-thought baselines on real Android tasks
Takeaways
- MIRAGE is a new framework for mobile AI agents that replaces verbose chain-of-thought text with silent latent reasoning vectors, cutting token output by 75%+.
- A generative world model objective aligns those latent vectors with future screenshots, teaching the agent to anticipate upcoming UI states before acting.
- On AndroidWorld, MIRAGE matches explicit CoT fine-tuning with 3-5x fewer decoded tokens and beats an instruction-tuned baseline by 10.2 points.
- On AndroidControl, it improves action grounding accuracy while generating over 75% fewer tokens than CoT baselines.
- Inference latency is 1.81 seconds per step for the 4B model, competitive with or faster than larger 7-8B models.
- No code or weights released yet; interpretability is reduced since reasoning lives in hidden states rather than readable text.
Getting an AI agent to reliably tap through your phone apps sounds simple until you watch it think out loud. Today's best mobile agents narrate every step in long chains of text before acting -- reasoning about what's on screen, what might happen next, and what to do about it. That verbosity has a real cost: slower responses, higher compute bills, and reasoning traces that are hard to supervise at scale. A new paper from researchers at Huawei and collaborating institutions proposes a different path: teach the agent to reason silently, in compressed internal representations, without ever writing a word.
The framework is called MIRAGE (Mobile Agents with Implicit Reasoning and Generative World Models). Its core idea is to take the explicit chain-of-thought (CoT) traces that a capable teacher model produces -- step-by-step rationales like "the settings icon is in the top right, I should tap it" -- and distill them into continuous latent vectors inside the model's hidden states. At inference time, the agent reasons in that compressed space instead of generating tokens, then acts directly.
Why verbose reasoning is a real bottleneck
Chain-of-thought reasoning became the dominant paradigm for complex AI tasks after OpenAI's o1 series showed that "think first, then answer" dramatically improves accuracy on hard problems. These models follow a "think first, then answer" approach, generating detailed CoT for self-reflection and reasoning before producing a final answer. For mobile GUI agents, multimodal reasoning is considered essential, as agents encounter unseen and complex tasks within mobile apps, with complexities increasing in dynamic contexts such as frequently changing app content -- allocating time for reasoning allows agents to adapt to environmental changes, correct mistakes, and determine the optimal path for completing a task.
But reasoning tokens are expensive. Each step in a multi-turn phone task requires the model to decode hundreds of tokens before it can act. Multiply that across dozens of steps and many parallel users, and latency and cost compound quickly. The MIRAGE paper frames this as a deployment problem: long rationales slow interaction, increase supervision cost, and complicate real-world rollout.