Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MIRAGE: Mobile Agents with Implicit Reasoning and Generative World Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MIRAGE, a mobile GUI agent that replaces visible textual reasoning with continuous latent slots refined by parallel Jacobi-style rounds (APLR) and regularized by a Q-Former head predicting next-screenshot features.
Problem
Mobile GUI agents externalize reasoning as long textual thoughts, which slows decoding, increases supervision cost, and reduces deployment efficiency.
Approach
Internalize explicit thoughts into continuous latent slots trained via a two-stage curriculum; approximate serial latent chain-of-thought with parallel Jacobi-style refinement (APLR); and align latent states with stop-gradient features of the next screenshot through a Q-Former world-model head, jointly with action cross-entropy.
Key results
- On AndroidWorld, MIRAGE-8B reaches 57.8 SR vs 47.6 for Qwen3-VL-8B-Instruct while averaging 27.0 tokens vs 108.0.
- MIRAGE-4B matches explicit CoT SFT (52.6 SR) on AndroidWorld with about 31 tokens per task, versus 103 for the Qwen3-VL-4B baseline.
- On AndroidControl low-level, MIRAGE-4B raises action accuracy from 75.15 to 91.09 while cutting generated tokens from 115.67 to 18.92.
Surprising findings
- Action-only SFT (no thought supervision) performs worse than the base model on AndroidWorld (31.0 vs 42.9 SR).
- The 8B MIRAGE model takes slightly more steps than its baseline on AndroidWorld, so gains are not from shorter trajectories.
Limitations
- Supervised-only training with no RL.
- World modeling operates at feature level rather than pixels or text.
- Requires next-frame supervision and privacy/action safeguards before deployment.
Experimental setup
Datasets: AMEX AndroidControl AndroidWorld
Benchmarks: AndroidControl AndroidWorld
Metrics: exact match action accuracy task success rate generated tokens average steps latency
Baselines: Qwen3-VL-4B-Instruct Qwen3-VL-8B-Instruct GPT-4o GUI-R1 UI-R1 ShowUI MAI-UI UI-Venus-Navi UI-TARS-7B-SFT Ferret-UI Lite
Builds on: Coconut BLIP-2 Q-Former Qwen3-VL
Closest papers by meaning
- K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control (ICLR 2026 · similarity 0.86)
- MobileIPL: Enhancing Mobile Agents Thinking Process via Iterative Preference Learning (ICLR 2026 · similarity 0.86)
- ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.86)
- Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation (AAAI 2026 · similarity 0.86)
- AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents (ICLR 2025 · similarity 0.85)
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.