Research map / Multi-agent systems and games / LLM agents
Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration
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 Mobile-Agent-v2, a training-free multi-agent framework (planning, decision, reflection agents plus a memory unit) for mobile device operation that improves task completion over single-agent Mobile-Agent.
Problem
Mobile device operation assistants built on single-agent MLLM architectures struggle to navigate task progress and focus content from long, interleaved image-text operation histories.
Approach
Split the single agent into three roles: a text-only planning agent that condenses history into task progress, a decision agent that acts and updates a memory unit of focus content, and a reflection agent that compares before/after screenshots to detect erroneous or ineffective operations.
Key results
- On the English scenario, Mobile-Agent-v2 raises advanced-instruction system-app success from 4/10 (Mobile-Agent) to 6/10, and external-app from 3/10 to 7/10.
- Ablation on 88 instructions: removing the planning agent drops advanced SR from 61.4 to 29.5, showing it is the most impactful component.
- Average success-rate improvement of about 27% over Mobile-Agent is reported across scenarios, with the abstract claiming over 30% task completion gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the Mobile-Agent-v2 multi-agent architecture instead of the single-agent Mobile-Agent | Mobile device operation assistant · Non-English and English mobile tasks | SR: 5/10 -> 9/10 (non-English system basic); 3/10 -> 6/10 (non-English system advanced); 2/10 -> 8/10 (non-English external | Helped |
| Inject manual operation knowledge | Mobile-Agent-v2 · English external apps, advanced instructions | SR: 7/10 -> 9/10 | Helped |
| Add the planning agent (ablation) | Mobile-Agent-v2 · Ablation evaluation, basic instructions | SR: 59.1 -> 88.6 | Helped |
| Add the memory unit (ablation) | Mobile-Agent-v2 · Ablation evaluation, basic instructions | SR: 77.3 -> 88.6 | Helped |
| Add the reflection agent (ablation) | Mobile-Agent-v2 · Ablation evaluation, basic instructions | SR: 86.4 -> 88.6 | Helped |
Surprising findings
- GPT-4V used directly as an end-to-end mobile operation assistant is almost infeasible (2.7 basic SR&DA), while the agent architecture makes it effective.
Limitations
- Evaluation uses only 88 instructions on real devices, with small per-app sample sizes (10 per app).
- Relies on closed-source GPT-4/GPT-4V APIs, limiting reproducibility and cost-efficiency.
- Knowledge injection results depend on manually written operation hints.
Experimental setup
Benchmarks: Mobile-Agent-v2 dynamic evaluation (88 instructions, Chinese HarmonyOS and English Android)
Metrics: success rate completion rate decision accuracy reflection accuracy
Baselines: Mobile-Agent GPT-4V w/o agent Gemini-1.5-Pro Qwen-VL-Max
Builds on: Mobile-Agent AppAgent GroundingDINO
Closest papers by meaning
- Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation (AAAI 2026 · similarity 0.89)
- ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.88)
- Agent-SAMA: State-Aware Mobile Assistant (AAAI 2026 · similarity 0.86)
- K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control (ICLR 2026 · similarity 0.86)
- SPA-BENCH: A COMPREHENSIVE BENCHMARK FOR SMARTPHONE AGENT EVALUATION (ICLR 2025 · similarity 0.86)
- Scaling, Benchmarking, and Reasoning of Vision-Language Agents for Mobile GUI Navigation (ICML 2026 · similarity 0.86)
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.