Research map / Multi-agent systems and games / LLM agents

Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 2 citations (OpenAlex)

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-RAG, a hierarchical multi-agent mobile automation framework with Manager-RAG (human-validated task plans) and Operator-RAG (app-specific subtask-screenshot-action retrieval), plus the Mobile-Eval-RAG benchmark of 50 multi-app tasks.

Problem

Mobile GUI agents achieve low success on long-horizon, cross-application tasks because they rely on static MLLM knowledge, causing strategic hallucinations in planning and operational errors in UI execution.

Approach

Separate retrieval knowledge for the two levels: a Manager-RAG that retrieves human-validated task plans for high-level planning, and an Operator-RAG that retrieves app-specific (subtask, screenshot, action) examples for atomic action generation, layered onto the Mobile-Agent-E hierarchy.

Key results

• Mobile-Agent-RAG with Gemini-1.5-Pro reaches 75.7% completion rate and 76.0% success rate on Mobile-Eval-RAG, versus 58.3% CR and 48.0% SR for Mobile-Agent-E.
• Reported gain of 11.0% task completion rate and 10.2% step efficiency over SoTA baselines.
• With GPT-4o and Claude-3.5-Sonnet, success rate reaches 84.0% versus 60.0% for Mobile-Agent-E.

Surprising findings

• The RAG benefit is reported to be negatively correlated with base model strength, with larger gains for weaker MLLMs.

Limitations

• Benchmark has only 50 tasks across a limited set of apps (about 7 in complex subset)
• Knowledge bases are manually curated and human-validated, limiting scalability
• Operator-RAG retrieves top-1 example, so gains depend on coverage of app-specific trajectories

Experimental setup

Datasets: Mobile-Eval-RAG Mobile-Eval-E

Benchmarks: Mobile-Eval-RAG

Metrics: Success Rate Completion Rate Operator Accuracy Reflector Accuracy Steps Efficiency

Baselines: AutoDroid AppAgent (Auto) AppAgent (Demo) Mobile-Agent Mobile-Agent-v2 Mobile-Agent-E Mobile-Agent-E+Evo

Builds on: Mobile-Agent-E AppAgent Mobile-Agent-v2 Contriever

Closest papers by meaning

• Mobile-Agent-v2: Mobile Device Operation Assistant with Effective Navigation via Multi-Agent Collaboration (NeurIPS 2024 · similarity 0.89)
• Agent-SAMA: State-Aware Mobile Assistant (AAAI 2026 · similarity 0.88)
• K²-Agent: Co-Evolving Know-What and Know-How for Hierarchical Mobile Device Control (ICLR 2026 · similarity 0.88)
• Agent S: An Open Agentic Framework that Uses Computers Like a Human (ICLR 2025 · similarity 0.88)
• ProactiveMobile: A Comprehensive Benchmark for Boosting Proactive Intelligence On Mobile Devices (CVPR 2026 · similarity 0.87)
• SPA-BENCH: A COMPREHENSIVE BENCHMARK FOR SMARTPHONE AGENT EVALUATION (ICLR 2025 · similarity 0.87)