Research map / Multi-agent systems and games / LLM agents
Agent-SAMA: State-Aware Mobile Assistant
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.
Agent-SAMA is a multi-agent mobile GUI framework that models app execution as a finite state machine built online, using it for planning, error recovery, and cross-task memory.
Problem
Existing mobile GUI agents are reactive: they reason over the current screen without a structured model of app navigation flow, which limits context understanding, detection of unexpected outcomes, and error recovery.
Approach
Model each app as an FSM whose states are UI screens (described in natural language with pre/post-conditions) and whose transitions are user actions; construct it incrementally during execution, use it to verify transitions and roll back to stable states, and store FSMs in long-term memory via a Mentor agent.
Key results
- On Mobile-Eval-E, Agent-SAMA reaches 84.0% success rate and 71.9% recovery success, versus 72.0% and 67.34% for Mobile-Agent-E+Evo.
- On SPA-Bench, Agent-SAMA reaches 80.0% success rate and 66.7% recovery success, versus 75.0% and 52.86% for the baseline.
- On AndroidWorld, Agent-SAMA achieves 63.7% success rate, above V-Droid (59.5%) and AgentS2 (54.3%).
Surprising findings
- Removing the planner agent caused the largest drop (SR from 84% to 52% on Mobile-Eval-E), larger than removing multi-plan selection or the Mentor.
Limitations
- Evaluation limited to Mobile-Eval-E, SPA-Bench and AndroidWorld, which may not cover dynamic content or interruptions.
- Baseline Mobile-Agent-E+Evo re-run results differ from original paper; variance handled by averaging five runs.
- Cross-app benchmark success relies on human evaluation with rubrics, and small task counts (25 and 20 tasks) limit statistical power.
Experimental setup
Datasets: Mobile-Eval-E SPA-Bench
Benchmarks: Mobile-Eval-E SPA-Bench AndroidWorld
Metrics: success rate satisfaction score action accuracy termination rate recovery success rate
Baselines: AppAgent Mobile-Agent Mobile-Agent-v2 Mobile-Agent-E+Evo AgentS2 V-Droid UI-TARS GPT-4o+UGround GPT-4o+Aria-UI
Builds on: Mobile-Agent-E Evo (long-term memory) GroundingDINO DBNet
Closest papers by meaning
- VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics (ICML 2026 · similarity 0.90)
- FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents (ICLR 2026 · similarity 0.90)
- D-GARA: A Dynamic Benchmarking Framework for GUI Agent Robustness in Real-World Anomalies (AAAI 2026 · similarity 0.89)
- ProBench: Benchmarking GUI Agents with Accurate Process Information (AAAI 2026 · similarity 0.89)
- AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents (ICLR 2025 · similarity 0.89)
- MMBench-GUI: A Unified Hierarchical Evaluation Framework for Multi-Platform GUI Agents (CVPR 2026 · similarity 0.89)
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.