Research map / Multi-agent systems and games / LLM agents
Executable Agentic Memory for GUI Agent
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 Executable Agentic Memory, a GUI logic knowledge graph built by DFS exploration and BPE-style action-group mining, with a Q-model-guided MCTS for path extraction that reaches 52.6% on AndroidWorld and cuts token cost ~6x vs GPT-4o.
Problem
GUI agents rely on step-wise LLM re-decision at every screen, which is fragile and costly in long-horizon tasks; in-context memory injection does not reliably yield executable paths.
Approach
Store GUI interaction logic as a knowledge graph of states and actions (with mined multi-step action groups), and extract executable paths by running MCTS over the graph guided by a lightweight Q-model trained with uniform-random-policy success probabilities, plus theoretical bias and sample-complexity bounds.
Key results
- EAM reaches 52.6% on AndroidWorld, 76.1% on MobileMiniWob++, and 86.1% on DroidTask, +19.6 points over UI-TARS-7B on AndroidWorld.
- Average per-step latency of 2.8 s and token cost of 8.3K versus 50.8K for GPT-4o (about 6x reduction).
- Self-training over four rounds improves success on AndroidWorld and DroidTask for 0.5B, 1.5B and 3B Q-models.
Surprising findings
- A Q-model trained on AndroidWorld transfers to DroidTask and MobileMiniWob++ and still improves over the baseline without in-environment training.
Limitations
- Assumes a relatively static UI; app updates can make the knowledge graph outdated.
Experimental setup
Datasets: AndroidWorld MobileMiniWob++ DroidTask
Benchmarks: AndroidWorld MobileMiniWob++ DroidTask
Metrics: success rate latency API token cost
Baselines: GPT-4o Qwen 2.5-VL-3B UI-TARS-2B UI-TARS-7B M3A AutoDroid-V2 AppAgentX GUI-Explorer
Builds on: Monte Carlo Tree Search Byte Pair Encoding AppAgentX GUI-Explorer UCT
Closest papers by meaning
- SE-GA: Memory-Augmented Self-Evolution for GUI Agents (ICML 2026 · similarity 0.89)
- Darwinian Memory: A Training-Free Self-Regulating Memory System for GUI Agent Evolution (ICML 2026 · similarity 0.88)
- Agent-SAMA: State-Aware Mobile Assistant (AAAI 2026 · similarity 0.87)
- MMBench-GUI: A Unified Hierarchical Evaluation Framework for Multi-Platform GUI Agents (CVPR 2026 · similarity 0.87)
- GUI-Xplore: Empowering Generalizable GUI Agents with One Exploration (CVPR 2025 · similarity 0.87)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.87)
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.