Reproduce “MapAgent: Trajectory-Constructed Memory-Augmented Planning for Mobile Task Automation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi Kong, Dianxi Shi, Guoli Yang, Zhang ke-di, Chenlin Huang, Xiaopeng Li, Songchang Jin
MapAgent enhances mobile task automation for LLM-based agents by integrating trajectory-constructed memory-augmented planning and a dual-LLM executor to mitigate knowledge bias and hallucinations. The framework achieved higher success rates on the SPA-Bench and CHOP datasets, particularly for complex and cross-application tasks, while maintaining a practical computational overhead.
Key insights
- Explicitly augmenting LLM planning with structured, context-aware memory from past trajectories effectively overcomes inherent knowledge biases and reduces hallucinations in dynamic GUI environments.
- A coarse-to-fine planning strategy, when combined with memory retrieval, enables efficient decomposition and execution of complex, multi-application tasks.
- A self-correcting dual-LLM execution framework, where one LLM makes decisions and another evaluates progress, significantly enhances the robustness and adaptability of agents in unpredictable mobile environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.