AIAny
Icon for item

MobilePA-Bench: Benchmarking Mobile Planner Agents on Complex Real-World Tasks

Assesses mobile planning agents' ability to call tools, plan long-horizon workflows, and coordinate sub-agents in realistic, interactive phone scenarios via a stateful executable sandbox. Covers 13 domains, 212 tools, evidence-based verification, and tests memory, skill usage, permission and runtime constraints.

Introduction

Most mobile-agent benchmarks focus either on surface-level GUI control or offline API matching, missing the middle ground where planning, tool selection, and runtime verification matter. MobilePA-Bench addresses that gap by pairing an interactive, stateful function-calling sandbox with evidence-based checks to evaluate whether an agent actually effected the intended side effects in realistic mobile workflows.

Key Findings
  • Benchmark design: Runs live app databases and returns structured feedback across 13 functional domains and 212 realistic mobile tools, enabling deterministic verification of task effects rather than heuristic screen-score judgments. This makes success measurable by side effects (database entries, tool outputs) rather than only visual cues.
  • Evaluation dimensions: Moves beyond single-action control to measure Sub-agent Collaboration (task decomposition and delegation), Memory Usage (recalling stored preferences and histories), and Skill Usage (invoking composite pre-packaged skills vs. planning primitive steps). These axes expose planning and orchestration weaknesses not visible in GUI-only tests.
  • Empirical result: State-of-the-art LLMs suffer sharp performance drops under strict tool ordering, permission limits, and unexpected runtime errors, indicating current models are brittle when faced with realistic mobile constraints. An interactive function-calling sandbox enables more diagnostic evaluations and supports agentic reinforcement learning research.
Who it's for and trade-offs

Great fit if you are a researcher or practitioner building mobile agents that must integrate tool use, long-horizon planning, and multi-agent decomposition, and if you need deterministic, evidence-based evaluation. Look elsewhere if your goal is purely GUI-level action prediction, lightweight synthetic evaluation, or benchmarking against proprietary closed-app ecosystems—the sandbox requires engineering to emulate app backends and may not capture every idiosyncrasy of closed commercial apps. The benchmark is diagnostic and research-oriented rather than a turnkey production validation for every commercial app.

Information

  • Websitearxiv.org
  • AuthorsYi Zhu, Xiongwei Wu, Qiyi Wang, Tingyu Qu, Jiajun Liu, Sihan Cao, Long Chen, Weigao Sun, Feida Zhu, Yiran Zhong
  • Published date2026/08/24

Categories

More Items

Develops methods to scale agentic AI for sustained, verifiable execution of complex long-horizon work by expanding executable environments and training coordinated agents with a shared execution harness (AgentOS) to maintain state, provenance, and failure recovery.

Proposes “Graph Engineering”: using explicit, dynamic graphs to represent tasks, agents, tools, and system state so LLM-based agent systems can coordinate, persist, and evolve. Surveys principles, methods, applications, and curates related resources.

Evaluates whether coding agents can modify real scientific software while preserving domain-specific scientific contracts. Contains 119 repository-level tasks across 98 GitHub projects and 20 scientific domains, measures reproducible edits in pinned Docker images, and analyzes recurring failure modes.