Modern mobile GUI agents can perceive the current screen and execute local actions, yet they lack persistent app‑specific navigation knowledge—where functions reside, which pages contain them, and which screens are equivalent across repeated visits. To address this, we introduce TAG‑Nav, a task‑agnostic graph navigation framework that wraps existing GUI‑agent backbones with a reusable destination memory. During offline exploration, the system distills app functions, target pages, and waypoint routes into compact records that are available before any instruction is seen. When a future task’s destination has already been covered, TAG‑Nav retrieves the target and provides route guidance without task‑specific demonstrations. This decouples mobile GUI destination navigation from full task completion: many tasks share the prefix of reaching the right page, while the final action may involve entering content, sending data, making a purchase, deleting information, or changing sensitive settings. To evaluate route finding without downstream execution confounds, we propose DestBench, a navigation‑only protocol that asks whether an agent reaches the correct pre‑action destination screen. DestBench enables controlled comparisons among no‑memory agents, random BFS, automated BFS, and clean graph memory. Experiments show that graph‑derived memory markedly improves navigation success: UI‑TARS‑1.5‑7B rises from 52.4% to 58.3%, and Qwen‑VL‑Max from 59.5% to 71.4% under the same conditions. The combination of TAG‑Nav and DestBench demonstrates that destination coverage and route cleanliness are key factors shaping the value of graph memory, making graph quality a measurable variable rather than an assumed benefit of exploration.
Review: TAG‑Nav builds a reusable target graph offline, enabling task‑agnostic navigation memory that substantially boosts cross‑task path success. DestBench offers a pure navigation benchmark, allowing researchers to quantify how graph quality impacts memory utility and opening new directions for mobile GUI automation.