Autonomous mobile agents promise to assist older adults with smartphone operations, yet most existing GUI benchmarks rely on explicit, goal‑driven commands and overlook natural language patterns of seniors such as indirect speech, referential ambiguity, and under‑specified requests. This gap can hinder reliable deployment of agents in real‑world elder care.
To fill the gap we introduce ElderBench, the first benchmark that evaluates mobile GUI agents in authentic elderly‑oriented scenarios. It comprises 249 naturally elicited smartphone tasks collected from older users across 20 popular applications.
We conduct a systematic linguistic comparison between elderly instructions and those used in current GUI benchmarks, examining syntactic, semantic, and pragmatic differences. Elderly commands tend to contain omissions, vague references, and context‑dependent phrasing.
We then evaluate mainstream GUI agents and vision‑language models in both online and offline settings, observing a substantial drop in performance when handling elderly‑oriented instructions.
Through controlled instruction normalization, failure case analysis, and fine‑grained linguistic feature correlation, we identify senior‑specific language patterns as the primary source of agent errors.
Based on these insights we propose design recommendations: improve models' ability to resolve omissions and ambiguous references, provide interpretable interaction feedback, and enrich training data with diverse elderly language. These steps aim to build more adaptive, interpretable, and age‑inclusive GUI agents.
Review