ElderBench - Offline: leaderboard
The offline protocol of ElderBench (Zhan et al., 2026), which tests mobile GUI agents on smartphone tasks worded the way older adults naturally ask for them, with indirect, ambiguous and under-specified requests. Agents replay 119 tasks in five Android apps against task graphs merged from human demonstrations and must reach the final state and declare it finished; success rate in percent, higher is better.
Metric: Task success rate (%; offline protocol: 119 tasks from older adults' naturally elicited instructions in 5 Android applications, replayed on task graphs merged from multiple human demonstration trajectories; success when the predicted action sequence reaches the graph's terminal state and the agent outputs finish (50-pixel click tolerance, fuzzy text match); at most 25 steps and 120 s per task). Source: arxiv.org. Saturation forecast: Around July 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 27.73 | #159 |
| 2 | Kimi K2.5 | 14.29 | #252 |
| 3 | GLM-4.6V | 11.76 | #698 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=elderbench-offline · How It Works · Data refreshed daily, snapshot 2026-11-12.