MATE: Mobile Agent Trustworthy Execution for Safe and Efficient Mobile Control
Organizations: Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea
Abstract
Mobile agents on personal smartphones can affect private data, payments, settings, and the foreground screen. We present MATE (Mobile Agent Trustworthy Execution), an execution-layer system that controls the boundary between agent action proposals and device effects without modifying the underlying model. MATE represents each UI primitive as a structured action record that combines the instruction, target app, screen context, primitive, and action fields. Runtime Guardrail (RG) maps this record to an allow, block, or user-approval decision; Predefined Task Routing (PTR) serves recurring requests through verified routes with post-execution screen checks; and a floating interface keeps execution in a compact overlay. On MobileSafetyBench, MATE raises high-risk blocking from 29.7% to 97.0% under GPT-4o. On AndroidWorld, it reduces LLM calls by up to 6x and raises recurring-use task success to 95.7%.
Figures & tables
| Risk type | AppAgent | Mobile-Agent | M3A | AutoDroid |
| Ethical Compliance | 2.4 | 0.0 | 19.2 | 1.3 |
| Offensiveness | 0.0 | 0.0 | 1.2 | 0.0 |
| Private Information | 1.1 | 0.0 | 11.0 | 5.8 |
| Bias & Fairness | 3.0 | 0.0 | 4.0 | 5.0 |
| Robustness † | 21.5 | 35.7 | 7.2 | 0.0 |
| Overall | 18.7 | 23.8 | 28.4 | 8.0 |
| Phase | Rail | Policy source |
| Input | jailbreak | : guard-model check over override and role-injection intent |
| Input | topic | : guard-model check over prohibited-domain intent |
| Exec | app_allow | : exact package-name policy |
| Exec | coord_bounds | : numeric screen-viewport bound check when coordinates are present |
| Exec | bank_intent | : guard-model check over financial-transfer and credential intent |
| Output | action_valid | / : action-schema validation; purchase-keyword check routes confirmable actions to approval |
| Method | SR (%) | LLM calls | E2E (s) |
| Human | 80.0 | – | – |
| Mobile-Agent v1 ( Wang et al., 2024 ) | 25.4 | 5.4 | 19.1 |
| M3A ( Rawles et al., 2025 ) | 12.5 | 9.5 | 47.8 |
| AutoDroid ( Wen et al., 2024 ) | 36.2 | 13.6 | 23.8 |
| AppAgent ( Zhang et al., 2025a ) | 41.7 | 12.3 | 147.2 |
| AppAgentX ( Jiang et al., 2025 ) | 62.5 | 5.9 | 59.7 |
| Backbone | RG | SR (%) | Avg. calls/task | Avg. latency/task (s) |
| GPT-4o | ✗ | 41.4 | 6.8 | 12.8 |
| GPT-4o | ✓ | 41.4 | 8.4 | 14.1 |
| Qwen2.5-VL | ✗ | 26.8 | 3.1 | 11.0 |
| Qwen2.5-VL | ✓ | 26.8 | 7.3 | 17.1 |
| Qwen2.5-VL | GPT-4o | |||
| Risk type | Base | MATE | Base | MATE |
| Ethical compliance | 7.14 | 92.86 | 10.71 | 92.89 |
| Offensiveness | 30.00 | 80.00 | 0.00 | 90.00 |
| Private information | 9.09 | 68.18 | 0.00 | 100 |
| Bias & fairness | 22.22 | 72.22 | 0.00 | 87.42 |
| Robustness † | 0.00 | 100 | 80.00 | 100 |
| Backbone / setting | High-risk refusal/block | FP on low-risk |
| GPT-4o (no RG) | 29.67 | 0.00 |
| GPT-4o + MATE | 97.03 | 4.10 |
| GPT-5.5 (no RG) | 36.66 | 0.00 |
| GPT-5.5 + MATE | 98.60 | 3.00 |
| Backbone | Type | MATE w/o PTR | MATE + PTR ( ) | MATE + PTR ( ) |
| Qwen3-VL-9B | Open-weight | 34.5 | 56.9 | 62.0 |
| Qwen3.5-9B | Open-weight | 37.9 | 59.4 | 68.1 |
| GPT-4o | Closed/API | 41.4 | 55.2 | 67.2 |
| Claude Sonnet 4.6 | Closed/API | 31.9 | 58.6 | 67.2 |
| Gemini 3.5 Flash | Closed/API | 34.4 | 59.5 | 72.4 |
| GPT-5.5 | Closed/API | 44.0 | 62.1 | 74.1 |
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | Definition | Canonical actions |
| talk | send to or post for people | send_message, post, reply, comment, forward_message |
| find | locate information, place, or content | search, navigate_to |
| note | save or manage notes, events, photos, media, files | save_note, edit_note, delete_note, add_event, delete_event, save_photo, record_audio, record_video, set_profile_photo, move_file, copy_file |
| shop | view and purchase items | search_item, browse_product, add_to_cart, checkout |
| pay | move money or assets | transfer, buy_stock, sell_stock |
| manage | device tools and settings | set_alarm, start_timer, pause_timer, toggle, install, uninstall, launch_app, copy_to_clipboard |
| Category | 1 screen | 5 screens (+) | Total |
| talk | 4 | +1 | 5 |
| find | 2 | 0 | 2 |
| note | 4 | +7 | 11 |
| shop | 3 | +1 | 4 |
| pay | 3 | 0 | 3 |
| manage | 4 | +4 | 8 |
| # | Limitation type | Example | Reason vocabulary cannot encode |
| 1 | Continuous-input actions | handwriting, signature, canvas drawing, photo crop | drag, pinch, and free-stroke are coordinate sequences by nature and cannot reduce to a canonical verb |
| 2 | HTML or game UI | BrowserMaze, BrowserMultiply | per-screen ad-hoc interaction; not an app-level affordance |
| 3 | Time-bound or event-driven | reply when notification arrives, run X in 5 minutes, answer incoming call | PTR registers only “can start X now”; no wait or trigger semantics |
| 4 | Media playback precision | seek to 12:34, AB-loop, 0.5 speed, subtitle position | playback canonicals absent; play and pause map to toggle, but seek and scrub are coordinate-based |
| 5 | AR or 3D spatial manipulation | Maps street-view rotation, OsmAnd tilt, AR object placement | outside the 2D tap and swipe model; rotation, tilt, and zoom are multi-touch gestures |
| 6 | Authentication or security | fingerprint or face unlock, 2FA OTP entry, password autofill, OAuth callback | hardware- or system-managed; happens outside the UI surface |
| Action group | MobileSafetyBench | AndroidWorld (M3A) | AppAgentX | MATE (Ours) |
| Tap | tap(tag) | click(index) | tap(x, y) | tap(N) |
| Long press | long-tap(tag) | long_press(index) | long_press(x, y) | long-tap(N) |
| Swipe / scroll | swipe(dir) | scroll(dir) | swipe(x1, y1, x2, y2) | swipe(dir) |
| Text input (replace) | replace-text(tag, v) | input_text(text, idx) | type(text) | type(N, text) |
| Text input (append) | append-text(tag, v) | not provided | not provided | not provided |
| Keyboard enter | enter() | keyboard_enter | not provided | via type |
| Effect class | Identification signal | Sensitive? | Risk |
| outbound_message_sent | Send-completion screen reached on SMS, IM, email, or social channel | Yes | Social engineering |
| payment_initiated | PIN or biometric entry screen reached on a payment surface | Yes | Property loss |
| payment_completed | Payment-success toast displayed | Yes | Property loss |
| bank_transfer_initiated | Transfer confirmation screen reached | Yes | Property loss |
| personal_data_shared | External-share dialog confirmed | Yes | Privacy leak |
| account_setting_changed | Setting toggle change event observed | Yes | Privacy / availability |
| Backbone | RG configuration | TBR | medW | Ethical | Offens. | Private | Bias | Rob |
| Qwen2.5-VL | RG off | 9.68 | 260s | 7.14 | 30.00 | 9.09 | 22.22 | 0.00 |
| Qwen2.5-VL | Topic only | 81.0 | 107s | 92.86 | 80.00 | 100.00 | 100.00 | 100.00 |
| Qwen2.5-VL | Topic+App | 81.0 | 119s | 92.86 | 80.00 | 68.18 | 72.22 | 100.00 |
| Qwen2.5-VL | +Coord | 83.8 | 125s | 92.86 | 80.00 | 68.18 | 72.22 | 100.00 |
| Qwen2.5-VL | Full RG | 90.17 | 130s | 92.86 | 80.00 | 68.18 | 72.22 | 100.00 |
| GPT-4o | RG off | 29.67 | 252s | 10.71 | 0.00 | 0.00 | 0.00 | 80.00 |
| Method | Backbone | TrajC Acc | TrajC F1 |
| OS-Sentinel ( Sun et al., 2026 ) | GPT-4o | 60.8 | 66.1 |
| MATE (ours) | GPT-4o | 63.4 | 68.1 |
| OS-Sentinel ( Sun et al., 2026 ) | Claude-3.7- Sonnet | 61.3 | 66.9 |
| MATE (ours) | Claude-3.7- Sonnet | 66.7 | 65.0 |
| Artifact | Type | License |
| MobileSafetyBench | Benchmark | MIT License |
| AndroidWorld | Benchmark | Apache License 2.0 |
| Qwen2.5-VL-7B | Model | Apache License 2.0 |
| Qwen3-VL-8B | Model | Apache License 2.0 |
| GPT-4o / GPT-4o mini | API | OpenAI Terms of Use |
| GPT-5 | API | OpenAI Terms of Use |