PowerLens: Taming LLM Agents for Safe and Personalized Mobile Power Management

Xingyu Feng; Chang Sun; Yuzhu Wang; Zhangbing Zhou; Chengwen Luo; Zhuangzhuang Chen; Xiaomin Ouyang; Huanqi Yang

cs.AIMar 20, 2026

PowerLens: Taming LLM Agents for Safe and Personalized Mobile Power Management

Authors: Xingyu Feng, Chang Sun, Yuzhu Wang, Zhangbing Zhou, Chengwen Luo, Zhuangzhuang Chen, Xiaomin Ouyang, Huanqi Yang

Organizations: China University of Geosciences (Beijing) · Shenzhen University · Hong Kong University of Science and Technology · City University of Hong Kong

Abstract

Battery life remains a critical challenge for mobile devices, yet existing power management mechanisms rely on static rules or coarse-grained heuristics that ignore user activities and personal preferences. We present PowerLens, a system that tames the reasoning power of Large Language Models (LLMs) for safe and personalized mobile power management on Android devices. The key idea is that LLMs' commonsense reasoning can bridge the semantic gap between user activities and system parameters, enabling zero-shot, context-aware policy generation that adapts to individual preferences through implicit feedback. PowerLens employs a multi-agent architecture that recognizes user context from UI semantics and generates holistic power policies across 18 device parameters. A PDL-based constraint framework verifies every action before execution, while a two-tier memory system learns individualized preferences from implicit user overrides through confidence-based distillation, requiring no explicit configuration and converging within 3--5 days. Extensive experiments on a rooted Android device show that PowerLens achieves 81.7% action accuracy and 38.8% energy saving over stock Android, outperforming rule-based and LLM-based baselines, with high user satisfaction, fast preference convergence, and strong safety guarantees, with the system itself consuming only 0.5% of daily battery capacity.

Figures & tables

Figure 1. Traditional power saver vs. PowerLens. Global rules degrade navigation by throttling GPS and dimming brightness; PowerLens preserves critical resources and learned preferences.

Panel of Figure 2 — Figure 2. Preliminary studies on challenges of mobile power management.

Category	Power ¶	Parameter	HW Component	Range	Default †	Control API	Granularity	Root	Latency	$\|\mathcal{S}\|$	Impact ‡
Display	42%	Brightness	OLED panel	0–4096	Auto	Settings.System	Continuous	—	< 1 s	4,097	H
		Refresh rate	Display controller	{30, 60, 90, 120} Hz	120 Hz	Settings.System	Discrete	—	< 1 s	4	H
		Screen timeout	Display controller	15 s–30 min	30 s	Settings.System	Discrete	—	Deferred	7	M
		Dark mode	OLED panel	On/Off	Off	UiModeManager	Binary	—	< 1 s	2	L
		Auto-rotation	Gyroscope/Accel.	On/Off	On	Settings.System	Binary	—	< 1 s	2	L
Connectivity	25%	Wi-Fi	Wi-Fi radio (WCN7851)	On/Off	On	Settings.Global	Binary	—	1–3 s	2	M

Table 1. Parameters controlled by PowerLens. 18 parameters across 5 categories, spanning a joint state space of

>

10^{17}

configurations.

Figure 3. Hardware activated during navigation. A single app engages components across the mainboard and daughter board, spanning all five parameter categories in Table 1 .

Figure 4. PowerLens system overview. Each cycle: ➊ Accessibility captures UI tree, ➋ Activity Agent recognizes context, ➌ Policy Agent generates strategy, ➍➎ Execution Agent verifies and applies via shell commands, ➏ Feedback Agent detects user overrides. The Memory System stores preferences for personalization.

Figure 5. Decision cycle example. The Policy Agent generates a structured JSON policy for an Instagram browsing session; the Execution Agent validates actions against PDL constraints and translates approved actions into root shell commands.

Figure 6. A Two-tier memory system. STM maintains session-scoped state; the Extractor distills observations into LPM with confidence-based rule promotion.

Figure 7. Personalization journey. Repeated user overrides accumulate confidence ( $c{=}0.50\to 0.665\to 0.818$ ) until the rule is promoted to LPM.

Figure 8. PowerLensBench overview. 7 app categories with representative screenshots.

Figure 9. Energy saving across app categories (relative to Stock Android, higher is better).

Method	Acc.	ES	Viol.	UES
Battery Saver	48.3	4.6	0.8	3.6
Rule-Based	63.5	19.9	1.2	3.4
Single-Agent LLM	52.1	48.4	12.5	2.5
PowerLens	81.7	38.8	0.6	4.3

Table 2. Comparison with baselines (all relative to Stock Android, ES = 0%). Acc.: Action Accuracy (%), ES: Energy Saving (%), Viol.: Violation Rate (%), UES: User Experience Score (1–5).

Panel of Figure 10 — Figure 10. Preference learning and adaptation.

Panel of Figure 11 — Figure 11. Memory effect.

Battery	Nav.	Video	Meet.	Social	Music	Feed	Read.	Avg.
High (80%)	23.6	26.4	18.7	38.2	21.5	42.3	48.1	31.3
Mid (45%)	27.4	30.8	22.3	45.7	25.1	49.6	56.8	36.8
Low (15%)	30.8	35.3	26.5	50.6	28.9	53.2	60.4	40.8

Table 3. Context sensitivity: energy saving (%) by battery level.

Figure 12. Activity recognition accuracy.

Panel of Figure 13 — Figure 13. Safety constraint verification.

Panel of Figure 14 — Figure 14. Ablation study.

Agent	Input Tok.	Output Tok.	Latency (s)	Cost ($)
Activity	2,049	162	1.9	0.0010
Policy	2,226	712	5.9	0.0025
Exec.-Verify	2,518	76	2.4	0.0009
Exec.-Action	614	174	2.0	0.0006
Total	7,407	1,124	12.2 †	0.0050
† Sequential; pricing: $0.30/M input,$ 2.50/M output (Gemini-2.5-Flash).

Table 4. System overhead per decision cycle.

Model	In Tok.	Out Tok.	Latency (s)	Cost ($)	Qual.
Gemini-2.5-Flash ⋆	7,407	1,124	12.2	0.0050	23
GLM-4.7	6,258	672	17.1	0.0049	22
GPT-4o-mini	6,253	1,003	20.8	0.0015	9 ∗
GPT-4o	6,650	1,343	39.0	0.0301	24
GLM-5	6,288	1,052	42.7	0.0066	20
MiniMax-M2.5	6,201	2,713	47.8	0.0049	20

Table 5. LLM backbone comparison. Frozen-snapshot evaluation (IDLE, 100% battery). Quality: 5 dimensions

\times

5 pts

=

25.

Panel of Figure 15 — Figure 15. User study results.

Language Model-Based Control

Language model-based control uses language models to interpret goals, choose actions, and steer AI systems or physical processes.

See all 180 papers

PALS: Power-Aware LLM Serving for Mixture-of-Experts Models
May 20, 2026Can Hankendi, Rana Shahout, Minlan Yu +1LLM Inference Efficiency LLM Serving
Boston University · Electrical and Computer Engineering Department · Boston, MA, USA +3
Fine-Tuning a 3B-Parameter LLM on a Smartphone: Characterizing Sustained Training
Oct 5, 2026Andrew Geyko, Marius Mosbach, André BrinkmannOn-Device Training LLM Fine-Tuning
Saarland University Saarbrücken, Germany
Sustainability Is Not Linear: Quantifying Performance, Energy, and Privacy Trade-offs in On-Device Intelligence
Mar 27, 2026Eziyo Ehsani, Luca Giamattei, Ivano Malavolta +1On-Device Language Model Inference Energy-Efficient ML
∗University of Naples Federico II, Naples, Italy · †Vrije Universiteit Amsterdam, Amsterdam, The Netherlands

May 20, 2026J/K move · Enter open · S save

PALS: Power-Aware LLM Serving for Mixture-of-Experts Models

Can Hankendi, Rana Shahout, Minlan Yu, Ayse K. Coskun

Boston University · Electrical and Computer Engineering Department · Boston, MA, USA+3

LLM Inference Efficiency LLM Serving Energy-Efficient ML

Open paper AskPDF arXiv

Large language model (LLM) inference has become a dominant workload in modern data centers, driving significant GPU utilization and energy consumption. While prior systems optimize throughput and latency by batching, scheduling, and parallelism, they largely treat GPU power as a static constraint rather than a controllable resource. In this paper, we present a power-aware runtime for LLM serving, PALS, that treats GPU power caps as a first-class control knob and jointly optimizes them with software parameters such as batch size. The system combines lightweight offline power-performance models with a feedback-driven controller to select configurations that satisfy throughput targets while maximizing energy efficiency. We implement PALS within an existing LLM serving framework, vLLM, demonstrating that it requires no model retraining or API changes. Across multi-GPU systems and both dense and mixture-of-experts (MoE) models, PALS improves energy efficiency by up to 26.3%, reduces QoS violations by 4x to 7x under power constraints, and tracks dynamic power budgets. These results highlight the potential of integrating power control directly into LLM inference runtimes, enabling energy-proportional and grid-interactive AI systems.

Abstract

Language Model-Based Control

Explore similar work

PALS: Power-Aware LLM Serving for Mixture-of-Experts Models

Fine-Tuning a 3B-Parameter LLM on a Smartphone: Characterizing Sustained Training

Sustainability Is Not Linear: Quantifying Performance, Energy, and Privacy Trade-offs in On-Device Intelligence