Energy-Efficient Small Language Model Inference for Mobile Agents via DVFS

Mobile agents powered by Small Language Models (SLMs) are increasingly deployed on edge devices to enhance privacy and minimize latency. However, the large computational workload of SLMs severely drains batteries and induces thermal throttling, rendering long-running tasks impractical. Existing Dynamic Voltage and Frequency Scaling (DVFS) schemes fail to accommodate the unique characteristics of SLMs due to the unknown relationship between processor frequencies, power and latency, as well as the hardware opacity that masks individual processor contributions. To address these, DVFSLM introduces workload-aware power and latency estimators that analyze core matrix operations and correlate them with hardware metadata, enabling precise estimations of how frequency adjustments affect power and latency. These estimations drive a runtime DVFS governor that dynamically adjusts frequencies across processors, minimizing energy use while meeting required token generation deadlines. Extensive experiments on a rich set of SLMs show that DVFSLM outperforms the state-of-the-art methods.

© 2026 Copyright held by the owner/author(s).