Xrbit: Toward Breaking the Thermodynamic Wall of Persistent LLMs in Mobile XR

초록

Deploying persistent Large Language Models (LLMs) on standalone XR devices is fundamentally constrained by the 'thermo-dynamic wall,' where on-device processing triggers thermal saturation and cybersickness. While cloud APIs offer an alternative, they impose prohibitive recurring costs for continuous conversational agents and raise privacy concerns regarding sensitive data. To address these barriers without the stochastic latency of cloud offloading, we propose Xrbit, a framework that repurposes a tethered smartphone as a Personal Edge Node. Xrbit employs a Contextual Bandit (LinUCB) offloading agent to dynamically switch between Wi-Fi and BLE, optimizing the trade-off between latency and thermal accumulation in real-time. Evaluations on Meta Quest 3 and Galaxy Z Fold 7 demonstrate that Xrbit sustains stable frame rates (>70 Hz) for 99.85% of the session, effectively minimizing thermal-induced jitter compared to on-device baselines. This work validates that intelligent, thermal-aware edge offloading can enable sustainable, cost-effective, and privacy-preserving always-on AI for mobile XR. © 2026 IEEE.

키워드