Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents

This paper introduces TSDS, a framework for edge LLM agents that combines a lightweight convergence probe with perplexity-based cloud deferral. Local reasoning stops when the intended action stabilizes, while uncertain actions are escalated to a cloud model. A multi-objective Learn-Then-Test calibration procedure operates on end-to-end episode trajectories and provides finite-sample guarantees for expected episode reward and cloud-call rate. On GSM8K, HotpotQA, MBPP, and a household-robot planning task, TSDS reduces per-episode thinking compute by 43%–73% versus deferral-only baselines on HotpotQA, MBPP, and the robot task while retaining certified guarantees.