Interruptible and Time-Aware Cognitive Loops for LLM Agents in Dynamic Environments

Abstract

Large Language Models are increasingly deployed as autonomous agents that must perceive, reason, and act within environments that evolve during inference. Standard turn-based pipelines treat reasoning as a blocking procedure, which encodes a frozen- world assumption and leads to brittle behavior under mid-episode updates. This pro- posal develops a principled framework for interruptible, intra-turn metacognition. We formalize the generation process as a stochastic dynamical system under exogenous control and introduce an Interruptible Cognitive Loop (ICL) that operates on seman- tic Cognitive Blocks rather than raw tokens. A scheduler observes time budgets and external events, estimates hazard, and decides when to continue, soft-pause, abort, or summarize. We provide Bayesian resumption methods to maintain semantic coherence after interrupts, a block-wise streaming runtime to create safe preemption points, and an evaluation protocol that perturbs static benchmarks with temporal and interactive dynamics. The research yields theory, algorithms, and an open-source runtime kernel, together with new benchmarks for robustness, responsiveness, and reasoning quality under interruption.

Keywords