LLM-Guided Heuristic Design from Simulation Traces for Dynamic Production and AGV Scheduling

A recent paper published on arXiv (ID 2608.09343) presents a heuristic design framework guided by LLMs that utilizes simulation traces to enhance scheduling for dynamic production and automated guided vehicles (AGVs). Unlike traditional simulation-based optimization (SBO), which views simulators as black boxes and relies on aggregate scores for candidate ranking without clarifying failures or recommending policy adjustments, this new framework employs iterative simulations for selection and event-level traces for diagnosis. A manager agent generates bottleneck hypotheses from the lowest-scoring simulation replay, while editing agents carry out parallel code-level modifications. Following execution checks and repeated assessments, only the best improvements are retained. This framework is tested within a discrete-event simulation for AGV scheduling and dynamic production, aiming to increase transparency and efficiency in SBO through detailed trace analysis.