Dynamic scheduling must dispatch jobs that arrive online while anticipating future arrivals. Poor decisions can amplify downstream costs under nonlinear tardiness objectives. We show that a pretrained large language model (LLM) can serve as the exogenous component of a structured world model, forecasting uncertain future job arrivals. We exploit an afterstate decomposition that separates deterministic shop transitions from uncertain exogenous arrivals, allowing the LLM to predict only the latter via a compact belief representation. Building on this, we propose Afterstate World‑Model Planning (AWMP), which couples an evolved belief‑conditioned dispatching strategy with short scenario rollouts. A surprise‑triggered refresh queries the LLM only when newly observed arrivals deviate significantly from prior forecasts, and online residual self‑calibration further conditions the LLM on recent forecast errors to handle non‑stationary arrivals. Across diverse arrival distributions, AWMP reduces weighted tardiness by 6‑21% on dynamic job‑shop benchmarks and by 22‑46% on flexible job‑shop benchmarks relative to the best of eight classical priority‑dispatching rules.
Review