Managing railway disruptions is a complex multi-agent routing problem where a single train failure can propagate delays across the network. Traditional approaches rely on heuristic optimization solvers, which are effective but assume access to a global system view and require substantial expert design, limiting their applicability and generalization. Reinforcement learning (RL) offers an alternative by learning adaptive strategies from interactions with the environment. In this paper, we show that none of the available paradigms is sufficient in isolation: (i) heuristic solvers encode valuable global expertise but cannot be deployed directly, (ii) world models improve sample efficiency but struggle to leverage expert knowledge, and (iii) pure RL can adapt policies but often lacks stability without strong guidance. We propose a hybrid framework that integrates the strengths of these approaches. First, imitation learning transfers knowledge from a global expert solver to initialize a neural policy. Then, model-based RL fine-tunes this policy using the DreamerV2 world model to enhance generalization and responsiveness to local perturbations. Our method builds on the Multi-Agent Model-Based Architecture (MAMBA) to model agent interactions and addresses the challenge of transferring expertise from global solvers to decentralized agents operating on local latent observations. Experiments on a train rescheduling problem using the Flatland environment show that our method outperforms MAMBA, improving performance by up to 23% on difficult instances. This highlights the benefit of combining imitation learning with world-model-based multi-agent RL for complex transportation networks. The code is available here https://github.com/corail-research/Imitation-Guided_World_Models .