Survey on Neural Routing Solvers

A consortium of ten researchers has published a landmark survey titled 'Survey on Neural Routing Solvers' on arXiv, offering a critical examination of AI's role in solving complex vehicle routing problems (VRPs). The paper highlights that Neural Routing Solvers (NRSs), which use deep learning to learn heuristic rules from data, show strong practical potential by reducing reliance on manual, handcrafted algorithms. However, the authors' primary contribution is a new, generalization-focused evaluation pipeline designed to address the shortcomings of conventional testing methods, which has uncovered a series of significant and previously unreported research gaps.

The survey systematically reviews existing NRSs through the lens of heuristic principles, introducing a hierarchical taxonomy for the field. By benchmarking representative models with both old and new pipelines, the research exposes critical weaknesses in how these AI systems generalize to unseen or more complex routing scenarios. This work is a vital reality check for the industry, indicating that while NRSs are promising, current approaches may not be as robust as previously thought when deployed in dynamic real-world environments like last-mile delivery or fleet management. The findings set a new standard for evaluation and provide a clear framework for future research aimed at building more reliable and scalable optimization AI.