Neural solvers for vehicle-routing now match classical optimization on cost — but tell dispatchers nothing about why they chose a particular route.
Researchers dissected six variants of a neural solver for the Multi-Attribute Vehicle Routing Problem, probing both the encoder (the part that reads the map and constraints) and the decoder (the part that picks each stop). They tested three attribution methods against five quality metrics — fidelity, concentration, stability, sanity, and actionability — across encoder and decoder combinations. The headline result: models trained with a Recourse decoder produced explanations that could tell a dispatcher the smallest change needed to make an infeasible route work. Models trained with a Hard-Mask decoder failed to generate those fixes even when researchers handed them infeasible alternatives directly.
That distinction matters because explainability in routing isn't just academic — logistics operators increasingly face regulatory or contractual pressure to justify automated dispatch decisions. A system that can say "add five minutes of slack to stop three and the route becomes legal" is categorically more useful than one that silently rejects a plan. The finding also reframes where engineers should focus: swapping in a fancier encoder architecture is less important than choosing the right training regime.
The Mixture-of-Experts encoder turned up its own wrinkle — it spreads constraint knowledge across many components rather than storing it in tidy, interpretable slots, which is the kind of result that looks good on a benchmark and quietly complicates any audit.