Two Black Boxes, One Solver: Encoder Probing and Decoder Attribution for Neural Multi-Attribute VRP under Hard-Mask and Recourse Decoders

Neural routing solvers lack transparency: The paper demonstrates that while neural autoregressive models for vehicle routing achieve competitive performance, they provide "no per-step justification," creating a critical gap when human dispatchers need to validate or compare routing decisions in real-world applications.

Graph structure improves both performance and explainability: Models with stronger graph inductive bias (architectural design that respects graph structure) showed improvements in both how predictably their internal representations encode routing constraints and in the sanity of their explanations—suggesting that better AI design can simultaneously enhance accuracy and interpretability.

Different training approaches produce fundamentally different decision-making patterns: The "Recourse" training regime produced policies that better represent infeasibility and generate actionable "make-feasible" explanations (counterfactuals showing minimal input changes to restore feasibility), whereas the "Hard-Mask" approach failed to produce these useful explanations even when given infeasible routing options—indicating that training methodology, not just architectural softness, shapes explainability.

Multiple explanation methods reveal different insights: Using three complementary attribution approaches (gradient-based, integrated gradients, and DeepLIFT) with three different reading angles (abductive, contrastive, and counterfactual) uncovered different aspects of model reasoning, suggesting that responsible AI evaluation requires multiple explanation techniques rather than relying on a single method.

Explainability can be systematically measured: The paper scored explanations across five dimensions (fidelity, concentration, stability, sanity, and actionability), providing a framework for practitioners to rigorously evaluate whether AI explanations are actually useful and trustworthy for human decision-making.