RELAGRAM: Recursive Latent Graph Model for the Capacitated Vehicle Routing Problem

TL;DR

RELAGRAM introduces a recursive latent graph model that uses latent recursion via a compact transformer to solve capacitated vehicle routing problems. It processes edge-based representations iteratively, producing a solution through an autoregressive decoder. The model achieves strong performance on CVRP-100 and generalizes zero-shot to larger instances without retraining.

Why It Matters

This work demonstrates that latent recursion can be effectively applied to complex combinatorial optimization problems like vehicle routing, offering a scalable and generalizable neural approach. It advances the use of recurrent latent representations in structured prediction tasks, potentially reducing reliance on handcrafted heuristics.

Key Findings

• RELAGRAM uses a recurrent transformer to update latent edge representations in a graph-based CVRP formulation.
• The model achieves 13.4% gap to a reference solver on CVRP-100 under greedy decoding, improving to 8.3% with 256-candidate search.
• It generalizes zero-shot to CVRP-500 and CVRP-1000 with gaps of 15.6% and 21.7%, respectively, under 256-candidate search.