A new pretraining trick helps AI routing solvers scale to problems far larger than anything in their training set.
Researchers built a self-supervised pretraining step for neural combinatorial optimization models, the kind of AI used to plan efficient routes through networks of points. Instead of training only on labeled routing solutions, the model first learns from rotated and mirror-flipped versions of the same maps, picking up patterns in spatial layout and relative distance before it ever tackles a real routing task. That geometric pretraining step is the paper's central idea, adapting a trick from computer vision and language models to a domain that has far less structure to exploit than pixels or text. On the classic Traveling Salesman Problem scaled up to 1,000 cities, the pretrained model cut tour lengths by 7.23 percent compared with models trained from scratch, and it ran up to two orders of magnitude faster than the exact solver Concorde.
That speed and generalization gap matters because routing problems show up everywhere: delivery networks, chip layout, wireless mesh design. AI solvers have struggled to hold up once problems get bigger than their training data, which limits their usefulness beyond toy benchmarks. Pretraining on geometric transformations, rather than more labeled data, is a cheap way to close that gap.
Still, this is a benchmark win on synthetic point sets, not a delivery company's actual road network. Whether the same trick holds up on messier, real-world routing graphs is the next test, and the code is public for anyone who wants to try.