GeoPAR Optimizes Large-Scale Multi-Agent Combinatorial Problems

Researchers have developed GeoPAR, a geometry-guided parallel autoregressive reinforcement learning framework designed to tackle large-scale multi-agent combinatorial optimization problems. These problems are notoriously challenging due to their NP-hard nature and the performance degradation of existing parallel solvers on larger instances. GeoPAR integrates three core components: a projection-window sparse geometry mechanism to create lightweight local candidate neighborhoods, sparse edge-biased attention to embed these geometric relations into node representations, and a cache-guided conflict-aware assignment process that reuses geometric information to prevent duplicate task selections. Experiments on complex problems like heterogeneous vehicle routing and multi-depot pickup-and-delivery demonstrate that GeoPAR significantly improves large-scale zero-shot generalization, reduces rollout steps, and maintains efficient inference.