DiCon: A Differential Attention Solver for Compositional Geometry Routing Problems
Researchers propose DiCon, a differential attention-assisted solver with contrastive learning, to address the Compositional Geometry Routing Problem (CGRP). CGRP unifies point-only, line-only, area-only, and hybrid task geometries, covering real-world routing scenarios. Non-point tasks introduce asymmetry, coupling travel routes with intrinsic paths, and expand the action space with many irrelevant options, challenging representation learning and decision-making. DiCon is a plug-and-play framework that tackles these issues from two angles: a differential attention mechanism suppresses probability mass on less competitive candidate actions, and a double-layer contrastive learning structure enhances representation. The work is published on arXiv (2605.18094).
Key facts
- CGRP is a unified superclass of traditional routing problems.
- It covers point-only, line-only, area-only, and arbitrary hybrid task geometries.
- Non-point tasks make CGRP inherently asymmetric.
- Travel routes are tightly coupled with intrinsic paths in CGRP.
- The action space is enlarged with numerous feasible yet often irrelevant options.
- DiCon uses differential attention to suppress less competitive candidate actions.
- DiCon employs double-layer contrastive learning.
- The paper is available on arXiv with ID 2605.18094.
Entities
Institutions
- arXiv