Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MixRoute: Rethinking Single-Distribution Training for Generalizable Neural Routing
Hang Yi, Ziwei Huang, Yining Ma, Zhiguang Cao
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Challenges the assumption that single-distribution training gives weak cross-distribution generalization in neural combinatorial optimization for routing problems, using the MixRoute approach.
Problem
Generalization of neural combinatorial optimization (NCO) routing solvers across diverse data distributions; existing fixes rely on complex multi-distribution training or meta-learning.
Approach
Questions the presumption that single-distribution training yields a weak zero-shot baseline, suggesting that a simpler single-distribution training scheme can generalize; the specific mechanism is only partially described in the abstract provided.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace POMO's normalization with MixNorm (IN+LN+BN), trained on uniform only | POMO · TSP100 zero-shot avg | optimality gap: 1.70% -> 0.87% | Improved |
| replace POMO's normalization with MixNorm | POMO · CVRP100 zero-shot avg | optimality gap: 2.83% -> 0.72% | Improved |
| replace POMO's normalization with MixNorm | POMO · CVRPTW100 zero-shot avg | optimality gap: 4.37% -> 2.65% | Improved |
| replace POMO's normalization with MixNorm | POMO · CVRPTW100 time-window Width shift | optimality gap: 7.73% -> 12.55% | Hurt |
| apply MixNorm to MatNet | MatNet · ACVRP100 demand shifts | optimality gap: -1.06% -> -1.63% | Improved |
| use MixNorm on CVRPLIB real-world sets | POMO · CVRPLIB seven sets | optimality gap: 11.80% -> 7.94% | Improved |
| use NoNorm instead of BN as third constituent of MixNorm (ablation) | MixRoute · CVRP100 | optimality gap: 0.88% -> 0.95% | No clear effect |
| use RMSNorm instead of BN as third constituent (ablation) | MixRoute · CVRP100 | optimality gap: 0.88% -> 1.41% | Hurt |
| use output-level MixNorm instead of Switchable Norm statistic mixing (ablation) | MixRoute · CVRP100 | optimality gap: 1.68% -> 0.88% | Improved |
| use MixRoute instead of multi-distribution AMDKD-POMO | MixRoute · TSP100 zero-shot avg | optimality gap: 1.05% -> 0.87% | Improved |
Surprising findings
- Single-distribution training may yield a stronger zero-shot baseline than commonly presumed, per the abstract's framing.
Closest papers by meaning
- URS: A Unified Neural Routing Solver for Cross-Problem Zero-Shot Generalization (ICML 2026 · similarity 0.87)
- Improving Generalization of Neural Combinatorial Optimization for Vehicle Routing Problems via Test-Time Projection Learning (NeurIPS 2025 · similarity 0.86)
- Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts (NeurIPS 2026 · similarity 0.85)
- SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver (NeurIPS 2026 · similarity 0.85)
- MVMoE: Multi-Task Vehicle Routing Solver with Mixture-of-Experts (ICML 2024 · similarity 0.85)
- Learning to Route Among Specialized Experts for Zero-Shot Generalization (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
