Research map / Architectures and efficiency / RNNs and normalization
Efficient Decoder Scaling Strategy for Constructive Neural Routing Solvers
Qing Luo, Fu Luo, Ke LI, Zhenkun Wang
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.
Studies scaling the decoder size of construction-based neural routing solvers for vehicle routing problems, rather than only the encoder.
Problem
Construction-based neural routing solvers for vehicle routing typically use small decoders (1-3M parameters), leaving the effects of decoder scaling largely unexplored.
Approach
Systematic study of decoder scaling in encoder-decoder constructive routing solvers, shifting parameters toward the decoder and examining how performance scales with decoder size, with an efficiency-oriented scaling strategy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| restructure to deep-narrow (D=96, W=32) at fixed ~1.3M parameters (ablation) | Decoder-only (D=6,W=128) · TSP100 | optimality gap (%): 0.464% -> 0.184% | Improved |
| restructure to deep-narrow (D=96, W=32) at fixed parameters (ablation) | Decoder-only (D=6,W=128) · TSP1000 | optimality gap (%): 2.172% -> 0.991% | Improved |
| increase depth 6 -> 42 instead of width, limited training data (0.1M) (ablation) | Decoder-only · TSP100 | optimality gap (%): 1.715% -> 0.547% | Improved |
| increase width 128 -> 512 with limited data (0.1M) (ablation) | Decoder-only · TSP100 | optimality gap (%): 1.715% -> 0.949% | Improved |
| scale depth instead of width in parameter scaling law (ablation) | Decoder-only · TSP100 | scaling exponent: depth 0.98-1.05 vs width 0.24-0.40 | Improved |
| scale depth (6 -> 42) for data efficiency exponent (ablation) | Decoder-only (W=128) · TSP100 | data scaling exponent: 0.47 -> 0.71 | Improved |
| scale depth 6 -> 42 for compute efficiency (ablation) | Decoder-only (W=128) · TSP100 beam search | compute exponent: 0.82 -> 1.71 | Improved |
| use medium-depth (24) model instead of wide model under 12-minute budget (ablation) | Decoder-only · TSP100 | optimality gap (%): 0.0138 -> 0.0044 | Improved |
| use D=42 instead of D=24 with sufficient inference budget (ablation) | Decoder-only (W=128) · TSP100 | optimality gap (%): 0.0002 -> 0.0001 | Improved |
| use deep model (42,128) with greedy decoding vs best baseline with beam16 | Decoder-only · TSP1000 | optimality gap (%): 1.35 -> 0.869 | Improved |
| fit single parameter-count power law to all 12 models (ablation) | Decoder-only · TSP100 | R^2 of fit: R^2 0.794 vs 0.989 decoupled | Hurt |
| scale to D=42, W=512 | Decoder-only · TSPLib (all) | optimality gap (%): 5.235 -> 1.288 | Improved |
Closest papers by meaning
- LINC: Decoupling Local Consequence Scoring from Hidden Matching in Constructive Neural Routing (NeurIPS 2026 · similarity 0.88)
- Rethinking Light Decoder-based Solvers for Vehicle Routing Problems (ICLR 2025 · similarity 0.88)
- SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver (NeurIPS 2026 · similarity 0.88)
- A Mixed-Curvature based Pre-training Paradigm for Multi-Task Vehicle Routing Solver (ICML 2025 · similarity 0.87)
- Improving Generalization of Neural Combinatorial Optimization for Vehicle Routing Problems via Test-Time Projection Learning (NeurIPS 2025 · similarity 0.86)
- MVMoE: Multi-Task Vehicle Routing Solver with Mixture-of-Experts (ICML 2024 · similarity 0.86)
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.
