Research map / Learning theory and algorithms / Combinatorial and search optimization
Neural Combinatorial Optimization with Heavy Decoder: Toward Large Scale Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace heavy-encoder POMO with heavy-decoder LEHD under same SL training (ablation) | constructive NCO · TSP1000 | optimality gap: 34.419 -> 3.168 | Helped |
| train LEHD with supervised learning instead of RL (ablation) | LEHD · TSP1000 | optimality gap: 46.441 -> 6.566 | Helped |
| remove normalization from attention layers (vs batch norm) (ablation) | LEHD · TSP1000 | optimality gap: 12.209 -> 3.168 | Helped |
| remove normalization from attention layers (vs instance norm) (ablation) | LEHD · TSP1000 | optimality gap: 222.730 -> 3.168 | Helped |
| remove normalization from POMO (ablation) | POMO · TSP1000 | optimality gap: 41.631 -> 45.747 | No clear effect |
| add RRC1000 to greedy LEHD | LEHD · TSP1000 | optimality gap: 3.168 -> 0.719 | Helped |
| add RRC100 to POMO instead of random sampling (ablation) | POMO · TSP1000 | optimality gap: 82.015 -> 25.378 | Helped |
| use LEHD instead of POMO with RRC1000 (ablation) | RRC · TSP1000 | optimality gap: 16.261 -> 0.735 | Helped |
| train LEHD with suboptimal OR-Tools labels plus RRC500 (ablation) | LEHD · CVRP100 | optimality gap: 10.428 -> 2.525 | Helped |
| self-improved training (RL then SL on RRC solutions) vs SL on optimal labels (ablation) | LEHD · TSP1000 | optimality gap: 1.195 -> 2.022 (RRC100) | Hurt |
| LEHD greedy vs BQ greedy | LEHD · TSP1000 | optimality gap: 3.965 -> 3.168 | Helped |
| LEHD greedy vs BQ greedy | LEHD · CVRP100 | optimality gap: 2.993 -> 3.648 | Hurt |
Closest papers by meaning
- Improving Generalization of Neural Combinatorial Optimization for Vehicle Routing Problems via Test-Time Projection Learning (NeurIPS 2025 · similarity 0.91)
- Neural Solver Selection for Combinatorial Optimization (ICML 2025 · similarity 0.90)
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.87)
- Learning to Insert for Constructive Neural Vehicle Routing Solver (NeurIPS 2025 · similarity 0.87)
- Rethinking Neural Combinatorial Optimization for Vehicle Routing Problems with Different Constraint Tightness Degrees (NeurIPS 2025 · similarity 0.87)
- Rethinking Light Decoder-based Solvers for Vehicle Routing Problems (ICLR 2025 · similarity 0.87)
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.