Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning to Insert for Constructive Neural Vehicle Routing Solver
Authors: Fu Luo, Xi Lin, Mengyuan Zhong, Fei Liu, Zhenkun Wang, Jianyong Sun, Qingfu Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we empirically evaluate the L2C-Insert framework on both synthetic and real-world benchmark TSP and CVRP datasets of various scales. We then contact ablation studies to verify the effect of the proposed framework s critical components. |
| Researcher Affiliation | Academia | 1 School of Automation and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen, China 2 Guangdong Provincial Key Laboratory of Fully Actuated System Control Theory and Technology, Southern University of Science and Technology, Shenzhen, China 3 Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China 4 School of Mathematics and Statistics, Xi an Jiaotong University, Xi an, China |
| Pseudocode | No | The paper describes the model structure and training scheme with figures and textual descriptions, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/CIAMGroup/L2C_Insert. |
| Open Datasets | Yes | To evaluate our method on real-world large-scale instances, we also collect 49 symmetric TSP instances and 100 CVRP instances with EUC_2D features and 1K nodes from TSPLIB [46] and CVRPLIB Set-X [47], respectively. |
| Dataset Splits | Yes | For each problem (TSP and CVRP), we create training datasets consisting of one million 100-node instances. ... For TSP, the TSP100 and TSP1K datasets contain 10,000 and 128 instances, respectively, while larger-scale datasets (10K, 50K, 100K nodes) each have 16 instances. CVRP datasets follow a similar structure, with the exception that CVRP1K contains 100 instances, as specified in [44]. |
| Hardware Specification | Yes | All experiments, including training, testing, and evaluation, are conducted on a single NVIDIA Ge Force RTX 4090 GPU with 24 GB of memory to ensure consistent computational conditions. |
| Software Dependencies | No | For TSP model training, we employ the Adam optimizer [37]. For CVRP model training, we employ the Adam W [48] optimizer, as we find it to be more stable. No specific versions for programming languages or libraries are provided. |
| Experiment Setup | Yes | The embedding dimension is set to d = 128, the hidden dimension of the feed-forward layer is set to 512, the query dimension dq is set to 16, and the head number in multi-head attention is set to 8. We set the number of attention layers in the decoder to L = 9 following [23]. ... The initial learning rate is 1 10-4, decaying by a factor of 0.97 after each epoch. The TSP and CVRP models undergo 50 and 15 training epochs, respectively, with a batch size of 1024. |