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..
Purity Law for Neural Routing Problem Solvers with Enhanced Generalizability
Authors: Wenzhao Liu, Haoran Li, Congying Han, Zicheng Zhang, Anqi Li, Tiande Guo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that PUPO can be seamlessly integrated with popular neural solvers, significantly enhancing their generalization performance without incurring additional computational overhead during inference. The code is available at https://github.com/Kejun0627/PUPO. [...] 6 Numerical Experiments To validate the effect of PUPO on enhancing generalization, we compare the performance of both vanilla policy optimization and PUPO across several state-of-the-art constructive neural solvers. Additionally, we conduct a comprehensive evaluation of generalization performance and purity metrics on well-known public datasets. |
| Researcher Affiliation | Collaboration | 1School of Mathematical Sciences, University of Chinese Academy of Sciences 2JD.com 3School of Mathematical Sciences, Nankai University |
| Pseudocode | Yes | The whole algorithm of PUPO is presented in Algorithm 1. [...] Algorithm 1 PUPO Training |
| Open Source Code | Yes | Extensive experiments demonstrate that PUPO can be seamlessly integrated with popular neural solvers, significantly enhancing their generalization performance without incurring additional computational overhead during inference. The code is available at https://github.com/Kejun0627/PUPO. |
| Open Datasets | Yes | The randomly generated dataset used in this paper is the same as that in INVi T [10], which is widely adopted to testify existing DRL approach. The real-world dataset we used is TSPLIB [39] and CVRPLIB [40]. |
| Dataset Splits | Yes | For each method, we conduct training on scales of 50 and 100. [...] For TSP, it contains 16 subsets and corresponding (near-)optimal solutions for TSP, including 4 distributions (uniform, clustered, explosion, and implosion, denoted as U, C, E, I) and 4 scales (100, 1000, 5000 and 10000). For CVRP, it contains 12 subsets and corresponding (near-)optimal solutions for CVRP, including 4 distributions (uniform, clustered, explosion, and implosion, denoted as U, C, E, I) and 3 scales (50, 500 and 5000). |
| Hardware Specification | Yes | All the numerical experiments are implemented on an NVIDIA GeForce RTX 3090 GPU with 24 GB of memory, paired with a 12th Gen Intel(R) Core(TM) i9-12900 CPU. |
| Software Dependencies | No | No specific software dependencies with version numbers are provided in the paper. The text only mentions using 'original network architectures and hyper-parameters provided in the source code'. |
| Experiment Setup | Yes | In each training paradigm, we use the original network architectures and hyper-parameters provided in the source code, without any modifications to modules. Due to the different magnitude in the policy gradient between the two, we only adjust the learning rates during PUPO. Details about the learning rates of each model are available in Appendix O. [...] For TSP, the learning rates are set to 0.0001, 0.00015, 0.0001, 0.00017, 0.0001, 0.00012, 0.00011, 0.00012 for POMO-50, POMO-100, PF-50, PF-100, ELG-50, ELG-100, INVi T-50, and INVi T-100, respectively. For CVRP, the learning rates are set to 0.0001, 0.00006, 0.0001, 0.0001, 0.00005, 0.0001 for POMO-50, POMO-100, ELG-50, ELG-100, INVi T-50, and INVi T-100. |