[NeurIPS 2025] MTL-KD: Multi-Task Learning via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver

Overview

This repository contains the code implementation of paper MTL-KD: Multi-Task Learning via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver. In this paper, we present a unified multi-task neural solver capable of solving 16 different Vehicle Routing Problem (VRP) variants. The model employs a heavy decoder and light encoder architecture, distilled from multiple single-task student models (each with heavy encoder and light decoder). This approach avoids the need for generating extensive supervised datasets and demonstrates excellent performance in both scale generalization and cross-task generalization.

Dependencies

Python=3.9.19
torch==2.0.1
numpy==1.24.4
matplotlib==3.7.5
tqdm==4.65.5
pytz==2022.1

Dataset Download

The test datasets can be downloaded from either of the following two links:

Google Drive：
https://drive.google.com/drive/folders/1E_pu4a7BWdRPhbGcLJkQ86-74B-X2ibh?usp=sharing

Baidu Netdisk:
https://pan.baidu.com/s/1YI9vI5EA64H3SifY6IX-6g?pwd=jjkf

After downloading, place the datasets in the MTLKD/data directory.

Training

To train the model, use the following command:

cd student_model
python run.py

Testing

Uniform Distribution Dataset

cd student_model
python test.py

You can modify the test_problem list and test_sizes list in main() to select specific problems and sizes for testing. For other configuration changes, please refer to test.py.

LIB Dataset

cd student_model
python test_lib.py

You can change the PROBLEM variable in test_lib.py to switch between testing CVRP or VRPTW problems. Remember to also switch the dataset path accordingly (e.g., ../data/CVRP-LIB or ../data/Vrp-Set-Solomon).

Citation

If this repository is helpful for your research, please cite our paper:

"Yuepeng Zheng, Fu Luo, Zhenkun Wang, Yaoxin Wu, Yu Zhou, MTL-KD: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver, The Thirty-ninth Annual Conference on Neural Information Processing Systems, NeurIPS 2025"

BibTeX:

@inproceedings{
Zheng2025mtlkd,
title={{MTL}-{KD}: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver},
author={Yuepeng Zheng and Fu Luo and Zhenkun Wang and Yaoxin Wu and Yu Zhou},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=rlH3e7VlY8}
}

Acknowledgements

The MTL-KD code implementation is based on the following works: LEHD,MVMoE,AMDKD. We express our sincere gratitude to the authors of these projects for their valuable contributions.