An Agentic Framework with LLMS for Solving Complex Vehicle Routing Problems

Our paper “An Agentic Framework with LLMs for Solving Complex Vehicle Routing Problems” has been accepted to ICLR 2026.

Generate code for solving VRPs

python main.py --path "path_to_instance_floder"

The generated code will be stored in the code directory. For example, to generate code for CVRP, you can run:

python main.py --path vrp/cvrp/50

Execute Generated Code for solving VRPs

python test.py --path "path_to_instance_folder" --problem "problem_name" --iteration <num_iterations>

For example, to solve the CVRP, you can run:

python test.py --path vrp/cvrp/50 --problem CVRP --iteration 1000

Reference

If you find this codebase useful, please consider citing the paper:

@inproceedings{zhang2026afl,
  title     = {An Agentic Framework with LLMs for Solving Complex Vehicle Routing Problems},
  author    = {Zhang, Ni and Cao, Zhiguang and Zhou, Jianan and Zhang, Cong and Ong, Yew-Soon},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://openreview.net/forum?id=BMOgYw4EhQ}
}