Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving (LLM2AD)

We study how modules from Large Language Models transfer to tokenized motion generation for autonomous driving.

Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving
Mingyi Wang*, Jingke Wang*, Tengju Ye, Junbo Chen, Kaicheng Yu.

Project Page
arXiv Paper

@article{wang2025llm2ad,
  title   = {Do LLM Modules Generalize? A Study on Motion Generation for Autonomous Driving},
  author  = {Wang, Mingyi and Wang, Jingke and Ye, Tengju and Chen, Junbo and Yu, Kaicheng},
  journal = {arXiv},
  year    = {2025},
  eprint  = {2509.02754},
  doi     = {10.48550/arXiv.2509.02754},
  url     = {https://arxiv.org/abs/2509.02754}
}

News & Updates

• Oct 2025 - Code has been released. Please follow the instruction document
• Sep 2025 — Codes coming soon.
• Sep 2025 — Paper preprint released on arXiv; project page and demo online.
• Aug 2025 — Our paper is accepted at CoRL 2025. Cheers!

Video

You can watch the short demo on the Project Page ▶.

Acknowledgement

We respectfully acknowledge MotionLM as a key inspiration. Our codebase is an independent, from-scratch implementation. We thank the authors for their outstanding contributions.