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..
VLMLight: Safety-Critical Traffic Signal Control via Vision-Language Meta-Control and Dual-Branch Reasoning Architecture
Authors: Maonan Wang, Yirong Chen, Aoyu Pang, Yuxin Cai, Chung Shue Chen, Yuheng Kan, Man On Pun
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that VLMLight reduces waiting times for emergency vehicles by up to 65% over RL-only systems, while preserving real-time performance in standard conditions with less than 1% degradation. VLMLight offers a scalable, interpretable, and safety-aware solution for next-generation traffic signal control. |
| Researcher Affiliation | Collaboration | 1The Chinese University of Hong Kong, Shenzhen, China 2Shanghai AI Laboratory, Shanghai, China 3Nanyang Technological University, Singapore 4Nokia Bell Labs, Paris, France 5Fourier Intelligence, Shanghai, China |
| Pseudocode | Yes | Algorithm 1 outlines the inference procedure of VLMLight. |
| Open Source Code | Yes | The open-source simulator offers a valuable tool for future research on perception-based traffic systems. |
| Open Datasets | No | We evaluate VLMLight on traffic data collected from three real-world intersections located in Songdo (South Korea), Yau Ma Tei (Hong Kong), and Massy (France), as illustrated in Figure 4. ... Each intersection is equipped with multi-directional cameras capturing 30 minutes of traffic data. The paper describes the collection of data but does not provide concrete access information (link, DOI, repository, etc.) for this collected data to be considered publicly available. |
| Dataset Splits | Yes | Each intersection is equipped with multi-directional cameras capturing 30 minutes of traffic data. The first 20 minutes are used for training the RL policy, while the remaining 10 minutes are reserved for testing. |
| Hardware Specification | Yes | The system is equipped with an Intel Xeon 6738P CPU, 256 GB of RAM, and five A100 GPUs, all running Ubuntu 20.04 LTS. |
| Software Dependencies | Yes | The experiments are conducted using a self-developed traffic simulation environment built on top of the SUMO (Version 1.22) framework. ... For the RL-based components of VLMLight, we use the PPO [28] algorithm, implemented via the Stable Baselines3 library. |
| Experiment Setup | Yes | The total number of environment steps is set to 3e5 and the batch size is configured to 64. The learning rate follows a linear schedule, starting at 1e-3 and gradually decreasing as the number of training steps increases. Additionally, the trajectory memory size is set to 3000. |