The International Conference on Machine Learning (ICML) 2026, taking place from July 6 to 11 in Seoul, South Korea, features a landmark contribution from the College of Civil Engineering (CCE) at Tongji University. For the first time, a paper authored by CCE researchers has been accepted at this premier venue, which is one of the “top-tier trio” AI conferences alongside the Conference on Neural Information Processing Systems (NeurIPS) and the International Conference on Machine Learning (ICLR).

The paper, titled “Stabilizing PPO via Latent-Space Regularization and KDE-Driven Exploration,” is authored by PhD student Du Meiyu, co-supervised by Prof. Wang Wei and Associate Prof. Gao Yuqing. Du, who earned his bachelor’s degree in Smart Construction at CCE and is currently pursuing his doctorate in the same program, focuses on reinforcement learning for structural optimization. Associate Prof. Gao led the team to Seoul for the poster session and oral presentation.

The research paper accepted by ICML 2026 tackles a fundamental challenge in reinforcement learning: the instability of Proximal Policy Optimization (PPO). The authors introduce SPPO, an enhanced framework that combines latent-space regularization with kernel density estimation (KDE)-driven exploration, achieving more stable training, efficient exploration, and stronger policy generalization. While the contribution advances foundational AI methodology, it also provides critical algorithmic support for downstream applications in structural intelligent design, smart construction, and robotic autonomous control, which are the domains at the heart of CCE’s interdisciplinary blueprint.

 This paper represents a flagship achievement of the TAIER Lab (Tongji Artificial Intelligence Engineering Research Lab), established by CCE to foster cross-disciplinary innovation. It reflects the college’s strategic investment in AI-forward research, talent cultivation, and the intelligent transformation of civil engineering disciplines.

 Notably, ICML, a CCF Class-A conference, is widely regarded as one of the most selective and influential gatherings in machine learning. Its acceptance marks a notable indicator of CCE’s growing strength at the intersection of artificial intelligence and engineering.