To forecast the future, spatiotemporal predictors are often updated at specific intervals. However, various events occurring between these updates can cause the predictors to fail. This paper formalizes this phenomenon from a decomposition perspective based on the joint density of concept drift and proposes an effective solution.

One measure of understanding the real world is the ability to simulate its evolution under given conditions, which becomes significantly more challenging in network-level spatiotemporal systems. This work proposes a benchmark for conditional future generation and a metric for network structural fidelity. Finally, we design a novel diffusion operator to achieve high-fidelity generation of network dynamics.

Embedding-based methods have demonstrated stronger potential than GNNs, particularly in spatiotemporal forecasting. Inspired by STAEformer, this work designs a frequency-domain embedding generation approach that surpasses previous state-of-the-art methods. Notably, while STAEformer originated from a lab where my current supervisor previously worked and involved close collaboration, I had not yet met my supervisor when this research was conducted.

AI infrastructure is the next frontier for scaling intelligence beyond large scale pretraining. While strong researchers are not always strong engineers, our focus is on urban infrastructure, which is often overlooked compared to AI despite its vulnerability to extreme weather. A robust spatiotemporal system should be able to quantify these impacts.

This is the conference version of our paper subsequently submitted to npj Urban Sustainability. The motivation stems from a research philosophy in applied fields: a high-quality application paper should be scenario-driven rather than technology-driven. We validated this through an extensive literature review and developed a multi-agent system to automatically recommend the most suitable methods based on the specific scenario. A demonstration video is available here.

In urban power generation systems, solar PV output is directly tied to weather conditions, making weather a valuable auxiliary variable. We design a multi-task learning (MTL) framework to incorporate weather information. This paper was selected as the cover article of the corresponding issue of Applied Energy and received 20 citations within one year.

Shifting the lens to the recent success of GenAI, hallucination mitigation has been a long-standing topic. ICLR 2026 reportedly rejects papers that cite hallucination literature. While the existence of hallucinations is well known, why they occur remains poorly understood, and more critically, comprehensive benchmarks are lacking. In this work, we design four sub-tasks (KE, KM, RE, and IFE), each targeting a distinct hallucination mechanism, enabling a thorough understanding of LLM hallucinations.