We propose a conformal prediction framework tailored to temporal point processes, resolving challenges posed by their continuous-time nature, temporal dependence, and violation of exchangeability assumptions. We establish rigorous coverage guarantees under mild assumptions of periodic stationarity and \(\beta\)-mixing, and extend the theoretical results to Hawkes processes and non-periodic stationary processes.
Published in Forty-second International Conference on Machine Learning (ICML), 2025
We introduce Residual TPP, a novel, unified, and lightweight approach that leverages the strengths of both classical temporal point processes (TPPs) and expressive neural TPPs to achieve superior performance. Specifically, we propose the Residual Events Decomposition (RED) technique for TPPs, which defines a weight function to quantify how well the intensity function captures the event characteristics. The RED serves as a flexible, plug-and-play module that can be integrated with any TPP model in a wide range of tasks.
We introduce a continuous memory framework for GUI agents, along with an auto-scaling data flywheel that autonomously discovers environments, generates tasks, collects trajectories, and performs quality evaluation. Our method achieves state-of-the-art results across multiple real-world GUI benchmarks.
Published in Journal of the American Statistical Association, 2026
We develop a comprehensive framework for analyzing computer-based problem-solving assessments that predicts students progression toward correct answers in real-time and summarizes it into probability trajectories. The resulting curves are robust, model-agnostic statistics that enable multiple downstream applications.
Published in Forty-third International Conference on Machine Learning (ICML), 2026
We study exact sparse time-series modeling via $\ell_0$-constrained least squares under temporal dependence. For stationary $\alpha$-mixing Gaussian processes with sparse coefficients, we establish high-probability RSC/RSS for the empirical quadratic loss and obtain nonasymptotic statistical guarantees for exact sparse algorithms such as iterative hard thresholding (IHT).