Tao Hu

Yifei Yu, Xiaoshan Wu, Xinting Hu et al.

Autoregressive (AR) diffusion enables streaming, interactive long-video generation by producing frames causally, yet maintaining coherence over minute-scale horizons remains challenging due to accumulated errors, motion drift, and content repetition. We approach this problem from a memory perspective, treating video synthesis as a recurrent dynamical process that requires coordinated short- and long-term context. We propose VideoSSM, a Long Video Model that unifies AR diffusion with a hybrid state-space memory. The state-space model (SSM) serves as an evolving global memory of scene dynamics across the entire sequence, while a context window provides local memory for motion cues and fine details. This hybrid design preserves global consistency without frozen, repetitive patterns, supports prompt-adaptive interaction, and scales in linear time with sequence length. Experiments on short- and long-range benchmarks demonstrate state-of-the-art temporal consistency and motion stability among autoregressive video generator especially at minute-scale horizons, enabling content diversity and interactive prompt-based control, thereby establishing a scalable, memory-aware framework for long video generation.

Chao Zhang, Dacheng Tao, Tao Hu et al.

In this paper, we study the generalization performance of regularized multi-task learning (RMTL) in a vector-valued framework, where MTL is considered as a learning process for vector-valued functions. We are mainly concerned with two theoretical questions: 1) under what conditions does RMTL perform better with a smaller task sample size than STL? 2) under what conditions is RMTL generalizable and can guarantee the consistency of each task during simultaneous learning? In particular, we investigate two types of task-group relatedness: the observed discrepancy-dependence measure (ODDM) and the empirical discrepancy-dependence measure (EDDM), both of which detect the dependence between two groups of multiple related tasks (MRTs). We then introduce the Cartesian product-based uniform entropy number (CPUEN) to measure the complexities of vector-valued function classes. By applying the specific deviation and the symmetrization inequalities to the vector-valued framework, we obtain the generalization bound for RMTL, which is the upper bound of the joint probability of the event that there is at least one task with a large empirical discrepancy between the expected and empirical risks. Finally, we present a sufficient condition to guarantee the consistency of each task in the simultaneous learning process, and we discuss how task relatedness affects the generalization performance of RMTL. Our theoretical findings answer the aforementioned two questions.