Understanding the structural dynamics of biomolecules is vital for elucidating biological function. With the increasing availability of molecular dynamics (MD) simulation data, deep generative models have been developed to synthesize realistic MD trajectories. However, existing approaches generate fixed-length trajectories by jointly denoising high-dimensional spatiotemporal representations, which conflicts with MD’s frame-by-frame integration process and fails to capture time-dependent conformational diversity. Motivated by the sequential nature of MD, we introduce a novel probabilistic autoregressive (\textbf{ProAR}) framework for trajectory generation. ProAR employs a dual-network system that explicitly models each frame as a multivariate Gaussian distribution and uses an anti-drifting sampling strategy to mitigate cumulative errors, thereby capturing conformational uncertainty and time-coupled structural changes while flexibly generating trajectories of arbitrary length. Experiments on ATLAS, a large-scale protein MD dataset, show that for the long trajectory generation task, our model achieves a 7.5\% reduction in reconstruction RMSE and an average 25.8\% improvement in conformation change accuracy over previous state-of-the-art methods. Regarding the conformation sampling task, it attains comparable performance to specialized time-independent models, offering a flexible and reliable alternative to conventional MD simulations.