Autoregressive models have emerged as a powerful alternative to diffusion for image and video generation, making them a natural fit for streaming and interactive video world models. However, because they generate content step by step and condition on their own previous outputs, they remain vulnerable to exposure bias—where early errors propagate and shape the entire generation process.

In this talk, Angela Yao (Associate Professor at the National University of Singapore) presents recent research from her group on mitigating exposure bias across the generation pipeline. The session covers key advances, including tokenizer and generator regularization, pixel-level quality alignment post-training, training-free guidance to reinforce historical consistency, and reinforcement learning strategies designed to maintain long-horizon stability in video world models.