The field of autoregressive image generation and segmentation is moving towards more efficient and effective models. Recent developments have focused on addressing the limitations of traditional autoregressive approaches, such as the quantization process and limited codebook size, by exploring continuous latent spaces and introducing novel training paradigms. Notably, the integration of discrete and continuous representations has shown promising results in improving generation quality and fidelity. Additionally, autoregressive modeling has been applied to image segmentation, framing it as a conditional autoregressive mask generation problem, which has opened up new avenues for spatial-aware vision systems. Furthermore, semantic context has been identified as a crucial factor in improving conditioning for autoregressive models, leading to better instruction adherence and visual fidelity. Overall, these advancements have the potential to significantly impact the field of computer vision and image processing. Noteworthy papers include: MixAR, which introduces a novel framework that leverages mixture training paradigms to inject discrete tokens as prior guidance for continuous AR modeling. Seg-VAR, which proposes a novel framework that rethinks segmentation as a conditional autoregressive mask generation problem, achieving state-of-the-art results on various segmentation tasks. SCAR, which introduces a Semantic-Context-driven method for Autoregressive models, achieving superior visual fidelity and semantic alignment on both instruction editing and controllable generation benchmarks. GloTok, which utilizes global relational information to model a more uniform semantic distribution of tokenized features, delivering state-of-the-art reconstruction performance and generation quality.