In 2012, AlexNet’s victory established that end-to-end learning outperforms manually designed pipelines, a principle that has driven advances in image classification, detection, and segmentation—but generative models have remained the exception. The most powerful generative models today, whether autoregressive or diffusion, are not trained end-to-end; they learn to predict one small step during training but unfold it repeatedly during inference, causing "exposure bias" as errors accumulate. A recent paper from UIUC and Harvard University introduces a new paradigm called Explorative Modeling (XM) to address this gap. The authors argue that generative models have a third amplifiable dimension beyond parameters and data. The core problem is that generation requires capturing many modes, but mainstream models use reconstruction loss such as squared error. When an input is randomly matched with many valid targets, the optimal solution is the average of those targets, which falls between modes and resembles none of them. The paper’s figure illustrates this issue. Project website: https://explorative-modeling.github.io. Paper link: https://arxiv.org/abs/2607.27372. Code repository: https://github.com/alexiglad/XM.