Generative models may be slow because of how we train them. When many outputs are valid, training a reconstructive model to predict one in a single step can pull its output toward an average of those possibilities. Diffusion and autoregressive models avoid this by breaking generation into smaller steps that must be repeated at inference. The simplest version of Explorative Modeling generates several possible outputs for each training example and learns from the one that best matches the target. This allows the model to capture different valid outputs rather than average them together. Added to a strong image-generation baseline, exploration reached the baseline’s final performance after processing 6.2x fewer training samples and using 4.1x fewer FLOPs. It also improved video and masked-diffusion language models. In separate scaling experiments, the gains increased as model size and training data grew. Fully end-to-end Explorative Models matched diffusion baselines on control tasks while using up to 256x fewer network evaluations at inference. Exploration may help models resolve more uncertainty during training so they need fewer steps to generate an output.