<p>Patch-wise learning is a common strategy for training neural networks on large-scale dense prediction problems, yet existing approaches assume uniform or fixed sampling distributions. This assumption is suboptimal when learning difficulty varies spatially and evolves with the model state during optimization. We reformulate patch-wise learning as a dynamic computation allocation problem and propose an adaptive patch sampling (APS) algorithm that learns where to sample by constructing model-state-dependent sampling distributions from voxel-wise uncertainty and prediction error. To learn what contextual information is encoded within sampled patches, we introduce a patch encoding (PE) block that infers implicit location information and modulates feature representations through context-dependent channel-wise attention, without relying on explicit spatial coordinates. Experiments on whole-body multi-organ PET-CT segmentation demonstrate faster convergence and consistent performance gains, with external validation on Synapse dataset confirming robustness. Mechanistic analyses of learning dynamics further characterize sampling behavior induced by APS and representation modulation driven by the PE block through attention analysis and causal channel pruning. Overall, this work contributes an efficient learning strategy for patch-wise training and provides insight into how dynamic sampling and contextual conditioning influence optimization in large-scale dense prediction tasks.</p>