We propose a novel architectural modification and post-training pipeline for enhancing large language model reasoning capabilities by teaching models to truncate forward passes early. Our approach augments the standard transformer architecture with an early-exit mechanism at intermediate layers, enabling the model to exit at shallower layers when tokens can be predicted without deep computation. Through a calibration stage followed by reinforcement learning, we incentivize the model to exit as early as possible while preserving task performance. Preliminary experiments on small reasoning models demonstrate adaptive computation reduction across tokens, suggesting that at appropriate scale, this approach can minimize excess computation for non-myopic planning using internal activations, reserving deep computation only for difficult-to-predict tokens.