Probe Guidance Improves Diffusion Language Model Generation via Internal States

Probe guidance utilizes frozen internal states of a diffusion model to generate guidance signals, eliminating the need for additional forward passes during inference. When applied to a 1.7B diffusion language model, this method achieves state-of-the-art results in unconditional generation and improves multiple-choice question answering performance.