Training-free steering of discrete diffusion models toward downstream rewards is increasingly important

steering them toward a downstream reward at inference time, without any retraining, is increasingly important

http://arxiv.org/abs/2608.26585v1