Proposed a training-free steering method for diffusion language models, exploiting Jacobian smoothness of pretrained embedding manifolds to maintain contextual coherence without fine-tuning. Uses dual-pass embedding manifold divergence to steer generation at inference time.