Adaptive scheduling steers diffusion L

Short summary

This paper reveals that different text attributes (topic, sentiment) solidify at distinct points during discrete diffusion language model generation. An adaptive scheduling algorithm concentrates steering interventions exactly when target attributes are forming, achieving 93% control strength without quality degradation—a major improvement over uniform intervention.

• •Discrete diffusion models generate text by parallel denoising of all positions; uniform steering at every step degrades output quality
• •Sparse autoencoders expose that topic commits early (first 2% of denoising) while sentiment emerges gradually (20% of process)
• •Adaptive scheduler concentrates interventions only when attributes are actively forming; reaches 93% steering strength on three-attribute control

Generated with AI, which can make mistakes.