Where Arbitrary Order Helps: Future-First Story Generation with Diffusion Language Models

Story generation allows future events to be chosen before constructing the actions and transitions that lead to them. Event plans specify such destinations, but arbitrary token order alone does not tell a diffusion language model how to form them and use them to guide the intervening prose. We introduce Single-Canvas Semantic Growth, instantiated as Self-SiC, which couples self-generated narrative waypoints with their textual realization on one persistent canvas. These waypoints encode sparse events or character states at their final output positions, giving the same frozen model visible destinations for generating the intervening prose. In a controlled placement study, in-canvas futures reduce dominant target restatement from 51/200 to 2/200. The complete Self-SiC pipeline is preferred over Standard Prompt for narrative-path quality in 20/32 paired stories.