Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models

In three linesDynamic Infilling Anchors (DIA) is a training-free method for diffusion large language models (dLLMs) that dynamically adjusts anchor positions to generate structured content (JSON, reasoning templates). Tested on GSM8K and MATH, DIA improves format compliance and answer accuracy by avoiding rigid spans.

Summary generated by Claude — human-verified