Diffusion language models generate text by denoising in parallel, committing several tokens per forward pass. Because those tokens are produced simultaneously, they cannot condition on each other. That is harmless for many sequences, but it breaks multi-token format units—such as specific keywords or syntax—that parsers need to recognize as a single structure.

The paper proposes intent-triggered anchoring: the model detects when a format unit is about to be generated and anchors the decoding so that all tokens in that unit are committed together, preserving their internal consistency. The authors evaluate the approach on ECHO, an open-weight one-step chest X-ray model, though the released abstract does not include quantitative results.