New GeoBlock method improves efficiency of parallel text generation in diffusion models by dynamically adapting block sizes based on token dependencies

GeoBlock analyzes attention-derived dependency patterns to determine optimal block boundaries during decoding, replacing fixed rules and heuristic-based approaches

The framework identifies regions with strong causal ordering that require sequential updates versus semantically cohesive regions suitable for parallel refinement

By adapting block granularity to the actual geometry of token dependencies, GeoBlock enables more efficient parallel decoding in diffusion language models