D5P4: Partition Determinantal Point Process for Diversity in Parallel Discrete Diffusion Decoding

This paper presents D5P4, a new decoding method for discrete diffusion models (an alternative to traditional text generation) that generates multiple text candidates in parallel while maintaining diversity. Unlike older methods that struggle to produce varied outputs, D5P4 uses a mathematical technique called Determinantal Point Process to balance between choosing text the model considers likely and ensuring the outputs are diverse, with minimal computational cost.