D5P4: Partition Determinantal Point Process for Diversity in Parallel Discrete Diffusion Decoding
Researchers have introduced D5P4, a novel decoding method designed for discrete diffusion models used in text generation. This method frames intermediate beam selection as MAP inference within a partitioned Determinantal Point Process. D5P4 aims to enhance the diversity and coverage of generated hypotheses while maintaining or improving quality and fidelity, as demonstrated in experiments across various generation tasks. AI
IMPACT Introduces a new decoding strategy that could improve the quality and diversity of text generated by discrete diffusion models.