ReFusion: A Novel Diffusion LLM Leveraging Parallel Decoding

This research introduces a novel architecture that merges diffusion models with large language models, aiming for improved efficiency. The parallel autoregressive decoding approach is particularly interesting for accelerating the generation process.