●builderThis approach could potentially reduce inference latency for text generation tasks by increasing the useful output per forward pass.
●researcherThis architecture explores a middle ground between standard autoregressive models and diffusion-based decoding by treating lines as parallel units.