Researchers introduce DCOCO, a mechanism that restores dependencies in parallel diffusion-based drafting to improve large language model inference throughput across four verifier backbones and ten benchmarks.
Carried by: arXiv, HF Daily Papers. First seen: .