EVENT DOSSIER
The researchers proposed a discrete diffusion-based approach aimed at providing a non-destructive improvement in inference speed for large language models (LLMs). By introducing a discretization mechanism, this method optimizes the computational process while maintaining the accuracy of model outputs, thereby significantly reducing latency and improving generation efficiency.
论文《通过离散扩散解锁大语言模型无损加速》发布后，作者 Nguyen Van Chien 与 Uno 团队就核心框架相似性产生争议。Uno 团队指出其方法 Uno 旨在保持架构不变以实现 AR 和扩散生成，而该论文修改了架构并添加了扩散注意力头及双向注意力机制。Nguyen Van Chien 回应称双方均保留冻结的 Transformer 骨干网络，认为“架构未变”与“架构改变”的区分不准确，并将在后续修订中补充讨论。