Abstract
Diffusion Language Models (DLMs) promise highly parallel text generation, yet their practical inference speed is often bottlenecked by suboptimal decoding schedulers. Standard approaches rely on 'scattered acceptance'-committing high confidence tokens at disjoint positions throughout the sequence. This approach inadvertently fractures the Key-Value (KV) cache, destroys memory locality, and forces the model into costly, repeated repairs across unstable token boundaries. To resolve this, we present the Longest Stable Prefix (LSP) scheduler, a training-free and model-agnostic inference paradigm based on monolithic prefix absorption. In each denoising step, LSP evaluates token stability via a single forward pass, dynamically identifies a contiguous left-aligned block of stable predictions, and snaps its boundary to natural linguistic or structural delimiters before an atomic commitment. This prefix-first topology yields dual benefits: systemically, it converts fragmented KV cache updates into efficient, contiguous appends; algorithmically, it preserves bidirectional lookahead over a geometrically shrinking active suffix, drastically reducing token flip rates and denoiser calls. Extensive evaluations on LLaDA-8B and Dream-7B demonstrate that LSP accelerates inference by up to 3.4x across rigorous benchmarks including mathematical reasoning, code generation, multilingual (CJK) tasks, and creative writing while matching or slightly improving output quality. By fundamentally restructuring the commitment topology, LSP bridges the gap between the theoretical parallelism of DLMs and practical hardware efficiency.
Abstract (translated)
扩散语言模型（DLM）承诺了高度并行的文本生成，但它们的实际推理速度常常因为不理想的解码调度程序而受限。标准方法依赖于“散列接受”——在整个序列中分散确定高置信度的令牌。这种方法无意中破坏了键值（KV）缓存的一致性，摧毁了内存局部性，并迫使模型在不稳定令牌边界处进行昂贵且重复的修复操作。 为了解决这一问题，我们提出了最长稳定前缀（LSP）调度程序，这是一种无训练、模型无关的推理范式，基于单体前缀吸收。在每次去噪步骤中，LSP通过一次正向传递评估令牌稳定性，动态识别一系列连续左对齐的稳定预测，并将其边界调整到自然语言或结构分隔符处后进行原子承诺。 这种以前缀为主的拓扑结构带来了双重好处：系统层面，它将碎片化的KV缓存更新转化为高效的连续追加；算法方面，它在整个活跃后缀几何收缩的过程中保留了双向前瞻的能力，大大减少了令牌翻转率和去噪器调用次数。在LLaDA-8B和Dream-7B上的广泛评估表明，LSP能够通过严格的基准测试（包括数学推理、代码生成、多语言（中日韩）任务以及创意写作），将推理速度提高多达3.4倍，并且输出质量与现有方法相当或略有提升。 通过从根本上重新构建承诺拓扑结构，LSP弥合了DLM理论并行性与实际硬件效率之间的差距。
URL
https://arxiv.org/abs/2603.05454
