首页 > AI前沿 > One Schedule Does Not Fit All: Adaptive Inference Schedules for Diffusion Language Models

One Schedule Does Not Fit All: Adaptive Inference Schedules for Diffusion Language Models

arXiv自然语言 2026-11-19 18:47 12 阅读 查看原文
用户现在需要把给定的纯文本新闻正文排版成语义化HTML片段,要符合要求:每个自然段用p,没有文档级标签,保留所有信息。首先看正文,先处理第一个段落?不对,先理清楚内容:首先开头的几句是一段?然后看列表部分,候选的那几个用ul,然后还有各种结果,重要的地方用strong?不对,要准确。首先,先分段落: 第一段:开头到“falls back to the full schedule.”?不对,不对,看内容: 首先,第一句到“guardrails against degeneration, TCDA uses simultaneous one-sided bootstrap bounds to certify which candidates meet both the utility and guardrail criteria.”?不对,应该自然分段。哦不对,首先: 第一个自然段:Diffusion language models typically use a fixed denoising schedule across workloads. We show that the cheapest schedule that preserves quality depends on the task, model, and sampler. We propose Task-Calibrated Denoising Allocation (TCDA), a deployment rule with an explicit error guarantee. Given candidate samplers, paired calibration data, a utility tolerance, and guardrails against degeneration, TCDA uses simultaneous one-sided bootstrap bounds to certify which candidates meet both the utility and guardrail criteria. It deploys the cheapest certified candidate or falls back to the full schedule. The probability of deploying an infeasible schedule is asymptotically at most alpha. 对,这部分是一段?不对,然后接下来是候选的列表,所以

之后是

Candidates include:

然后
  • early prefix decoding
  • state-reusing continuation
  • reduced time grids
  • confidence-threshold decoding
?不对,原文是“Candidates include early prefix decoding, state-reusing continuation, reduced time grids, and confidence-threshold decoding.”所以可以写成

Candidates include:

然后
    ,对吧,因为是列表内容。然后接下来是实验结果的段落:

    Across seven datasets and two model families, TCDA certifies reductions in denoiser evaluations by factors of 3.4 on XSum and 2.0 on WMT14 for ELF-B; 8.0 on WMT14 and 4.1 on GSM8K for LLaDA-8B; and 1.33 on OpenWebText for both ELF-M and ELF-L.

    然后下一段:

    A validation protocol fixed before generation confirmed every selection on fresh seeds. Across 500 calibration splits, false certification remained below 2% in every setting at a nominal level of 5%.

    然后下一段:

    Certified schedules vary across tasks and model scales. Trajectory measurements suggest that when decoded tokens commit is more informative for schedule selection than the remaining latent discretization error.

    然后最后一段:

    We also find that the ELF reference sampler draws its time grid at random, so reproducible schedule comparisons for unconditional generation require fixing the grid.

    等下,要注意不要增删改内容,所以原文的Candidates include后面的内容,要准确拆分,不能改。哦不对,原文是“Candidates include early prefix decoding, state-reusing continuation, reduced time grids, and confidence-threshold decoding.”所以正确的处理是,要么写成

    Candidates include:

    然后
      里的四个li,每个对应原文的四个项,这样没有增删,对吧?对,因为是列表内容,所以用ul是对的。然后重要的名词比如TCDA可以加strong,实验数据也可以加?不对,不要乱加,用户说重要句子可用strong,所以合理的重要术语和关键数据可以加。等下,再核对原文所有信息: 现在逐句对应: 1. Diffusion language models typically use a fixed denoising schedule across workloads. We show that the cheapest schedule that preserves quality depends on the task, model, and sampler. We propose Task-Calibrated Denoising Allocation (TCDA), a deployment rule with an explicit error guarantee. Given candidate samplers, paired calibration data, a utility tolerance, and guardrails against degeneration, TCDA uses simultaneous one-sided bootstrap bounds to certify which candidates meet both the utility and guardrail criteria. It deploys the cheapest certified candidate or falls back to the full schedule. The probability of deploying an infeasible schedule is asymptotically at most alpha. → 这个是第一个p,TCDA是重要术语,加strong,对吧?所以

      Diffusion language models typically use a fixed denoising schedule across workloads. We show that the cheapest schedule that preserves quality depends on the task, model, and sampler. We propose Task-Calibrated Denoising Allocation (TCDA), a deployment rule with an explicit error guarantee. Given candidate samplers, paired calibration data, a utility tolerance, and guardrails against degeneration, TCDA uses simultaneous one-sided bootstrap bounds to certify which candidates meet both the utility and guardrail criteria. It deploys the cheapest certified candidate or falls back to the full schedule. The probability of deploying an infeasible schedule is asymptotically at most alpha.

      2. 接下来是Candidates include的内容,所以

      Candidates include:

      然后
      • early prefix decoding
      • state-reusing continuation
      • reduced time grids
      • confidence-threshold decoding
      ,对吗?对,因为原文是这四个项,所以列表是对的,没有改内容。 3. 然后下一句:Across seven datasets and two model families, TCDA certifies reductions in denoiser evaluations by factors of 3.4 on XSum and 2.0 on WMT14 for ELF-B; 8.0 on WMT14 and 4.1 on GSM8K for LLaDA-8B; and 1.33 on OpenWebText for both ELF-M and ELF-L. → 这个用p,关键数据可以加strong?比如

      Across seven datasets and two model families, TCDA certifies reductions in denoiser evaluations by factors of 3.4 on XSum and 2.0 on WMT14 for ELF-B; 8.0 on WMT14 and 4.1 on GSM8K for LLaDA-8B; and 1.33 on OpenWebText for both ELF-M and ELF-L.

      4. 接下来:A validation protocol fixed before generation confirmed every selection on fresh seeds. Across 500 calibration splits, false certification remained below 2% in every setting at a nominal level of 5%. →

      A validation protocol fixed before generation confirmed every selection on fresh seeds. Across 500 calibration splits, false certification remained below 2% in every setting at a nominal level of 5%.

      5. 接下来:Certified schedules vary across tasks and model scales. Trajectory measurements suggest that when decoded tokens commit is more informative for schedule selection than the remaining latent discretization error. →

      Certified schedules vary across tasks and model scales. Trajectory measurements suggest that when decoded tokens commit is more informative for schedule selection than the remaining latent discretization error.

      6. 最后:We also find that the ELF reference sampler draws its time grid at random, so reproducible schedule comparisons for unconditional generation require fixing the grid. →

      We also find that the ELF reference sampler draws its time grid at random, so reproducible schedule comparisons for unconditional generation require fixing the grid.

      等下,有没有需要h2或者h3的地方?原文没有小节,所以不用。有没有引用或者代码?没有,所以不用blockquote或者pre code。现在检查有没有增删改内容?所有信息都保留了,对吧。再核对: 原文的每个内容都在: - 开头的固定调度,最便宜的调度取决于任务模型采样器,提出TCDA,部署规则,给定的输入,用单边自助法界认证,部署最便宜的认证