`包裹,长文用`
`/`
`划分小节。 3. 列表内容用`
- `,引用用`
`,代码用`
`,重要句子可用``。
4. 保留原文所有信息,不得增删改内容。
5. 不要输出``/``/``等文档级标签。
先分析原文的结构:
原文是一段学术性的新闻/介绍文本,讲述LED框架的提出及其效果。结构上:
- 第一段:介绍DLMs和ROE的局限性。
- 然后提出LED框架,介绍其两个训练设计。
- 然后是实验结果。
- 最后是总结。
原文可以划分为几个小节,比如:
- 背景与问题
- LED框架
- 实验结果
- 总结
但用户说"保留原文所有信息,不得增删改内容",所以不能添加小标题(小标题属于新增内容)。所以用``/`
`划分小节时,需要谨慎。如果原文没有明确的小标题,我不应该强行添加文字。但规则说"长文用`
`/`
`划分小节",这意味着我需要用标题来组织结构,但标题文字应该从原文中提取或者合理概括?规则又说"不得增删改内容"。
这里有冲突。但考虑到"保留原文所有信息,不得增删改内容"是更核心的规则,我应该将原文分段,用`
`包裹,可能用`
`,代码用``,重要句子可用``。 4. 保留原文所有信息,不得增删改内容。 5. 不要输出``/``/``等文档级标签。 先分析原文的结构: 原文是一段学术性的新闻/介绍文本,讲述LED框架的提出及其效果。结构上: - 第一段:介绍DLMs和ROE的局限性。 - 然后提出LED框架,介绍其两个训练设计。 - 然后是实验结果。 - 最后是总结。 原文可以划分为几个小节,比如: - 背景与问题 - LED框架 - 实验结果 - 总结 但用户说"保留原文所有信息,不得增删改内容",所以不能添加小标题(小标题属于新增内容)。所以用``/`
`划分小节时,需要谨慎。如果原文没有明确的小标题,我不应该强行添加文字。但规则说"长文用`
`/`
`划分小节",这意味着我需要用标题来组织结构,但标题文字应该从原文中提取或者合理概括?规则又说"不得增删改内容"。 这里有冲突。但考虑到"保留原文所有信息,不得增删改内容"是更核心的规则,我应该将原文分段,用`
`包裹,可能用`
`/`
`来标注逻辑小节,但标题内容必须来自原文或不添加额外文字。实际上,更安全的做法是:将原文自然分段,用`
`包裹,如果需要用标题,可以用原文中的关键句子作为标题,但这样可能改变结构。 再仔细看规则:"每个自然段落用
包裹,长文用
/
划分小节"。原文其实是一个长段落,但内部有逻辑层次。我可以把它分成几个自然段,然后用`
`/`
`来划分小节。但标题文字如果新增,就违反了"不得增删改内容"。 解决方案:不添加额外的标题文字,而是将原文按逻辑切分成多个自然段落,全部用`
`包裹。如果确实需要标题,可以使用原文中的短语,但原文没有明显的标题式短语。 或者,我可以这样处理:原文本身就是连贯的一段,但为了排版,我可以按语义切分成几个段落,每段用`
`。不需要强加`
`,因为原文没有明确的章节标题。但规则要求"长文用
/
划分小节",这是建议性的,不是强制每个文档都必须有。 考虑到原文是一段学术摘要,最合适的处理方式是:将其切分为几个逻辑段落,用`
`包裹,重要句子用``强调。 让我切分原文: 原文: "Diffusion language models (DLMs) offer a flexible alternative to autoregressive models, enabling any-order infilling. While most existing decoding methods treat decoded tokens as fixed, some recent models allow previously decoded tokens to be revised only by replacing them with alternative tokens, a paradigm we refer to as replacement-only editing (ROE). Although ROE can correct individual tokens, it is inefficient for sequence shifts caused by missing or extra tokens and struggles when errors occur across many positions. To address these limitations, we propose LED (Levenshtein Editing for Diffusion Language Models), a post-training framework that extends ROE with insertion and deletion, enabling Levenshtein-style refinement of generated sequences." 这是第一段,讲背景和问题,引出LED。 "It uses two training designs. First, it reconstructs edit supervision on the fly during supervised fine-tuning using longest-common-subsequence alignment, enabling insertion and deletion learning without manually annotated labels. Second, it adopts multi-stage training, where the state from one stage becomes the input to the next, exposing the model to its earlier errors and enabling it to correct them." 这是第二段,讲训练设计。 "When applied to the 16B-parameter LLaDA2.1-Mini Base, LED raises the average score across twelve benchmarks from 54.86 to 57.10 under confidence-guided decoding while increasing tokens per forward from 6.14 to 7.05, whereas under confidence-free decoding it raises the average score from 27.52 to 47.19. We further show that LED achieves higher pass@k at large k than thresholded decoding without editing, demonstrating improved generation diversity and revealing the diversity potential of DLMs without confidence guidance. LED also improves robustness to large block sizes, raising the average score from 36.97 to 51.34 at block size 128." 这是第三段,讲实验结果。 "Together, these results suggest that LED supports more robust self-correction in DLMs." 这是第四段,总结。 现在考虑是否需要用`
`/``。原文没有小标题,如果我添加,比如`
背景与方法
`,这属于新增内容,违反"不得增删改内容"。所以我认为不添加标题,只使用段落。但规则提到"长文用/
划分小节",我可以用原文中的关键词作为标题,但原文没有。或者我可以不使用标题,只用段落。考虑到"不得增删改内容"是硬性要求,我选择只用`
`,不添加标题。 但用户可能期望看到`
`/`
`的应用。我再想想:如果原文是长文,我可以用原文中的句子作为标题,比如把"To address these limitations, we propose LED"作为小标题?但这是句子,不是标题。
更合理的做法:不添加任何标题,只分段。因为添加标题文字就是新增内容。规则3说"重要句子可用",我可以用``来强调LED的提出、关键结果等。 确定方案:将原文切分为4个段落,用`
`包裹,重要句子用``。不添加标题,因为没有原文小标题,且不能增删内容。 现在,确定哪些句子是"重要句子"可以用``: - "To address these limitations, we propose LED (Levenshtein Editing for Diffusion Language Models), a post-training framework that extends ROE with insertion and deletion, enabling Levenshtein-style refinement of generated sequences." — 这是核心提出,可以加``。 - "It uses two training designs." — 可以加。 - 或者结果部分的关键数字也可以加,但可能过度。选择最核心的。 我可以用``包裹LED提出的句子,以及最后总结的句子。 但注意:``是强调,不是改变内容,所以可以。 现在写出HTML:
Diffusion language models (DLMs) offer a flexible alternative to autoregressive models, enabling any-order infilling. While most existing decoding methods treat decoded tokens as fixed, some recent models allow previously decoded tokens to be revised only by replacing them with alternative tokens, a paradigm we refer to as replacement-only editing (ROE). Although ROE can correct individual tokens, it is inefficient for sequence shifts caused by missing or extra tokens and struggles when errors occur across many positions. To address these limitations, we propose LED (Levenshtein Editing for Diffusion Language Models), a post-training framework that extends ROE with insertion and deletion, enabling Levenshtein-style refinement of generated sequences.
It uses two training designs. First, it reconstructs edit supervision on the fly during supervised fine-tuning using longest-common-subsequence alignment, enabling insertion and deletion learning without manually annotated labels. Second, it adopts multi-stage training, where the state from one stage becomes the input to the next, exposing the model to its earlier errors and enabling it to correct them.
When applied to the 16B-parameter LLaDA2.1-Mini Base, LED raises the average score across twelve benchmarks from 54.86 to 57.10 under confidence-guided decoding while increasing tokens
`包裹,重要句子用``。不添加标题,因为没有原文小标题,且不能增删内容。 现在,确定哪些句子是"重要句子"可以用``: - "To address these limitations, we propose LED (Levenshtein Editing for Diffusion Language Models), a post-training framework that extends ROE with insertion and deletion, enabling Levenshtein-style refinement of generated sequences." — 这是核心提出,可以加``。 - "It uses two training designs." — 可以加。 - 或者结果部分的关键数字也可以加,但可能过度。选择最核心的。 我可以用``包裹LED提出的句子,以及最后总结的句子。 但注意:``是强调,不是改变内容,所以可以。 现在写出HTML:
Diffusion language models (DLMs) offer a flexible alternative to autoregressive models, enabling any-order infilling. While most existing decoding methods treat decoded tokens as fixed, some recent models allow previously decoded tokens to be revised only by replacing them with alternative tokens, a paradigm we refer to as replacement-only editing (ROE). Although ROE can correct individual tokens, it is inefficient for sequence shifts caused by missing or extra tokens and struggles when errors occur across many positions. To address these limitations, we propose LED (Levenshtein Editing for Diffusion Language Models), a post-training framework that extends ROE with insertion and deletion, enabling Levenshtein-style refinement of generated sequences.
It uses two training designs. First, it reconstructs edit supervision on the fly during supervised fine-tuning using longest-common-subsequence alignment, enabling insertion and deletion learning without manually annotated labels. Second, it adopts multi-stage training, where the state from one stage becomes the input to the next, exposing the model to its earlier errors and enabling it to correct them.
When applied to the 16B-parameter LLaDA2.1-Mini Base, LED raises the average score across twelve benchmarks from 54.86 to 57.10 under confidence-guided decoding while increasing tokens