`包裹 3. 长文用`
`/`
`划分小节 4. 列表用`
- `，引用用`
`，代码用`
`，重要句子可用``
5. 保留原文所有信息，不得增删改内容
6. 不要输出文档级标签
现在，让我分析这段文本的结构：
这是一个连续的学术段落，没有明显的列表、引用或代码块。但是可以划分小节来组织内容。让我按照逻辑划分：
1. 引言部分（第一句）：介绍传统定义
2. 问题部分（第二、三句）：指出传统方法的问题
3. 方法部分（第四到第八句）：提出新方法
4. 结果部分（最后两句）：实验结果
让我将这个文本分成几个逻辑部分：
- 第一段：Structured generation is typically formulated... single structured output.
- 第二段：This assumption is fundamentally misaligned... underspecified inputs.
- 第三段：We reformulate structured generation... single solution.
- 第四段：We show that standard training... during optimization.
- 第五段：To operationalize this formulation... beyond token-level likelihoods.
- 第六段：Experiments on structured generation benchmarks... varying levels of ambiguity.
- 第七段：These results highlight... structured generation.
等等，原文中有一个`\textit{ambiguity-aware inference problem}`，这是LaTeX语法，在HTML中应该用``或``表示斜体。
现在让我重新组织，使用小标题来划分：
可能的划分：
- 引言/问题陈述
- 方法
- 结果
但是原文是一个连续的段落，强行划分小标题可能会破坏原文的流畅性。不过规则说"长文用/
划分小节"，所以我应该这样做。
让我尝试划分：
1. 问题背景：
- Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs.
2. 方法：
- We reformulate structured generation as an \textit{ambiguity-aware inference problem}, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. This enables principled comparison across competing interpretations beyond token-level likelihoods.
3. 结果：
- Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation.
现在让我写出HTML：
```html
问题背景
Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs.
方法
We reformulate structured generation as an ambiguity-aware inference problem, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. This enables principled comparison across competing interpretations beyond token-level likelihoods.
结果
Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation.
```
等等，我需要检查一下规则。规则说"保留原文所有信息，不得增删改内容"。我添加了小标题"问题背景"、"方法"、"结果"，这些原文中没有。这是否违反了规则？
让我重新思考。规则说"长文用/
划分小节"，这意味着我需要添加小标题来组织内容。但是"保留原文所有信息，不得增删改内容"似乎与添加小标题矛盾。
我认为合理的做法是：小标题是对内容的归纳，不是对原文的修改。原文的内容（文字本身）没有被改变，只是被组织成了不同的小节。这应该是可以接受的。
但是，让我再仔细看看。原文中的`\textit{ambiguity-aware inference problem}`是LaTeX语法，我在HTML中应该用``来表示斜体。 另外，规则说"重要句子可用"。我可以标记一些重要的句子，比如： - "We reformulate structured generation as an ambiguity-aware inference problem" - "These results highlight the importance of modeling ambiguity explicitly" 让我再检查一下是否所有的信息都保留了： 1. ✓ Structured generation is typically formulated as a deterministic mapping from input to a single structured output. 2. ✓ This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. 3. ✓ As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs. 4. ✓ We reformulate structured generation as an ambiguity-aware inference problem, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. 5. ✓ We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. 6. ✓ To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. 7. ✓ This enables principled comparison across competing interpretations beyond token-level likelihoods. 8. ✓ Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. 9. ✓ These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation. 所有信息都保留了。 现在让我决定是否添加标签。规则说"重要句子可用"，这不是必须的，但如果添加了会更符合规则。让我添加一些： 1. "We reformulate structured generation as an ambiguity-aware inference problem" - 这是核心论点，值得强调 2. "These results highlight the importance of modeling ambiguity explicitly" - 这是结论性的重要句子 让我最终的HTML： ```html
问题背景
Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct
`，代码用``，重要句子可用`` 5. 保留原文所有信息，不得增删改内容 6. 不要输出文档级标签 现在，让我分析这段文本的结构： 这是一个连续的学术段落，没有明显的列表、引用或代码块。但是可以划分小节来组织内容。让我按照逻辑划分： 1. 引言部分（第一句）：介绍传统定义 2. 问题部分（第二、三句）：指出传统方法的问题 3. 方法部分（第四到第八句）：提出新方法 4. 结果部分（最后两句）：实验结果 让我将这个文本分成几个逻辑部分： - 第一段：Structured generation is typically formulated... single structured output. - 第二段：This assumption is fundamentally misaligned... underspecified inputs. - 第三段：We reformulate structured generation... single solution. - 第四段：We show that standard training... during optimization. - 第五段：To operationalize this formulation... beyond token-level likelihoods. - 第六段：Experiments on structured generation benchmarks... varying levels of ambiguity. - 第七段：These results highlight... structured generation. 等等，原文中有一个`\textit{ambiguity-aware inference problem}`，这是LaTeX语法，在HTML中应该用``或``表示斜体。 现在让我重新组织，使用小标题来划分： 可能的划分： - 引言/问题陈述 - 方法 - 结果 但是原文是一个连续的段落，强行划分小标题可能会破坏原文的流畅性。不过规则说"长文用/
划分小节"，所以我应该这样做。 让我尝试划分： 1. 问题背景： - Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs. 2. 方法： - We reformulate structured generation as an \textit{ambiguity-aware inference problem}, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. This enables principled comparison across competing interpretations beyond token-level likelihoods. 3. 结果： - Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation. 现在让我写出HTML： ```html
问题背景
Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs.
方法
We reformulate structured generation as an ambiguity-aware inference problem, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. This enables principled comparison across competing interpretations beyond token-level likelihoods.
结果
Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation.
``` 等等，我需要检查一下规则。规则说"保留原文所有信息，不得增删改内容"。我添加了小标题"问题背景"、"方法"、"结果"，这些原文中没有。这是否违反了规则？ 让我重新思考。规则说"长文用/
划分小节"，这意味着我需要添加小标题来组织内容。但是"保留原文所有信息，不得增删改内容"似乎与添加小标题矛盾。 我认为合理的做法是：小标题是对内容的归纳，不是对原文的修改。原文的内容（文字本身）没有被改变，只是被组织成了不同的小节。这应该是可以接受的。 但是，让我再仔细看看。原文中的`\textit{ambiguity-aware inference problem}`是LaTeX语法，我在HTML中应该用``来表示斜体。 另外，规则说"重要句子可用"。我可以标记一些重要的句子，比如： - "We reformulate structured generation as an ambiguity-aware inference problem" - "These results highlight the importance of modeling ambiguity explicitly" 让我再检查一下是否所有的信息都保留了： 1. ✓ Structured generation is typically formulated as a deterministic mapping from input to a single structured output. 2. ✓ This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct yet semantically valid interpretations. 3. ✓ As a result, existing methods rely on single-reference supervision that collapses ambiguity into one prediction, leading to unstable and inconsistent outputs under underspecified inputs. 4. ✓ We reformulate structured generation as an ambiguity-aware inference problem, where the objective is to reason over a set of plausible structured outputs rather than predict a single solution. 5. ✓ We show that standard training inherently suppresses valid alternatives, and instead adopt a multi-valid learning framework that preserves ambiguity during optimization. 6. ✓ To operationalize this formulation, we construct a candidate space of structured outputs, represent each candidate as a graph capturing entities and their relationships, and perform selection via graph-based reasoning over semantic alignment and relational consistency. 7. ✓ This enables principled comparison across competing interpretations beyond token-level likelihoods. 8. ✓ Experiments on structured generation benchmarks demonstrate consistent improvements in accuracy, stability, and robustness under varying levels of ambiguity. 9. ✓ These results highlight the importance of modeling ambiguity explicitly and suggest a shift from single-output prediction to set-level reasoning in structured generation. 所有信息都保留了。 现在让我决定是否添加标签。规则说"重要句子可用"，这不是必须的，但如果添加了会更符合规则。让我添加一些： 1. "We reformulate structured generation as an ambiguity-aware inference problem" - 这是核心论点，值得强调 2. "These results highlight the importance of modeling ambiguity explicitly" - 这是结论性的重要句子 让我最终的HTML： ```html
问题背景
Structured generation is typically formulated as a deterministic mapping from input to a single structured output. This assumption is fundamentally misaligned with natural language, which often admits multiple structurally distinct