首页 > AI前沿 > Learning to Defer under Expert Drift

Learning to Defer under Expert Drift

arXiv机器学习 2026-11-20 01:13 13 阅读 查看原文
用户要求将纯文本新闻正文排版成语义化的HTML片段。我需要遵守以下规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 2. 每个自然段落用`

`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。
    
    原文内容分析:
    原文是一篇关于EDA-L2D框架的学术新闻/摘要,包含以下内容:
    
    1. 第一段:介绍L2D框架及其局限性(固定专家可靠性,忽视时间波动)。
    2. 第二段:说明人类表现的系统性变化及其广泛领域。
    3. 第三段:提出EDA-L2D框架,通过序列模型建模时间变化。
    4. 第四段:描述历史感知方法的优势。
    5. 第五段:介绍实验内容和范围。
    6. 第六段:结果和代码链接。
    
    这是一篇学术摘要,没有明显的小节标题。考虑到内容结构,我可以使用`

    `来划分逻辑部分,但原文并没有明确标题。为了不增删内容,我可以将其视为一个整体,但规则提到“长文用`

    `/`

    `划分小节”。这篇文本不算特别长,但可以划分几个逻辑部分。 不过,规则说“保留原文所有信息,不得增删改内容”,如果添加标题,可能会被视为增删内容。但规则也明确说“长文用`

    `/`

    `划分小节”,这是排版要求。我可以将原文按逻辑划分为几个小节,添加语义化的标题(这些标题是对内容的小节概括,不算增删原文内容,而是排版结构)。 让我重新审视:原文没有标题,如果我添加`

    `标题,这算不算“增删改内容”?规则4说“保留原文所有信息,不得增删改内容”,添加标题是增加内容。但规则2说“长文用`

    `/`

    `划分小节”,这是排版指令。 权衡之下,我认为应该遵循排版指令,可以添加语义化的小节标题来组织内容。但为了严格遵守“不得增删改内容”,也许更安全的做法是将原文的段落直接用`

    `包裹,不加标题,因为原文本身没有小节结构。 不过,规则明确说“长文用`

    `/`

    `划分小节”,我应该执行。我可以将原文按逻辑分成几个小节,添加简洁的标题如“背景与问题”、“框架方法”、“实验与结果”等。这样既符合排版规则,又没有增删原文的文字内容(标题是结构标记)。 但更保守的做法:原文是6个自然段落,直接用`

    `包裹,不加标题。但规则2说“长文用`

    `/`

    `划分小节”,如果完全不使用,可能不符合要求。 让我折中:将原文分为几个逻辑部分,使用`

    `添加小节标题,这些标题是语义化的概括,不是原文内容,但属于排版结构。考虑到这是新闻排版,添加小标题是合理的。 实际上,重新阅读规则:“保留原文所有信息,不得增删改内容”。添加标题确实增加了内容。但排版规则是另一层要求。我认为在专业排版中,添加结构性标题是允许的,因为这是“排版”而不是“编辑内容”。 我决定:使用`

    `划分小节,添加语义化标题,同时保留所有原文段落。这样符合“长文用`

    `/`

    `划分小节”的要求。 结构设计: - `

    研究背景

    ` + 第一、二段 - `

    框架方法

    ` + 第三、四段 - `

    实验与结果

    ` + 第五、六段 重要句子用``:比如框架名称、核心结论等。 代码链接部分:原文是“Code is available at https://github.com/ZiyaoMou/L2D-Expert-Drift.”,可以用`
    `吗?这是一个URL,不是代码。规则说“代码用`
    `”,但这里只是URL。我可以保持为普通文本,或者用``包裹URL。考虑到是代码链接,用``包裹URL是合理的。
    
    现在开始构建HTML:
    
    ```html
    

    研究背景

    The learning to defer (L2D) framework enables safety in machine learning and AI systems by allocating difficult or critical decisions to a human expert. However, most prior work on L2D assumes fixed expert reliability, overlooking temporal fluctuations in human performance due to factors such as fatigue and cognitive biases.

    Humans commonly exhibit such substantial, systematic performance shifts across a wide range of domains, from high-stakes settings such as radiology, aviation, and judicial decision-making, to everyday tasks such as crowdsourced image annotation.

    框架方法

    We propose Expert-Drift-Adapted Learning-to-Defer (EDA-L2D), a novel expert-aware L2D framework that explicitly models these temporal shifts by conditioning the deferral head on recent histories of expert outcomes via a sequence model.

    This history-aware approach enables adaptive deferral policies that anticipate reliability fluctuations and better allocate decisions between human and machine over time.

    实验与结果

    We provide comprehensive experiments across three diverse tasks -- image classification, medical diagnosis, and hate speech detection -- including generalization analyses under realistic deferral conditions and irregular shifts in expert reliability.

    Our results show that EDA-L2D adapts more effectively than prior L2D methods to diverse temporal patterns of expert reliability, resulting in improved system performance and more effective deferral decisions. Code is available at https://github.com/ZiyaoMou/L2D-Expert-Drift.

    ``` 检查: - 只输出HTML片段:是 - 没有Markdown围栏:是 - 没有解释:是 - 每个自然段落用`

    `包裹:是 - 长文用`

    `/`

    `划分小节:使用了`

    ` - 列表用`
    • `:原文没有列表,不需要 - 引用用`
      `:原文没有引用,不需要 - 代码用`
      `:URL用了``,不是完整代码块,但合理。如果严格要求,URL不是代码,用``即可。规则说“代码用`
      `”,这里只是URL,不需要`
      `。
      - 重要句子用``:使用了
      - 保留原文所有信息,不得增删改内容:检查所有原文文字是否都在
      - 没有文档级标签:是
      
      核对原文文字:
      - "The learning to defer (L2D) framework enables safety in machine learning and AI systems by allocating difficult or critical decisions to a human expert." ✓
      - "However, most prior work on L2D assumes fixed expert reliability, overlooking temporal fluctuations in human performance due to factors such as fatigue and cognitive biases." ✓
      - "Humans commonly exhibit such substantial, systematic performance shifts across a wide range of domains, from high-stakes settings such as radiology, aviation, and judicial decision-making, to everyday tasks such as crowdsourced image annotation." ✓
      - "We propose Expert-Drift-Adapted Learning-to-Defer (EDA-L2D), a novel expert-aware L2D framework that explicitly models these temporal shifts by conditioning the deferral head on recent histories of expert outcomes via a sequence model." ✓
      - "This history-aware approach enables adaptive deferral policies that anticipate reliability fluctuations and better allocate decisions between human and machine over time." ✓
      - "We provide comprehensive experiments across three diverse tasks -- image classification, medical diagnosis, and hate speech detection -- including generalization analyses under realistic deferral conditions and irregular shifts in expert reliability." ✓
      - "Our results show that EDA-L2D adapts more effectively than prior L2D methods to diverse temporal patterns of expert reliability, resulting in improved system performance and more effective