包裹，长文用
/
划分小节；列表内容用
- ，引用用
，代码用
，重要句子可用；保留原文所有信息，不得增删改内容；不要输出//等文档级标签。
正文如下：
Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data. The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness. In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates. We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome. Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings. Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly. Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional. Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.
首先，我需要分析这段正文的结构。看起来它是一段连续的学术摘要或介绍，没有明显的小标题或列表。不过，根据规则，我可以用或
来划分小节。可能需要按主题分节。比如，开头介绍activation steering，然后讲方法框架，再讲估计器，最后是实验。
需要检查是否有列表内容，这里似乎没有。引用或代码也没有。重要句子可能需要加，比如关键方法或结论。
用户可能希望这段排版成结构清晰的HTML片段，适合嵌入到新闻网站或学术页面中。需要确保每个自然段落用包裹，并且适当使用标题划分小节。
现在，我需要将原文分成几个部分，可能按以下结构：
1. 引言：介绍activation steering及其应用。
2. 方法论：区分estimand和estimator，将现有方法归类为plug-in estimators，以及扩展方法。
3. 估计器讨论：讨论灵活模型可能存在的问题，引入Stein score估计器，以及正交化估计器。
4. 实验应用：说明在基准数据集上的应用。
需要确保每个段落都用
包裹，并且标题用
或
。例如，可能将“引言”作为
，然后每个主要部分作为
，但原文中并没有明确的标题，所以可能需要自己添加合适的标题，或者不添加标题，直接按段落划分。
但用户可能希望保持原文结构，所以可能需要不添加额外标题，而是将原文分成几个段落，每个段落用
包裹。不过，原文是一段连续的文字，可能需要拆分成多个段落。 需要检查原文是否有多个自然段落。看起来原文是一段，但可以按逻辑拆分成几个部分。例如： - 第一句到“without a target concept, such as truthfulness.”：介绍activation steering及其标准方法。 - 接下来的部分：“In this paper, we separate...” 到“with the binary contrastive label fixed as the outcome.”：论文的主要贡献，区分estimand和estimator，并将现有方法框架化。 - 然后：“Using this framing...” 到“such as log-probabilities or oracle ratings.”：扩展方法到连续结果。 - 接下来：“Flexible ML models...” 到“target the gradient directly.”：讨论灵活模型的问题，引入直接估计梯度的动机。 - 然后：“Drawing on the debiased...” 到“the Riesz representer of the gradient functional.”：开发基于Stein score的估计器。 - 最后：“Combining the two routes...” 到“applied to a number of benchmark datasets.”：结合两种方法得到正交化估计器，并应用到基准数据集。 需要将这些部分拆分成段落，每个用
包裹。可能需要添加小标题，但用户没有明确要求，所以可能不需要。不过，根据规则，长文可以用
/
划分小节。所以可能需要添加小标题来划分。
但原文中没有标题，所以可能需要自己创建合适的标题，或者不添加标题。需要确保不增删改内容，所以可能不能添加新标题。因此，可能只能将原文分成多个段落，每个用
包裹，而不添加标题。 需要检查是否有重要句子需要加。例如，“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression”可能是一个重要句子，可以加。但需要确保不改变内容，只是强调。 现在，开始处理原文： 原文第一句：“Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data.” 这是一个独立的句子，可以作为第一段。 接下来：“The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness.” 这可以作为第二段。 然后：“In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates.” 这是第三段。 接下来：“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome.” 这里可能需要拆分成两个句子，但原文是一个句子，所以保留原样。 然后：“Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings.” 这是另一个段落。 接下来：“Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly.” 这是下一个段落。 然后：“Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional.” 这是下一个段落。 最后：“Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.” 这是最后一段。 现在，需要将这些段落用
包裹。需要确保每个自然段落独立，所以可能需要根据句子之间的逻辑进行拆分。 现在，开始构建HTML片段：
Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data.
The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness.
In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates.
We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome.
Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings.
Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly.
Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional.
Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.
现在，检查是否需要添加小标题。根据规则，长文可以用
/
划分小节。但原文中没有明确的标题，所以可能需要自己添加，但用户要求不增删改内容，所以可能不能添加新标题。因此，可能只能保持段落结构，不添加标题。 另外，需要检查是否有重要句子需要加。例如，“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression”可能可以加，但需要确保不改变内容。 最后，确保没有使用文档级标签，只输出HTML片段本身。 所以，最终的HTML片段应该是：
Activation steering is a
，代码用，重要句子可用；保留原文所有信息，不得增删改内容；不要输出//等文档级标签。 正文如下： Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data. The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness. In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates. We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome. Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings. Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly. Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional. Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets. 首先，我需要分析这段正文的结构。看起来它是一段连续的学术摘要或介绍，没有明显的小标题或列表。不过，根据规则，我可以用或
来划分小节。可能需要按主题分节。比如，开头介绍activation steering，然后讲方法框架，再讲估计器，最后是实验。 需要检查是否有列表内容，这里似乎没有。引用或代码也没有。重要句子可能需要加，比如关键方法或结论。 用户可能希望这段排版成结构清晰的HTML片段，适合嵌入到新闻网站或学术页面中。需要确保每个自然段落用
包裹，并且适当使用标题划分小节。 现在，我需要将原文分成几个部分，可能按以下结构： 1. 引言：介绍activation steering及其应用。 2. 方法论：区分estimand和estimator，将现有方法归类为plug-in estimators，以及扩展方法。 3. 估计器讨论：讨论灵活模型可能存在的问题，引入Stein score估计器，以及正交化估计器。 4. 实验应用：说明在基准数据集上的应用。 需要确保每个段落都用
包裹，并且标题用
或
。例如，可能将“引言”作为
，然后每个主要部分作为
，但原文中并没有明确的标题，所以可能需要自己添加合适的标题，或者不添加标题，直接按段落划分。 但用户可能希望保持原文结构，所以可能需要不添加额外标题，而是将原文分成几个段落，每个段落用
包裹。不过，原文是一段连续的文字，可能需要拆分成多个段落。 需要检查原文是否有多个自然段落。看起来原文是一段，但可以按逻辑拆分成几个部分。例如： - 第一句到“without a target concept, such as truthfulness.”：介绍activation steering及其标准方法。 - 接下来的部分：“In this paper, we separate...” 到“with the binary contrastive label fixed as the outcome.”：论文的主要贡献，区分estimand和estimator，并将现有方法框架化。 - 然后：“Using this framing...” 到“such as log-probabilities or oracle ratings.”：扩展方法到连续结果。 - 接下来：“Flexible ML models...” 到“target the gradient directly.”：讨论灵活模型的问题，引入直接估计梯度的动机。 - 然后：“Drawing on the debiased...” 到“the Riesz representer of the gradient functional.”：开发基于Stein score的估计器。 - 最后：“Combining the two routes...” 到“applied to a number of benchmark datasets.”：结合两种方法得到正交化估计器，并应用到基准数据集。 需要将这些部分拆分成段落，每个用
包裹。可能需要添加小标题，但用户没有明确要求，所以可能不需要。不过，根据规则，长文可以用
/
划分小节。所以可能需要添加小标题来划分。 但原文中没有标题，所以可能需要自己创建合适的标题，或者不添加标题。需要确保不增删改内容，所以可能不能添加新标题。因此，可能只能将原文分成多个段落，每个用
包裹，而不添加标题。 需要检查是否有重要句子需要加。例如，“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression”可能是一个重要句子，可以加。但需要确保不改变内容，只是强调。 现在，开始处理原文： 原文第一句：“Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data.” 这是一个独立的句子，可以作为第一段。 接下来：“The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness.” 这可以作为第二段。 然后：“In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates.” 这是第三段。 接下来：“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome.” 这里可能需要拆分成两个句子，但原文是一个句子，所以保留原样。 然后：“Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings.” 这是另一个段落。 接下来：“Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly.” 这是下一个段落。 然后：“Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional.” 这是下一个段落。 最后：“Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.” 这是最后一段。 现在，需要将这些段落用
包裹。需要确保每个自然段落独立，所以可能需要根据句子之间的逻辑进行拆分。 现在，开始构建HTML片段：
Activation steering is a popular approach for shifting LLM output toward a desired behavior, and increasingly for generating realistic synthetic data.
The standard recipe estimates a steering direction by contrasting LLM activations on prompts with and without a target concept, such as truthfulness.
In this paper, we separate the estimand (the steering direction these methods target) from the estimator used to recover it, two notions the literature often conflates.
We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression, with the binary contrastive label fixed as the outcome.
Using this framing, we can immediately incorporate flexible modeling approaches and extend existing methods to allow for continuous outcomes such as log-probabilities or oracle ratings.
Flexible ML models that fit the outcome well, however, might not estimate the gradient well, which motivates estimators that target the gradient directly.
Drawing on the debiased machine learning literature, we develop a complementary estimator built on the Stein score, the Riesz representer of the gradient functional.
Combining the two routes gives an orthogonalized estimator with classical double-robustness properties. We illustrate the framework on activation steering applied to a number of benchmark datasets.
现在，检查是否需要添加小标题。根据规则，长文可以用/
划分小节。但原文中没有明确的标题，所以可能需要自己添加，但用户要求不增删改内容，所以可能不能添加新标题。因此，可能只能保持段落结构，不添加标题。 另外，需要检查是否有重要句子需要加。例如，“We frame nearly all existing methods as plug-in estimators of the average gradient of an outcome regression”可能可以加，但需要确保不改变内容。 最后，确保没有使用文档级标签，只输出HTML片段本身。 所以，最终的HTML片段应该是：
Activation steering is a