包裹，长文用
/
划分小节。正文是一个段落，没有明显的小节标题。所以，我应该检查是否有自然段落。原文是一个连续的段落，没有分节。因此，我可能只需要用
包裹整个段落。但规则说“长文用
/
划分小节”，这里没有小节，所以可能不需要标题。或者，如果我认为它是长文，我可以尝试划分，但原文没有明显的标题。我应该保持原样，只包裹在
中。 3. 列表内容用
- ，引用用
，代码用
，重要句子可用。原文中没有列表、引用、代码。重要句子：我可以加到重要部分，但规则说“重要句子可用”，意思是可选的。我需要判断哪些句子重要。原文是一篇摘要，可能整个段落都重要，但我不应该增删内容，所以最好不加，除非明显重要。规则说“可用”，所以不是必须。为了安全，我可以不加，或者加一些。 4. 保留原文所有信息，不得增删改内容。所以，我不能改变文字内容，只添加HTML标签。 5. 不要输出//等文档级标签。只输出片段。 原文是： "Transformer-based large language models (LLMs) impose substantial computation and memory costs as context sizes increase, motivating methods that compress a context into a shorter representation before inference. Existing soft-prompt compression methods typically use the same Transformer backbone as used in the target model, where producing the compressed representation leads to significant latency overheads. This restricts them to the offline setting where context is compressed once and then reused across many queries. We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM) that encodes arbitrary-length context in a single linear-time pass, producing a representation on which we condition a frozen target LLM. The SSM is trained by distillation, matching the next-token logit distribution from the LLM ingesting the full context (as teacher) to the same LLM ingesting the compressed representation (as the student). At equal compression ratios, Boa matches or surpasses the performance of existing state-of-the-art offline compression baselines, while being an order of magnitude faster than Transformer-based implementations, thereby reducing end-to-end latency rather than adding compression time overhead. We demonstrate that Boa also scales well to longer contexts and to compression ratios as aggressive as 16x. By enabling effective compression that is cheap enough to run on every query at inference time, Boa becomes a practical and general tool for efficient long-context inference." 这是一个段落。没有列表、引用、代码。所以，我只需要用包裹整个文本。 但是，规则说“长文用
/
划分小节”。这里没有小节，所以可能不需要标题。或者，我可以视为长文，但原文没有标题，所以最好不加。 重要句子：我可以考虑加到关键部分，比如“Boa Compressor”或一些关键点，但规则说“可用”，意思是可选。为了不改变内容，我应该不加，除非明显重要。用户说“重要句子可用”，所以我可以选择加一些。例如，可能加到“Boa Compressor”或“plug-and-play online context compressor”。但为了安全，我可能不加，因为规则是“可用”，不是必须。 回顾规则：4. 保留原文所有信息，不得增删改内容。所以，添加可能被视为修改，因为它强调了部分内容。但是语义标签，不改变文本内容，只是强调。在HTML中，它表示重要性，但文本内容不变。所以，我认为可以加到重要部分，而不改变信息。 但为了保守，我可以不加任何，因为规则说“可用”，意思是如果需要才用。原文没有指定哪些重要，所以可能不加。 用户是新闻排版工程师，所以我应该根据内容判断。这是一个技术摘要，可能“Boa Compressor”是关键。 但规则强调“不得增删改内容”，所以添加标签是排版，不是修改内容。内容是文本，标签是结构。 我认为可以加到一些关键点，比如“Boa Compressor”和“plug-and-play online context compressor”。 但为了简单，我先尝试不加，只包裹
。 现在，输出HTML片段。 片段应该是：
Transformer-based large language models (LLMs) impose substantial computation and memory costs as context sizes increase, motivating methods that compress a context into a shorter representation before inference. Existing soft-prompt compression methods typically use the same Transformer backbone as used in the target model, where producing the compressed representation leads to significant latency overheads. This restricts them to the offline setting where context is compressed once and then reused across many queries. We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM) that encodes arbitrary-length context in a single linear-time pass, producing a representation on which we condition a frozen target LLM. The SSM is trained by distillation, matching the next-token logit distribution from the LLM ingesting the full context (as teacher) to the same LLM ingesting the compressed representation (as the student). At equal compression ratios, Boa matches or surpasses the performance of existing state-of-the-art offline compression baselines, while being an order of magnitude faster than Transformer-based implementations, thereby reducing end-to-end latency rather than adding compression time overhead. We demonstrate that Boa also scales well to longer contexts and to compression ratios as aggressive as 16x. By enabling effective compression that is cheap enough to run on every query at inference time, Boa becomes a practical and general tool for efficient long-context inference.
但规则说“长文用/
划分小节”。这里是一个段落，没有小节，所以可能不需要。或者，如果我认为它是长文，我可以添加标题，但原文没有标题，所以添加标题会增删内容。 规则4说“不得增删改内容”，所以添加标题会改变内容，因为标题是新内容。 因此，我应该不加标题，只包裹
。 现在，关于：我可以加到“Boa Compressor”和“plug-and-play online context compressor”，因为它们是关键术语。 例如：We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM)... 但规则说“重要句子可用”，是句子，不是单词。所以，可能加到整个句子。 但句子很长。例如，“We propose the Boa Compressor...” 这个句子重要。 但为了不改变，我可以不加。 用户指定“重要句子可用”，所以我应该判断。 在这个摘要中，可能整个段落都重要，但我想加到一些部分。 为了安全，我不加任何，因为规则是“可用”，不是必须。 或者，我可以加到“Boa Compressor”作为关键实体。 在HTML中，通常用于强调，不改变内容。 我认为可以加。 但让我检查规则： “重要句子可用” – 重要句子可以用。 所以，如果我认为某个句子重要，我可以包裹它。 例如，“We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM) that encodes arbitrary-length context in a single linear-time pass, producing a representation on which we condition a frozen target LLM.” 这个句子是核心提议。 或者整个段落。 但为了精确，我加到包含“Boa Compressor”的句子。 在原文中，句子是连续的，没有句号分隔？不，有句号。 原文有句号分隔句子。 让我看： "We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM) that encodes arbitrary-length context in a single linear-time pass, producing a representation on which we condition a frozen target LLM."