相关但不完整：指称悬空作为硬提示压缩中的范式级失败模式

We need to translate markdown content from English to Simplified Chinese, preserving markdown formatting, links, etc. Need produce only translation.

Let's parse content. It's an academic paper abstract/sections. Need translate into Chinese.

Important: preserve headings, lists, emphasis, inline code, links. URLs unchanged. Math? There is some inline math using LaTeX-like notation (e.g., 0.300.30, 3434 to 54%54%). Actually the source text has weird duplication due to formatting? It says "compression ratio0.300.30,Beaver, which ranks coherent chunks using Qwen3-0.6B embeddings, leaves the answer path incomplete in3434to54%54\\%of bridge examples" This seems like numbers repeated because of arXiv HTML conversion. We should translate naturally, likely "0.30" and "34% to 54%". Need infer.

Let's examine sections.

Abstract:
"Hard prompt compression reduces long-context inference cost by scoring tokens, sentences, or chunks independently and retaining the highest-scoring units under a budget. We identify a structural failure in this procedure: independent selection can split dependent evidence pairs, retaining one member while deleting the other. When the retained text span contains an answer but the deleted span defines the entity needed to interpret it, we call the result referential dangling. At compression ratio 0.30, Beaver, which ranks coherent chunks using Qwen3-0.6B embeddings, leaves the answer path incomplete in 34% to 54% of bridge examples across three multi-hop question answering (QA) datasets. The failure is not specific to that implementation: on a shared HotpotQA bridge set, all six hard compressors we test exhibit dangling at rates reaching 60%, and every document in LongBench-v2 Single-Document QA contains at least one dangling reference. We then test whether selecting different content at the same budget helps. On dangling examples evaluated with Qwen3-8B, reinserting the missing supporting paragraph and offsetting its tokens by removing paragraphs not annotated as supporting the answer improves accuracy by 29 to 34 points (p<10^-4), recovering at least 88% of the gap to contexts that retain both annotated supporting paragraphs. Stronger answer models do not absorb the loss: on MuSiQue, GPT-5.5 is 8.8 points less accurate on the compressed contexts than on contexts retaining both supporting paragraphs. Finally, we train a compact classifier to rank sentences the compressor omitted by whether they are needed to interpret retained text, and reinsert the top-ranked candidates without using support annotations at inference. On HotpotQA with Qwen3-8B as the downstream answer model, this automatic restoration improves accuracy by 4.7 points while changing the compression ratio only from 0.30 to 0.31. Hard compressors should optimize relevance and referential completeness."

Need translate.

Section 1 Introduction. Includes figure caption. "Refer to caption Figure 1: Referential dangling with a missing bridge. ..."

Need preserve.

Let's translate all.

We need be careful with "hard prompt compression" -> "硬提示压缩" or "硬性提示压缩"? "Hard prompt compression" means select/discard tokens (discrete) as opposed to soft. Chinese term maybe "硬提示压缩". We can use "硬性提示压缩" or "硬提示压缩". Let's use "硬提示压缩".

"referential dangling" -> "指称悬空" (referential = 指称, dangling = 悬空). Could also "指代悬空" but "referential" includes referring expressions. Let's use "指称悬空".

"bridge examples" -> "桥梁示例" or "桥接示例". In multi-hop QA, "bridge" is a type where connect entities. Use "桥接示例".

"answer path incomplete" -> "答案路径不完整".

"coherent chunks" -> "连贯块".

"support annotations" -> "支持标注".

"paradigm-level failure mode" -> "范 paradigm-level failure mode"? Title: "Referential Dangling as a Paradigm-Level Failure Mode in Hard Prompt Compression" -> "硬提示压缩中的指称悬空：一种范式级失败模式".

Need translate sections.

Section 2 Related Work.

Need preserve citations links.

Let's translate carefully.

Potential issue: The source text has weird "3434to54%54\\%" due to LaTeX. We'll fix numbers.

Let's write output with markdown preserved.

Need include all text. Let's translate.

We'll produce final.

Let's start:

# 硬提示压缩中的指称悬空：一种范式级失败模式

Source: https://arxiv.org/html/2608.04569

Zhengpei Hu1\equalcontrib, Kai Li2\equalcontrib, Dapeng Fu3, Xuechao Zou2, Yuanhao Tang1, Yue Li1, Tengfei Cao1, Jianqiang Huang1\corresponding

Need preserve author line? Yes, translate nothing but preserve \equalcontrib etc. Actually "\\equalcontrib" is LaTeX. Keep as is? Since markdown, we can keep. The instruction says preserve markdown. We'll include the line.

###### Abstract -> ###### 摘要

Then translated.

Need use Chinese punctuation. Keep citations as links.

Let's translate all paragraphs.

Important math: "0.300.30" -> "0.30". "3434" -> "34", "54%54\\%" -> "54%". "p<10−4p<10^\{\-4\}" -> "p<10^{-4}"? We can write "p<10^-4" or "p<10^{-4}". Preserve inline math? There's no $, just raw. We'll use "p<10^{-4}" or "p<10^-4". Better "p<10^{-4}" because it's LaTeX-ish.

"2929to3434points" -> "29到34个百分点". "8\.88\.8points" -> "8.8个百分点". "4\.74\.7points" -> "4.7个百分点".

"0\.300\.30to0\.310\.31" -> "0.30到0.31".

"88%88\\%" -> "88%".

Need preserve URLs exactly.

Let's translate.

Section 1:

"Refer to captionFigure 1:Referential dangling with a missing bridge\. Independent scoring retains the query subject and the answer string but removes the fact that Tim DuBois was born in Southwest City, leaving the inference chain incomplete\.In recent frontier models..."

Need note "Refer to caption" likely "Figure 1 caption". We'll render as "图1："

Caption text should be in italics? Markdown maybe "Figure 1: ..." We'll use "图1：...".

Then paragraphs.

Let's do:

# 1 引言

"最近的前沿模型..." etc.

Need translate "In recent frontier models, the context windows of large language models (LLMs) have expanded from a few thousand tokens to the million-token scale." -> "在最近的前沿模型中，大语言模型（LLM）的上下文窗口已从数千个token扩展到百万token量级。"

Need maintain "token" as is.

"Yet long contexts remain expensive: the quadratic cost of self-attention inflates prefill latency, and models underuse relevant content when it appears away from prompt boundaries (Liu et al. 2024)." -> "但长上下文仍然代价高昂：自注意力的二次方成本推高了预填充延迟，而相关内容远离提示边界时模型会利用不足（Liu等人，2024）。"

Need "prefill latency" -> "预填充延迟".

"Prompt compression addresses this tension by pruning the input before it reaches the target model while trying to preserve the information needed for the task (Jiang et al. 2023b; Pan et al. 2024; Li et al. 2023)." -> "提示压缩通过在输入到达目标模型之前对其进行剪枝，同时尽量保留任务所需信息，来缓解这一矛盾（Jiang等人，2023b；Pan等人，2024；Li等人，2023）。"

Then compressors paragraph.

"Many hard prompt compressors follow a common selection procedure: they score each token, sentence, or chunk for importance and retain the highest scoring fragments within a token budget." -> "许多硬提示压缩器遵循一种常见的选择程序：对每个token、句子或块的重要性进行评分，并在token预算内保留得分最高的片段。"

"These scores may reflect query relevance or perplexity (Jiang et al. 2024), embedding similarity (Hu et al. 2026), information content or learned classifier scores (Li et al. 2023; Pan et al. 2024), syntactic salience (Mao et al. 2025), or importance derived from attention (Zhao et al. 2025)." -> "这些分数可能反映查询相关性或困惑度（Jiang等人，2024）、嵌入相似度（Hu等人，2026）、信息量或学习到的分类器分数（Li等人，2023；Pan等人，2024）、句法显著性（Mao等人，2025），或从注意力推导出的重要性（Zhao等人，2025）。"

"Although these methods differ in architecture and supervision, they share the objective of retaining fragments that appear important. This objective assumes that the selected fragments collectively form a usable prompt. However, an important fragment need not be self-contained, because its meaning may depend on an earlier definition, an antecedent, or a bridge fact that the compressor removes independently." -> "尽管这些方法在架构和监督信号上有所不同，它们都共同追求保留看似重要的片段。这一目标假设所选片段共同构成一个可用的提示。然而，重要的片段未必是自包含的，因为其含义可能依赖于前文定义、先行词或桥接事实，而压缩器可能独立地删除这些依赖。"

Then formalize.

"We call this failure referential dangling and formalize it in Section 3. ..." -> "我们将这种失败称为指称悬空，并在第3节中形式化定义它。"

"Figure 1 illustrates this mechanism. ..." -> "图1展示了这一机制。"

Need translate.

Then "It remains unclear whether dangling is systematic, fixed-budget reselection recovers accuracy, or restoration can be automated." -> "目前尚不清楚指称悬空是否具有系统性、固定预算下的重新选择能否恢复准确率，以及恢复过程能否自动化。"

Then experiments.

"We first examine the official implementation of Beaver (Hu et al. 2026). At compression ratio 0.30, it leaves answer paths incomplete in 34% to 54% of bridge examples across HotpotQA (Yang et al. 2018), 2WikiMultiHopQA (Ho et al. 2020), and MuSiQue (Trivedi et al. 2022). A human audit finds 95% precision." -> "我们首先检验Beaver（Hu等人，2026）的官方实现。在0.30的压缩比下，它在HotpotQA（Yang等人，2018）、2WikiMultiHopQA（Ho等人，2020）和MuSiQue（Trivedi等人，2022）中分别导致34%至54%的桥接示例答案路径不完整。人工审计发现95%的准确率。"

Note "finds 95% precision" means human audit confirms our detection? Actually "finds 95% precision" means the audit found the dangling annotations have 95% precision. Translate "人工审计的查准率为95%。"

"The diagnosis extends beyond this setting: all six compressors tested exhibit dangling at rates from 32% to 60% on the shared bridge set, and all 80 documents in LongBench-v2 Single-Document QA (Bai et al. 2025) contain at least one dangling reference." -> "这一诊断并不限于该实现：在共享桥接集合上，所有六个被测试的压缩器的指称悬空率都在32%到60%之间；LongBench-v2单文档问答（Bai等人，2025）中的所有80篇文档都至少包含一处指称悬空引用。"

Then "We next conduct an annotation-guided reselection experiment..." -> "接下来，我们进行了一项由标注指导的重新选择实验：将遗漏的支持段落重新插入，并通过删除未被标注为支持答案的低分段落来抵消其token数。在Qwen3-8B评估的悬空示例上，这一干预将准确率提高了29到34个百分点（p<10^{-4}），且不增加token预算，并恢复了原始压缩上下文与保留两个标注支持段落的上下文之间准确率差距的至少88%。"

Then "Finally, we generate candidate sentences from the omitted text and train a compact classifier..." -> "最后，我们从被省略的文本中生成候选句子，并训练一个紧凑的分类器，根据是否需要用于解释保留文本来对它们排序。在推理时不使用支持段落标注，该方法重新插入排名最高的候选。在HotpotQA上以Qwen3-8B作为下游答案模型，这种自动恢复将准确率提高了4.7个百分点，而压缩比仅从0.30变为0.31。硬压缩器应当同时优化相关性和指称完整性。"

Need "released the code..." -> "我们已发布代码、单个示例结果和训练好的模型" + link.

Now Section 2.

"#### Prompt and context compression." -> "#### 提示与上下文压缩。"

Then paragraph. Need translate.

"Prompt compression reduces long-context inference cost by shortening the input before it reaches the target model. It complements model-side efficiency methods, including elastic subnetworks, compact discrete semantic tokens for multimodal inputs, and efficient encoder–decoder or state-space architectures (Li and Luo 2024; Li et al. 2025b, 2022, a)." -> "提示压缩通过在输入到达目标模型之前缩短输入来降低长上下文推理成本。它与模型侧效率方法互补，包括弹性子网络、用于多模态输入的紧凑离散语义token，以及高效的编码器-解码器或状态空间架构（Li和Luo，2024；Li等人，2025b，2022，a）。"

Need decide "elastic subnetworks" -> "弹性子网络". "compact discrete semantic tokens" -> "紧凑离散语义token".

"We study hard compression, which selects tokens, sentences, chunks, or parse nodes under a budget using self-information or perplexity (Li et al. 2023; Jiang et al. 2024), learned token classifiers (Jiang et al. 2023b; Pan et al. 2024), embedding similarity (Hu et al. 2026), syntactic salience (Mao et al. 2025), attention-based signals (Zhao et al. 2025), or training-free sentence and fragment selection (Tang et al. 2025)." -> "我们研究硬压缩，即在预算下使用自信息或困惑度（Li等人，2023；Jiang等人，2024）、学习型token分类器（Jiang等人，2023b；Pan等人，2024）、嵌入相似度（Hu等人，2026）、句法显著性（Mao等人，2025）、基于注意力的信号（Zhao等人，2025）或无需训练的句子和片段选择（Tang等人，2025）来选择token、句子、块或解析节点。"

"Task-aware selectors add reinforcement-learning rewards or key-information density objectives (Shandilya et al. 2025; Lin et al. 2025), whereas soft or latent approaches use gist tokens, autoencoders, semantic source coding, or learned special-token representations (Mu et al. 2023; Chevalier et al. 2023; Ge et al. 2023; Fei et al. 2024; Li et al. 2025d)." -> "任务感知选择器会加入强化学习奖励或关键信息密度目标（Shandilya等人，2025；Lin等人，2025），而软压缩或潜在方法则使用要点token、自编码器、语义源编码或学习到的特殊token表示（Mu等人，2023；Chevalier等人，2023；Ge等人，2023；Fei等人，2024；Li等人，2025d）。"

"A further line of work compresses hidden activations or KV caches instead of source text (Zhang et al. 2025a, 2023); because these methods never produce a reduced text sequence, the failure we study does not arise in the same form and we do not evaluate them." -> "另一类工作压缩隐藏激活或KV缓存，而不是源文本（Zhang等人，2025a，2023）；由于这些方法从不产生缩减的文本序列，我们研究的失败不会以相同形式出现，因此我们不评估它们。"

"Hard methods differ in supervision, granularity, and query access but select fragments by salience or relevance. Existing surveys organize prompt-compression methods by hard versus soft strategies and discuss their downstream adaptations (Li et al. 2025c)." -> "硬压缩方法在监督信号、粒度和查询访问方面有所不同，但都通过显著性或相关性选择片段。现有综述按硬策略与软策略组织提示压缩方法，并讨论其下游适配（Li等人，2025c）。"

"Comparative studies further show that downstream performance and information preservation vary with the compression method, task, and compression setting (Jha et al. 2024; Łajewska et al. 2025), while rate-distortion analysis formalizes the budget–performance trade-off and the role of query-aware selection (Nagle et al. 2024)." -> "比较研究进一步表明，下游性能和信息保持随压缩方法、任务和压缩设置而变化（Jha等人，2024；Łajewska等人，2025），而率失真分析形式化了预算-性能权衡以及查询感知选择的作用（Nagle等人，2024）。"

"We study how independent hard selection can retain a fragment but remove its required definition or bridge." -> "我们研究独立硬选择如何可能保留一个片段，却移除其所需的定义或桥接信息。"

"#### Dependency-preserving selection." -> "#### 保持依赖的选择。"

"Extractive summarization identified this dependency and built a constraint for it: Durrett et al. (2016) add anaphoricity constraints to the selection objective, so a sentence may not be extracted unless the text its pronouns depend on is extracted with it." -> "抽取式摘要已识别出这种依赖性并为其构建了约束：Durrett等人（2016）在选择目标中加入了回指性约束，即一个句子的代词所依赖的文本如果没有被同时抽取，则该句子不能被抽取。"

"Hard prompt compressors inherit the selection problem without the constraint: they score units by salience or query relevance and keep the highest-scoring units under a budget, with no term rewarding the joint retention of a unit and the text needed to interpret it." -> "硬提示压缩器继承了选择问题，却没有这一约束：它们按显著性或查询相关性对单元评分，并在预算下保留得分最高的单元，没有任何一项奖励单元及其解释所需文本的共同保留。"

"Other work approaches the dependency from different angles. Coreference resolution identifies anaphoric links (Lee et al. 2017), but on complete documents; resolving links in the source does not indicate whether the retained subset stays interpretable once the antecedent is deleted." -> "其他工作从不同角度处理这一依赖。共指消解识别回指链接（Lee等人，2017），但这是在完整文档上；在源文本中解析链接并不能说明当前件被删除后，保留的子集是否仍可解释。"

"Faithfulness evaluation asks the converse question, whether a generated summary is supported by its source (Maynez et al. 2020), rather than whether the retained source is self-contained." -> "忠实度评估提出的是相反的问题，即生成的摘要是否得到其来源的支持（Maynez等人，2020），而非保留的源文本是否自包含。"

"Closest to our setting, Deng et al. (2025) report boundary and information-path failures under gist-based compression." -> "与我们的设置最接近的是Deng等人（2025）在基于要点（gist）的压缩下报告的边界和信息路径失败。"

"Retrieval-augmented generation retrieves, rewrites, or attends over evidence to favor passages that support generation (Lewis et al. 2020; Xu et al. 2024; Liu et al. 2023; Choi et al. 2024), but scoring by answer support still lets a passage qualify while remaining uninterpretable once its definition is removed." -> "检索增强生成会检索、改写或关注证据，以偏向支持生成的段落（Lewis等人，2020；Xu等人，2024；Liu等人，2023；Choi等人，2024），但按答案支持度评分仍然可能让一个段落合格，而其定义一旦被删除就变得不可解释。"

"Long-context studies show that accuracy also depends on where relevant information sits and on surrounding distractors (Liu et al. 2024; Shi et al. 2023), factors that fixed-budget reselection changes and that we therefore control." -> "长上下文研究表明，准确率还取决于相关信息的位置及周围的干扰项（Liu等人，2024；Shi等人，2023），这些因素会随固定预算的重新选择而变化，因此我们对其加以控制。"

"We measure what the missing constraint costs: how often independent selection splits a dependency pair, and how much downstream accuracy is lost when it does." -> "我们衡量缺失约束的代价：独立选择拆散依赖对的频率，以及拆散时下游准确率损失多少。"

Section 3.

"#### Referential dangling." -> "#### 指称悬空。"

Then formal definitions. This is math heavy. Need translate surrounding text, preserve formulas.

Let's render formulas in plain text as close as possible. Since markdown, no LaTeX rendering, but we can use inline code? Better to keep as original LaTeX-ish.

Original:

"LetC=(c1,...,c|C|) denote a tokenized context,Q∈Q a query, andτ∈N a token budget."

Actually formatting: \(C=(c_1,\ldots,c_{\|C\|})\). We can write "设 \(C=(c_1,\ldots,c_{\|C\|})\) 表示一个 token 化的上下文，\(Q\in\mathcal{Q}\) 表示查询，\(\tau\in\mathbb{N}\) 表示 token 预算。" Need preserve math. We'll use LaTeX in text.

But user said preserve markdown. It's okay.

Need translate all definitions.

Let's parse each formula:

"LetC=\(c1,...,c\|C\|\)C=\(c_{1},\ldots,c_{\|C\|}\)denote a tokenized context,Q∈Q\\in\\mathcal\{Q\}a query, andτ∈N\\tau\\in\\mathbb\{N\}a token budget."

We'll write: "设 \(C=(c_1,\ldots,c_{\|C\|})\) 表示一个 token 化的上下文，\(Q\in\mathcal{Q}\) 表示一个查询，\(\tau\in\mathbb{N}\) 表示 token 预算。"

"We restrict the formal definition to extractive hard compressors." -> "我们将形式化定义限于抽取式硬压缩器。"

"For any extractive outputAA, letPosC⁡\(A\)⊆\{1,...,\|C\|\}\\operatorname\{Pos\}\_\{C\}\(A\)\\subseteq\\\{1,\\ldots,\|C\|\\\}denote the source-token positions emitted inAA, each emitted at most once and in source order." -> "对于任何抽取式输出 \(A\)，令 \(\operatorname{Pos}_C(A)\subseteq\{1,\ldots,\|C\|\}\) 表示 \(A\) 中输出的源 token 位置，每个位置至多输出一次，且按源文本顺序排列。"

"An extractive compressorffreturnsC~=f\(C,Q,τ\)\\tilde\{C\}=f\(C,Q,\\tau\)with\|C~\|≤τ<\|C\|\|\\tilde\{C\}\|\\leq\\tau<\|C\|, where\|⋅\|\|\\cdot\|denotes the number of tokens." -> "抽取式压缩器 \(f\) 返回 \(\tilde{C}=f(C,Q,\tau)\)，满足 \(\|\tilde{C}\|\le\tau<\|C\|\)，其中 \(\|\cdot\|\) 表示 token 数量。"

"A compressor that does not useQQis query agnostic. The downstream LLM receivesQQandC~\\tilde\{C\}but notCC." -> "不使用 \(Q\) 的压缩器为查询无关（query agnostic）。下游 LLM 接收 \(Q\) 和 \(\tilde{C}\)，但不接收 \(C\)。"

"A sentence relevant to the task may depend on explicit support elsewhere in the context, and several alternative support paths may be valid." -> "与任务相关的句子可能依赖于上下文中其他位置的显式支持，并且可能存在多条替代的有效支持路径。"

"LetSent⁡\(C\)\\operatorname\{Sent\}\(C\)denote the indexed source-sentence occurrences inCC. For any indexed source fragmentyy, letI\(y\)⊆\{1,...,\|C\|\}\\mathcal\{I\}\(y\)\\subseteq\\\{1,\\ldots,\|C\|\\\}contain its source-token positions." -> "令 \(\operatorname{Sent}(C)\) 表示 \(C\) 中带索引的源句子出现位置。对于任何带索引的源片段 \(y\)，令 \(\mathcal{I}(y)\subseteq\{1,\ldots,\|C\|\}\) 包含其源 token 位置。"

"LetTC,Q⊆Sent⁡\(C\)\\mathcal\{T\}\_\{C,Q\}\\subseteq\\operatorname\{Sent\}\(C\)contain the task-relevant sentences whose interpretation or use in answeringQQmay require such support." -> "令 \(\mathcal{T}_{C,Q}\subseteq\operatorname{Sent}(C)\) 包含任务相关句子，这些句子在回答 \(Q\) 时的解释或使用可能需要此类支持。"

"We define exact sentence retention by

RetC⁡\(A\)=\{s∈Sent⁡\(C\):I\(s\)⊆PosC⁡\(A\)\}\.\\operatorname\{Ret\}\_\{C\}\(A\)=\\left\\\{s\\in\\operatorname\{Sent\}\(C\):\\mathcal\{I\}\(s\)\\subseteq\\operatorname\{Pos\}\_\{C\}\(A\)\\right\\\}.
"

We can write:

"我们将精确句子保留定义为
\[
\operatorname{Ret}_C(A)=\left\{s\in\operatorname{Sent}(C):\mathcal{I}(s)\subseteq\operatorname{Pos}_C(A)\right\}.
\]"

Need use display math. Since markdown, maybe use $$...$$. We'll do.

Next:

"Fors∈TC,Qs\\in\\mathcal\{T\}\_\{C,Q\}andD⊆Sent⁡\(C\)∖\{s\}D\\subseteq\\operatorname\{Sent\}\(C\)\\setminus\\\{s\\\}, letSuffC,Q⁡\(s,D\)\\operatorname\{Suff\}\_\{C,Q\}\(s,D\)hold when the source-ordered text formed byD∪\{s\}D\\cup\\\{s\\\}, together withQQ, contains all explicit information fromCCneeded to interpretssand use it in an evidence chain forQQ."

Translate: "对于 \(s\in\mathcal{T}_{C,Q}\) 和 \(D\subseteq\operatorname{Sent}(C)\setminus\{s\}\)，当由 \(D\cup\{s\}\) 按源顺序组成的文本与 \(Q\) 一起包含 \(C\) 中解释 \(s\) 并将其用于 \(Q\) 的证据链所需的全部显式信息时，\(\operatorname{Suff}_{C,Q}(s,D)\) 成立。"

"Sufficiency is assessed jointly onD∪\{s\}D\\cup\\\{s\\\}rather than one dependency edge at a time." -> "充分性是在 \(D\cup\{s\}\) 上联合评估的，而不是一次评估一条依赖边。"

"Define the complete family of inclusion-minimal sufficient support sets

DC,Q\(s\)=\{D:\\displaystyle\\mathcal\{D\}\_\{C,Q\}\(s\)=\\bigl\\\{D:D⊆Sent⁡\(C\)∖\{s\},SuffC,Q⁡\(s,D\),\\displaystyle D\\subseteq\\operatorname\{Sent\}\(C\)\\setminus\\\{s\\\},\\quad\\operatorname\{Suff\}\_\{C,Q\}\(s,D\),∄D′⊊DwithSuffC,Q\(s,D′\)\}\.\\displaystyle\\nexists D^\{\\prime\}\\subsetneq D\\ \\text\{with\}\\ \\operatorname\{Suff\}\_\{C,Q\}\(s,D^\{\\prime\}\)\\bigr\\\}.
"

This is garbled. Need output clean math. The intended:

\(\mathcal{D}_{C,Q}(s)=\{D\subseteq\operatorname{Sent}(C)\setminus\{s\}: \operatorname{Suff}_{C,Q}(s,D), \nexists D'\subsetneq D\text{ with }\operatorname{Suff}_{C,Q}(s,D')\}\).

We'll write:

"定义包含最小充分支持集的完全族
\[
\mathcal{D}_{C,Q}(s)=\left\{D\subseteq\operatorname{Sent}(C)\setminus\{s\}:\operatorname{Suff}_{C,Q}(s,D),\ \nexists D'\subsetneq D\ \text{with}\ \operatorname{Suff}_{C,Q}(s,D')\right\}.
\]"

Then "For everys∈TC,Qs\\in\\mathcal\{T\}\_\{C,Q\}, we assumeDC,Q\(s\)≠∅\\mathcal\{D\}\_\{C,Q\}\(s\)\\neq\\varnothing. BecauseSent⁡\(C\)\\operatorname\{Sent\}\(C\)is finite, every sufficientDDcontains an inclusion-minimal sufficient subset, so retaining a sufficient support set is equivalent to retaining some member of this min"

Need continue? The input seems truncated at "some member of this min" because section ends? Actually provided stops at "min" at end of Section 3. Need translate what is present. Perhaps complete? It says "so retaining a sufficient support set is equivalent to retaining some member of this min" maybe there is more but cut off. We'll translate exactly: "因此，保留一个充分支持集等价于保留该最小族中的某个成员。" But "min" incomplete. Maybe "minimal family" should be "minimal family." We'll infer "this minimal family" and complete the sentence. Since it says "this min" likely "this minimal family". We'll translate as "该最小族" to close sentence.

Need ensure no extra content.

The user said "Translate the given markdown article content" This is an excerpt ending at mid-sentence. We should translate what's there, complete the sentence logically.

Now final answer: translated markdown.

Need maintain comments? No.

Let's produce.# 硬提示压缩中的指称悬空：一种范式级失败模式

Source: https://arxiv.org/html/2608.04569

Zhengpei Hu1\equalcontrib, Kai Li2\equalcontrib, Dapeng Fu3, Xuechao Zou2, Yuanhao Tang1, Yue Li1, Tengfei Cao1, Jianqiang Huang1\corresponding

###### 摘要

硬提示压缩通过对 token、句子或块进行独立评分，并在预算内保留得分最高的单元来降低长上下文推理成本。我们识别出这一流程中的结构性失败：独立选择可能拆散相互依赖的证据对，保留其中一员而删除另一员。当保留的文本片段包含答案，但被删除的片段定义了解释该答案所需的实体时，我们将结果称为*指称悬空*。在压缩比为0.30时，使用 Qwen3-0.6B 嵌入对连贯块进行排序的 Beaver，在三个多跳问答（QA）数据集的桥接示例中，分别导致34%至54%的答案路径不完整。该失败并不特定于该实现：在共享的 HotpotQA 桥接集合上，我们测试的全部六个硬压缩器都表现出最高达60%的指称悬空率；LongBench-v2 单文档问答中的每篇文档都至少包含一处指称悬空的引用。随后，我们测试在相同预算下选择不同内容是否有帮助。在使用 Qwen3-8B 评估的悬空示例上，重新插入缺失的支持段落，并通过删除未被标注为支持答案的段落来抵消其 token 数，可将准确率提高29至34个百分点（p<10^{-4}），并恢复至少88%的、与保留两个标注支持段落的上下文之间的差距。更强的答案模型无法吸收这一损失：在 MuSiQue 上，GPT-5.5 在压缩上下文上的准确率比保留两个支持段落的上下文低8.8个百分点。最后，我们训练一个紧凑分类器，对被压缩器省略的句子按是否需要用于解释保留文本进行排序，并在推理时不使用支持标注地重新插入排名最高的候选。在 HotpotQA 上，以 Qwen3-8B 作为下游答案模型，这种自动恢复将准确率提高了4.7个百分点，而压缩比仅从0.30变为0.31。硬压缩器应当同时优化相关性和指称完整性。

## 1 引言

参见图1：缺少桥接信息的指称悬空。独立评分保留了查询主语和答案字符串，却删除了 Tim DuBois 出生于 Southwest City 这一事实，使得推理链不完整。

在最近的前沿模型中，大语言模型（LLM）的上下文窗口已从数千个 token 扩展到百万 token 量级。这些扩展的窗口使模型能够对整篇文档和文档集合进行推理。但长上下文仍然代价高昂：自注意力的二次方成本推高了预填充延迟，而且当相关内容远离提示边界时，模型对其利用不足（Liu 等人，2024）。提示压缩通过在输入到达目标模型之前对输入进行剪枝，同时尽力保留任务所需信息，来缓解这一矛盾（Jiang 等人，2023b；Pan 等人，2024；Li 等人，2023）。

许多硬提示压缩器采用一种常见的选择程序：对每个 token、句子或块的重要性进行评分，并在 token 预算内保留得分最高的片段。这些分数可能反映查询相关性或困惑度（Jiang 等人，2024）、嵌入相似度（Hu 等人，2026）、信息量或学习到的分类器分数（Li 等人，2023；Pan 等人，2024）、句法显著性（Mao 等人，2025），或由注意力推导出的重要性（Zhao 等人，2025）。尽管这些方法在架构和监督方式上有所不同，但它们的共同目标是保留看似重要的片段。这一目标假设所选片段共同构成一个可用的提示。然而，重要的片段未必是自包含的，因为其含义可能依赖于前文定义、先行词或桥接事实，而压缩器可能独立地删除这些依赖。

我们将这种失败称为*指称悬空*，并在第3节中对其进行形式化定义。当压缩后的提示保留了一个相关片段，但删除了解释该片段所需的依赖，使该片段对目标模型不可用时，就出现了指称悬空。图1展示了这一机制。压缩器保留了提及 Tim DuBois 的一个片段，以及说明 Southwest City 位于 McDonald County 的另一个片段，却删除了说明 DuBois 出生于 Southwest City 的桥接信息。尽管答案字符串仍留在压缩提示中，但保留的片段已不再支持从查询主语到答案的推理。这个例子说明，仅靠相关性并不能保证压缩上下文可用。目前仍不清楚指称悬空是否具有系统性、固定预算下的重新选择能否恢复准确率，以及恢复能否自动化。

我们首先检验 Beaver（Hu 等人，2026）的官方实现。在0.30的压缩比下，它在 HotpotQA（Yang 等人，2018）、2WikiMultiHopQA（Ho 等人，2020）和 MuSiQue（Trivedi 等人，2022）的桥接示例中，分别导致34%至54%的答案路径不完整。人工审计发现其标记的查准率为95%。这一诊断并不限于该设置：在共享的桥接集合上，所有六个被测试的压缩器都表现出32%至60%的指称悬空率；LongBench-v2 单文档问答（Bai 等人，2025）中的所有80篇文档都至少包含一处指称悬空的引用。接下来，我们进行了一项由标注指导的重新选择实验：将遗漏的支持段落重新插入，并通过删除未被标注为支持答案且得分最低的段落来抵消其 token 数。在使用 Qwen3-8B 评估的悬空示例上，这一干预在不增加 token 预算的情况下将准确率提高了29至34个百分点（p<10^{-4}），并恢复了原始压缩上下文与保留两个标注支持段落的上下文之间准确率差距的至少88%。最后，我们从被省略的文本中生成候选句子，并训练一个紧凑分类器，根据这些句子是否被需要用于解释保留文本对它们排序。该方法在推理时不使用支持段落标注，而是重新插入排名最高的候选。在 HotpotQA 上以 Qwen3-8B 作为下游答案模型时，这种自动恢复将准确率提高了4.7个百分点，而压缩比仅从0.30变为0.31。我们已发布代码、单个示例的结果以及训练好的模型¹¹https://cslikai.cn/Referential-Dangling/。

## 2 相关工作

#### 提示与上下文压缩。

提示压缩通过在输入到达目标模型之前缩短输入来降低长上下文推理成本。它与模型侧效率方法互补，包括弹性子网络、用于多模态输入的紧凑离散语义 token，以及高效的编码器-解码器或状态空间架构（Li 和 Luo，2024；Li 等人，2025b，2022，a）。我们研究硬压缩，即在预算下使用自信息或困惑度（Li 等人，2023；Jiang 等人，2024）、学习型 token 分类器（Jiang 等人，2023b；Pan 等人，2024）、嵌入相似度（Hu 等人，2026）、句法显著性（Mao 等人，2025）、基于注意力的信号（Zhao 等人，2025），