citation-coverage-audit

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Citation Coverage Audit

引用覆盖审核

Audit whether a draft paper cites the works it should cite. This skill is about reference completeness: classic foundations, closest prior work, current concurrent work, benchmark/dataset/method attributions, and reviewer-sensitive omissions.

Use this skill when:

the user has a draft paper but suspects the reference list is incomplete
a related work section may miss classic or recent work
the paper needs to cite concurrent arXiv/OpenReview papers before submission
the paper makes novelty claims that require stronger citation support
the user wants to reduce reviewer complaints like "missing important related work"

Do not use this as a replacement for

citation-audit

. Use

citation-audit

after this skill to verify that added BibTeX entries and citation keys are correct.

Pair this skill with

research-project-memory

when missing citations affect novelty, baseline, related-work, or reviewer-risk tracking.

审核论文草稿是否引用了应当引用的研究工作。此技能聚焦于参考文献的完整性：涵盖经典基础研究、最相近的前期工作、当前的同期研究、基准/数据集/方法的归属，以及审稿人关注的遗漏内容。

在以下场景中使用此技能：

用户拥有论文草稿，但怀疑参考文献列表不完整
相关工作部分可能遗漏了经典或近期研究
论文在提交前需要引用arXiv/OpenReview上的同期研究
论文提出的创新性声明需要更有力的引用支持
用户希望减少审稿人诸如“遗漏重要相关工作”的意见

请勿将此技能作为

citation-audit

的替代工具。使用完此技能后，应使用

citation-audit

来验证添加的BibTeX条目和引用键是否正确。

当缺失的引用会影响创新性、基准、相关工作或审稿风险跟踪时，可将此技能与

research-project-memory

配合使用。

Skill Directory Layout

技能目录结构

text

<installed-skill-dir>/
├── SKILL.md
└── references/
    ├── coverage-taxonomy.md
    ├── discovery-protocol.md
    ├── memory-model.md
    └── report-template.md

text

<installed-skill-dir>/
├── SKILL.md
└── references/
    ├── coverage-taxonomy.md
    ├── discovery-protocol.md
    ├── memory-model.md
    └── report-template.md

Progressive Loading

渐进式加载

Always read

references/coverage-taxonomy.md

and

references/discovery-protocol.md

Read
```
references/report-template.md
```
when writing the audit output.
Read
```
references/memory-model.md
```
when saving or reusing topic-specific citation coverage knowledge.
Use current web search for recent and concurrent work. Do not rely on static memory for "latest" or "recent" citations.

务必阅读

references/coverage-taxonomy.md

和

references/discovery-protocol.md

。

撰写审核输出时，阅读
```
references/report-template.md
```
。
保存或复用特定主题的引用覆盖知识时，阅读
```
references/memory-model.md
```
。
使用当前网络搜索查找近期和同期研究。请勿依赖静态记忆获取“最新”或“近期”的引用。

Core Principles

核心原则

Missing close citations are submission risks. Treat them as reviewer-facing problems, not bibliography trivia.
Separate classic foundations, closest prior work, and recent concurrent work. They serve different rhetorical and reviewer functions.
Search from the paper's claims, not just keywords. A method paper, benchmark paper, theory paper, and empirical study need different citation coverage.
Prefer primary sources: official papers, proceedings pages, arXiv, OpenReview, ACL Anthology, PMLR, CVF, ACM, IEEE, Springer, Semantic Scholar, DBLP, and authors' project pages.
Do not pad the bibliography. Recommend citations only when they support a specific sentence, section, baseline, dataset, method component, theorem, or novelty boundary.
Mark uncertainty. If a candidate looks relevant but cannot be verified, list it as
```
needs-author-review
```
.

缺失相近研究的引用会带来提交风险。将其视为面向审稿人的问题，而非参考文献琐事。
区分经典基础研究、最相近的前期工作和近期同期研究。它们在修辞和满足审稿人需求方面发挥不同作用。
从论文的声明出发进行搜索，而非仅依赖关键词。方法类论文、基准类论文、理论类论文和实证研究需要不同的引用覆盖范围。
优先选择原始来源：官方论文、会议论文集页面、arXiv、OpenReview、ACL Anthology、PMLR、CVF、ACM、IEEE、Springer、Semantic Scholar、DBLP以及作者的项目页面。
不要堆砌参考文献。仅当引用能支持特定句子、章节、基准、数据集、方法组件、定理或创新性边界时，才推荐引用。
标注不确定性。如果候选研究看似相关但无法验证，将其标记为
```
needs-author-review
```
（需作者审核）。

Step 1 - Define Scope

步骤1 - 定义范围

Identify:

target venue and year
topic and subtopic
paper archetype: method, theory, empirical study, benchmark/dataset, systems, analysis, application, or hybrid
paper stage: outline, full draft, final submission, rebuttal, camera-ready
current source format: LaTeX, PDF, Markdown, notes
search horizon for recent work, such as last 6 months, last 12 months, or since the previous conference deadline
whether the user wants a report only or also BibTeX/prose patch suggestions

Default search horizon:

for active ML/AI topics: last 12 months
for fast-moving topics such as foundation models, diffusion, reasoning, agents, alignment, or multimodal learning: last 6-9 months plus current OpenReview submissions if accessible
for theory or mature topics: include classics and last 24 months

确定：

目标会议及年份
主题和子主题
论文类型：方法类、理论类、实证研究类、基准/数据集类、系统类、分析类、应用类或混合类
论文阶段：大纲、完整草稿、最终提交、回复审稿意见、终稿
当前来源格式：LaTeX、PDF、Markdown、笔记
近期研究的搜索时间范围，如过去6个月、过去12个月，或自上一届会议截止日期以来
用户是否仅需要报告，还是同时需要BibTeX/文本修改建议

默认搜索时间范围：

针对活跃的ML/AI主题：过去12个月
针对快速发展的主题，如基础模型、扩散模型、推理、Agent、对齐或多模态学习：过去6-9个月，加上可获取的当前OpenReview提交论文
针对理论或成熟主题：涵盖经典研究和过去24个月的内容

Step 2 - Extract Paper Claims and Existing Coverage

步骤2 - 提取论文声明与现有引用覆盖情况

Read the draft and extract:

title and abstract
contribution bullets
introduction novelty claims
related work categories
method components
theorem assumptions or proof techniques
datasets, benchmarks, metrics, and baselines
experiment comparison points
limitation and scope claims
current citation keys and BibTeX entries

Build a paper citation map:

markdown

undefined

阅读草稿并提取：

标题和摘要
贡献要点
引言中的创新性声明
相关工作分类
方法组件
定理假设或证明技术
数据集、基准、指标和基线
实验对比点
局限性和范围声明
当前的引用键和BibTeX条目

构建论文引用图谱：

markdown

undefined

Paper Citation Map

论文引用图谱

Main claim:
Core method/theory:
Closest related work currently cited:
Classic foundations currently cited:
Recent/concurrent work currently cited:
Datasets/benchmarks currently cited:
Baselines currently cited:
Novelty claims requiring citation support:
Sections with thin citation coverage:


If the draft is incomplete, audit the available claims and mark missing context.

核心声明:
核心方法/理论:
当前已引用的最相近相关工作:
当前已引用的经典基础研究:
当前已引用的近期/同期研究:
当前已引用的数据集/基准:
当前已引用的基线:
需要引用支持的创新性声明:
引用覆盖薄弱的章节:


如果草稿不完整，审核现有声明并标注缺失的上下文。

Step 3 - Classify Required Citation Categories

步骤3 - 分类所需的引用类别

Read

references/coverage-taxonomy.md

For the paper, define required buckets:

foundational classics
closest prior work
direct competitors
concurrent work
benchmark/dataset/metric sources
method/tooling components
theory/proof technique sources
negative or limitation-related citations
surveys or taxonomy papers
venue-specific or community-standard citations

For each bucket, state why it matters for reviewer perception.

阅读

references/coverage-taxonomy.md

。

针对论文定义所需的类别：

经典基础研究
最相近的前期工作
直接竞争对手
同期研究
基准/数据集/指标来源
方法/工具组件
理论/证明技术来源
与局限性相关的引用
综述或分类学论文
特定会议或社区标准要求的引用

针对每个类别，说明其对审稿人看法的重要性。

Step 4 - Discover Missing Work

步骤4 - 发现缺失的研究工作

Read

references/discovery-protocol.md

Search iteratively:

use the paper's own title/claim keywords
search each method component and baseline
search closest cited papers forward and backward when possible
search recent arXiv/OpenReview/proceedings for the topic
search venue-specific accepted/oral/spotlight papers if the target venue matters
inspect surveys or benchmark leaderboards for canonical citations

For recent/concurrent work, include access dates and search queries. Search results change over time.

Do not recommend a paper only because it shares keywords. Each candidate must map to a citation role.

阅读

references/discovery-protocol.md

。

迭代式搜索：

使用论文自身的标题/声明关键词
搜索每个方法组件和基线
尽可能对已引用的最相近论文进行正向和反向搜索
搜索主题相关的近期arXiv/OpenReview/会议论文集
如果目标会议重要，搜索该会议已接收的口头报告/亮点报告论文
查阅综述或基准排行榜以获取标准引用

针对近期/同期研究，记录访问日期和搜索查询。搜索结果会随时间变化。

不要仅因关键词匹配就推荐某篇论文。每个候选研究都必须对应一个明确的引用作用。

Step 5 - Evaluate Candidate Relevance

步骤5 - 评估候选研究的相关性

For each candidate missing paper, classify:

```
must-cite
```
: close prior work, direct competitor, foundational citation, dataset/metric/source attribution, or work required to support/qualify a novelty claim
```
should-cite
```
: relevant recent work, adjacent strong baseline, useful survey, or important related line reviewers may expect
```
optional
```
: context or breadth citation that improves related work but is unlikely to affect review
```
do-not-cite
```
: keyword match but not scientifically relevant
```
needs-author-review
```
: relevance plausible but not enough evidence to decide

For every

must-cite

should-cite

, specify:

where to cite it
what claim it supports or qualifies
how it changes novelty framing
whether it needs BibTeX
whether it suggests new experiments or baselines

针对每篇缺失的候选研究，分类为：

```
must-cite
```
（必须引用）：最相近的前期工作、直接竞争对手、经典基础引用、数据集/指标/来源归属，或支持/限定创新性声明所需的研究
```
should-cite
```
（应当引用）：相关的近期研究、相近的强基线、有用的综述，或审稿人可能期望的重要相关研究方向
```
optional
```
（可选引用）：能丰富相关工作内容但不太可能影响评审的背景或广度引用
```
do-not-cite
```
（无需引用）：仅关键词匹配但无科学相关性
```
needs-author-review
```
（需作者审核）：看似相关但证据不足无法确定

针对每个

must-cite

或

should-cite

，需明确：

引用位置
支持或限定的声明
如何改变创新性框架
是否需要BibTeX条目
是否建议新增实验或基线

Step 6 - Check Novelty and Reviewer Risk

步骤6 - 检查创新性与审稿风险

For each novelty claim, ask:

Is the closest prior work cited?
Is the difference stated correctly?
Would a reviewer expect this missing paper?
Does the missing paper weaken the novelty claim?
Does it require a new baseline, ablation, theorem comparison, or discussion?

Risk levels:

```
blocking
```
: missing citation could make a novelty claim look false or unethical
```
high
```
: reviewer likely complains about missing related work
```
medium
```
: omission weakens positioning but probably not fatal
```
low
```
: useful completeness citation

针对每个创新性声明，询问：

是否引用了最相近的前期工作？
差异描述是否准确？
审稿人是否会期望引用这篇缺失的论文？
缺失的论文是否会削弱创新性声明？
是否需要新增基线、消融实验、定理对比或讨论？

风险等级：

```
blocking
```
（阻碍性）：缺失的引用可能导致创新性声明看起来虚假或不道德
```
high
```
（高风险）：审稿人很可能会抱怨遗漏相关工作
```
medium
```
（中风险）：遗漏会削弱定位但可能不致命
```
low
```
（低风险）：有助于提升完整性的引用

Step 7 - Produce Coverage Report

步骤7 - 生成覆盖报告

Use

references/report-template.md

The report should include:

paper citation map
search strategy and search date
candidate missing citations
must-cite and should-cite recommendations
exact insertion points
novelty-framing updates
baseline/experiment implications
unresolved author decisions

If the user wants a saved report and gives no path, use:

text

docs/reports/citation_coverage_audit_YYYY-MM-DD.md

使用

references/report-template.md

。

报告应包含：

论文引用图谱
搜索策略和搜索日期
候选缺失引用
必须引用和应当引用的建议
精确的插入位置
创新性框架的更新建议
基线/实验的影响
待作者解决的问题

如果用户需要保存报告但未指定路径，使用：

text

docs/reports/citation_coverage_audit_YYYY-MM-DD.md

Step 8 - Optional BibTeX and Text Patches

步骤8 - 可选的BibTeX与文本修改

If the user asks for edits:

add BibTeX only from verified metadata
insert citations at the smallest safe location
do not make broad related-work rewrites unless asked
do not cite papers you have not verified enough to identify
after edits, run or recommend
```
citation-audit
```
to verify keys and metadata

When exact prose is uncertain, write a suggested sentence in the report instead of editing the paper.

如果用户要求编辑：

仅从已验证的元数据添加BibTeX条目
在最安全的最小位置插入引用
除非被要求，否则不要大幅重写相关工作部分
不要引用未充分验证的论文
编辑完成后，运行或建议使用
```
citation-audit
```
验证引用键和元数据

当不确定具体措辞时，在报告中写出建议句子，而非直接编辑论文。

Step 9 - Update Citation Coverage Memory

步骤9 - 更新引用覆盖记忆

Read

references/memory-model.md

When a topic scan was performed:

update

.agent/citation-coverage/topics/<topic-slug>.md

update

.agent/citation-coverage/project-coverage.md

record search queries, access dates, candidate papers, and decisions

Memory must separate:

canonical papers for the topic
recent/concurrent papers with access dates
papers already cited
papers intentionally not cited
unresolved author decisions

If the project uses

research-project-memory

, also update:

```
memory/risk-board.md
```
: missing closest-work, classic, benchmark, baseline, or concurrent-work risks
```
memory/action-board.md
```
: add-citation, check-bibtex, add-baseline, revise-novelty, or author-review actions
```
memory/claim-board.md
```
: novelty claims that must be weakened or qualified
```
memory/evidence-board.md
```
: citation evidence for claims when a paper directly supports or limits a claim
```
paper/.agent/paper-status.md
```
: related-work sections or insertion points that need edits

Use

needs-verification

for candidate papers not yet read deeply enough.

阅读

references/memory-model.md

。

当完成主题扫描后：

更新

.agent/citation-coverage/topics/<topic-slug>.md

更新

.agent/citation-coverage/project-coverage.md

记录搜索查询、访问日期、候选论文和决策

记忆内容必须区分：

主题的标准论文
带有访问日期的近期/同期论文
已引用的论文
有意不引用的论文
待作者解决的问题

如果项目使用

research-project-memory

，还需更新：

```
memory/risk-board.md
```
：缺失最相近研究、经典研究、基准、基线或同期研究的风险
```
memory/action-board.md
```
：添加引用、检查BibTeX、添加基线、修改创新性声明、作者审核等行动项
```
memory/claim-board.md
```
：必须弱化或限定的创新性声明
```
memory/evidence-board.md
```
：当论文直接支持或限定某声明时，该声明的引用证据
```
paper/.agent/paper-status.md
```
：需要编辑的相关工作部分或引用插入位置

对于未深入阅读的候选论文，标记为

needs-verification

（需验证）。

Final Sanity Check

最终合理性检查

Before finalizing:

classic, closest, and recent/concurrent work were considered separately
search queries and dates are recorded
every recommended paper has a citation role
must-cite recommendations include insertion points
novelty risks are explicit
uncertain candidates are not presented as verified
any edits are followed by a recommendation to run
```
citation-audit
```

完成前需确认：

已分别考虑经典研究、最相近研究和近期同期研究
已记录搜索查询和日期
每篇推荐论文都有明确的引用作用
必须引用的建议包含精确插入位置
创新性风险已明确标注
不确定的候选研究未被当作已验证内容呈现
任何编辑操作后都建议运行
```
citation-audit
```