analyze-paper

Paper Analysis — One-click deep analysis of a paper

Enter a paper URL, PDF path, or title, and automatically collect information across seven dimensions in parallel, then output structured research notes after cross-analysis.

When to Use

Triggered when the user requests the following operations:

"Analyze this paper https://arxiv.org/abs/..." / "Analyze the xxx paper"
"paper analysis: Attention Is All You Need"
"Read this paper" / "Help me analyze this paper"
"What is this paper about" / "Paper notes xxx"

Source credibility rules (global, applicable to all Phases/Agents)

Primary source priority principle

The original paper is the sole authority for facts. All key numbers, method descriptions, and experimental resultsmust be traceable to specific sections/figures/tables/formulas in the original paper. External sources are only used to supplement interpretation and community feedback.

Source layering

Tier	Source type	Purpose	Annotation requirement
L1 Primary source	Original paper (arXiv HTML/PDF), official conference pages (USENIX/ACM/IEEE/NeurIPS), author homepages	Key facts, numbers, method descriptions	Annotate Section/Figure/Table numbers
L2 Strongly credible	Google Scholar, Semantic Scholar, DBLP, OpenReview official reviews, lab homepages	Citation counts, h-index, review comments	Annotate page URL + retrieval date
L3 Supplementary interpretation	Well-known technical blogs (such as the author's own blog, official institutional blogs), academic community discussions (original Reddit/Twitter posts)	Community feedback, accessible interpretation	Clearly annotate as "external interpretation", not to be used as a factual basis
L4 Low credibility	Press releases, aggregator sites, media paraphrases, exchange/marketing websites	Must not be used to support key facts, and may only be cited in the community feedback dimension as a "public opinion signal"	Annotate as "low-credibility source"

In-paragraph citations and assertion annotation

The "pile links at the end" pattern is prohibited. All key assertions must have their sources annotated in place within the text:

Facts from the original paper:（论文 Section 3.2, Figure 4）
Externally verifiable data:（Google Scholar, 2026-03-14 抓取, 1,234 citations）
External interpretation:（[来源名称](url), 外部解读）
Analyst inference: must be marked with wording such as "the analysis believes"/"inferred", and must not be disguised as fact
Time-sensitive metrics (citation counts, stars, upvotes):must annotate retrieval date + source URL; if they cannot be obtained from L1/L2, do not write a number, and instead write "as of {date}, no reliable public data found"

Phase 0: Paper positioning and input parsing

Hard constraint (cannot be skipped): Detecting whether an existing analysis file exists locally is prohibited. Regardless of whether there is a file with the same name or a similar file in the working directory, the complete analysis process must be executed from scratch and the report regenerated.

Based on the user input, identify the input type and execute the corresponding process:

Input path

Input Type	Processing Flow
URL	WebFetch retrieves the page and extracts the title, authors, abstract, and institutions
PDF Path	Read reads the first few pages of the PDF and extracts the title, authors, and abstract
Title/Keywords	WebSearch searches and locates, then WebFetch retrieves the details

URL Input Processing

arXiv Link

Step 1: WebFetch https://arxiv.org/abs/{id}
        → 提取标题、作者、摘要、提交日期
Step 2: 尝试 WebFetch https://arxiv.org/html/{id} 获取论文 HTML 全文（用于后续 Agent 引用具体章节）
Step 3: WebSearch "{paper_title} conference accepted" 或 "{paper_title} published venue"
        → 确认发表会议/期刊

Other URL(Semantic Scholar, ACL Anthology, OpenReview, etc.)

WebFetch {url} → 提取标题、作者、摘要、机构、venue

PDF Input Processing

Read {pdf_path} (前 5 页)
→ 提取标题、作者、摘要、机构
→ 记录论文结构（章节标题列表）供后续引用
→ 如信息不全，WebSearch "{提取到的标题}" 补充

Title/Keyword Input Processing

Step 1: WebSearch "{论文标题/关键词} arxiv OR paper OR conference"
        → 找到论文链接
Step 2: WebFetch {找到的链接}
        → 提取完整信息
Step 3: 如果找到 arXiv 链接，尝试获取 HTML 全文

If the search results are unclear, confirm with the user before continuing.

Obtain the original paper text (key step, cannot be skipped)

Before entering Phase 1,mustattempt to obtain the original paper content as the primary source of facts:

arXiv HTML(preferred):WebFetch https://arxiv.org/html/{id} — obtain the structured full text
PDF: if the user provided a PDF path,Read {pdf_path} read the full text
Official conference page：WebFetch {conference_url} — some conferences provide an HTML version

Store the obtained original paper content (or key section summaries) into the {paper_content} variable and pass it to each Agent. If the full text cannot be obtained, note in {paper_content} "full paper text not obtained, Agent needs to supplement via search".

Generate the paper location map (required, cannot be skipped)

From {paper_content} extract the paper's structural skeleton and generate {paper_locator_map}, and all subsequent Agents must reference the numbers in this map:

从论文原文中提取：
1. Section 编号与标题（如 "Section 3.2: DualPath Architecture"）
2. 所有 Figure 编号与标题（如 "Figure 3: System overview"）
3. 所有 Table 编号与标题（如 "Table 2: Offline throughput comparison"）
4. 所有 Algorithm 编号与标题（如有）
5. References 中被引用 ≥ 2 次的前置论文编号与标题

输出格式：
## paper_locator_map
### Sections
- Section 1: Introduction
- Section 2: Background
- Section 3: Method
  - Section 3.1: ...
  - Section 3.2: ...
- ...
### Figures
- Figure 1: ...
- Figure 2: ...
### Tables
- Table 1: ...
- Table 2: ...
### Key References
- [12] Attention Is All You Need (Vaswani et al., 2017)
- [15] ...

If the paper HTML has no numbering (some arXiv papers), use the Section title as the location anchor (e.g., "Evaluation > Offline Throughput"）。

The output of this step {paper_locator_map} must be passed to each Agent, as the "list of valid numbers" for citations. The Section/Figure/Table numbers cited by Agents must come from this map.

Determine key variables

The following variables are needed by all subsequent Agents:

{paper_title} — full paper title
{authors} — author list
{institution} — institution/team
{year} — publication year
{venue} — publication conference/journal (if unpublished, mark as arXiv preprint)
{abstract} — abstract
{paper_content} — original paper content or key sections (passed to each Agent as the primary source of facts)
{paper_locator_map} — paper structural skeleton (list of Section/Figure/Table/Algorithm numbers; Agent citations must come from this map)
{paper_type} — paper type (academic paper / technical report / white paper)
{lang} — output language (user-specified or default Chinese)
{date} — analysis date, format YYYY-MM-DD
{paper_name_en} — English filename form (lowercase, hyphen-separated)

Paper type adaptation tips

Pass to each sub-agent:

Academic paper → "This is an academic paper; focus on analyzing the research methods, experimental design, and theoretical contributions"
Technical report → "This is a technical report; focus on analyzing the system design, engineering practices, and performance metrics"
White paper → "This is a white paper; focus on analyzing industry insights, solution design, and business value"
Survey/review → "This is a survey paper; focus on analyzing the classification framework, research lineage, and future directions"

If the paper type or key information cannot be determined, first use a single WebSearch for a quick lookup,do not skip this step。

Phase 1: Parallel Data Collection (7 Subagents)

Key requirement: The following 7 Agents must be placed inthe same messageto ensure parallel execution.

Each Agent uses subagent_type: "general-purpose"。

Global Agent instructionsdivided into two parts, placed respectively at thebeginningandend：

{preceding reference constraints}(placed in each Agent promptbefore the task description, as a hard constraint prefix):

## 硬约束（违反将导致输出被丢弃）

**可用的论文定位编号**: {paper_locator_map}
你引用的 Section/Figure/Table 编号**必须来自上述 map**，不得编造不存在的编号。

1. 每个关键数字/断言后必须**就地**括号标注论文位置: `（论文 Section X.Y, Table/Figure Z）`
   - ✅ `离线吞吐提升 1.87×（论文 Table 2, Section 5.1）`
   - ❌ `离线吞吐提升 1.87×（来源: arxiv html）`
   - ❌ `离线吞吐提升 1.87×` （无标注）
   - ❌ 段落末尾统一放一个链接或"参见论文 Section 5" — 这是**段末堆链接**，不算引用
   - HTML 无编号时用 Section 标题: `（论文 "Evaluation > Offline Throughput"）`
2. **非数字的技术断言也需标注**：具体机制（如"使用 RDMA 直接传输"）、设计决策（如"采用 io_uring 旁路内核"）、性能特征（如"近线性扩展"）→ 必须标注论文中描述该机制的 Section/Figure
3. L4 来源（新闻稿/聚合站/交易所）不得支撑任何事实断言
4. 无法定位到论文具体位置的数字或技术细节 → 不写，改为定性描述或省略
5. 分析推断必须用"分析认为/推断"标记

{output format requirements}(placed in each Agent promptat the end）：

## 来源与引用规则（必须遵守）

1. **一手源优先**：论文原文是唯一事实权威。关键数字和方法描述必须能定位到论文的具体 Section/Figure/Table。
2. **段内引用**：每个关键断言后必须括号标注来源，格式：
   - 论文原文：（论文 Section X.Y, Figure Z）
   - 外部验证数据：（来源名, URL, 抓取日期）
   - 外部解读：（[来源名](url), 外部解读）
3. **来源分层**：
   - L1（论文原文、会议官方页、作者主页）→ 可支撑关键事实
   - L2（Google Scholar、Semantic Scholar、OpenReview）→ 可支撑引用量等指标
   - L3（技术博客、社区讨论）→ 仅作补充解读，标注"外部解读"
   - L4（新闻稿、聚合站）→ 不得用于关键事实
4. **不确定就不写**：无法从 L1/L2 来源验证的具体数字，不写数字，改为"未找到可靠数据"。
5. **区分事实与推断**：论文原文事实直接陈述；分析者推断必须用"推断/分析认为"标记。
6. **字数上限 800 字**：超出时优先保留有一手源支撑的内容，砍掉无法验证的描述。

## 必须的输出格式

在你的分析文本之后，**必须**附加一个断言-位置映射表:

### 断言-位置映射
| 断言/数字 | 论文位置 | 来源层级 |
|-----------|---------|---------|
| {每个关键数字或事实断言} | Section X, Table Y / 来源URL | L1/L2/L3/L4 |

要求: 至少覆盖你输出中的所有阿拉伯数字。无法定位的填"未定位"。

Agent 1: Background and Motivation

description: "Research the background and motivation of {paper_title}"

prompt template:

{前置引用约束}

你是学术研究分析专家。请研究论文《{paper_title}》（{authors}, {year}）的研究背景与动机。

论文摘要：{abstract}
论文原文（如有）：{paper_content}
论文定位映射：{paper_locator_map}
论文类型提示：{type_hint}
输出语言：{lang}

任务:
1. **优先从论文原文提取**：从 {paper_content} 中定位 Introduction / Related Work 部分，提取：
   - 研究领域现状和前人工作
   - 核心问题和动机
   - 标注具体 Section 编号
2. **竞争路线图谱**：从 Related Work / Introduction 中识别解决同一核心问题的 3-5 条主要技术路线（如扩窗口 / 外部存储 / 压缩记忆 / 稀疏检索 / 混合架构），每条路线：名称、核心思路一句话、主要代价一句话、1-2 篇代表工作（标注论文 References 编号）
3. 使用 WebSearch 搜索补充信息（至少 1 次）:
   - "{paper_title} motivation background related work"
   → 补充论文未明确说明的领域背景和竞争路线
4. 使用 WebFetch 访问至少 1 个 L1/L2 来源页面验证
5. **时机分析**：为什么这项工作出现在现在？识别 2-3 条正在汇合的技术/行业路线，说明它们如何共同催生了本文的切入点

整理以下信息:

   **研究领域现状:**
   - 该领域的发展脉络和现状（标注信息来源）
   - 前人工作的主要贡献和局限（标注论文 Section）

   **竞争路线地图（必须输出，供 Phase 2 合成技术地图）:**

   | 路线 | 核心思路 | 主要代价 | 代表工作 | 论文 Section |
   |------|---------|---------|---------|-------------|
   | {路线一} | {一句话} | {一句话} | {[N] 论文名} | Related Work / Section X |
   | {路线二} | ... | ... | ... | ... |
   | ... | ... | ... | ... | ... |

   本文属于哪条路线: {路线名} — 做了什么独特选择: {一句话}

   **核心问题与动机:**
   - 论文要解决的核心问题（标注论文 Section）
   - 现有方法为什么不够好（标注论文中的论据）
   - 作者的切入角度和研究假设

   **为什么是现在:**
   - 使这项工作变得可能或必要的 2-3 条汇合趋势（如技术成熟度、行业需求、前置工作积累）
   - 每条趋势的当前状态和与本文的关系（标注来源）

{输出格式要求}

Agent 2: Team and Author Background

description: "Research the team background of {paper_title}"

prompt template:

{前置引用约束}

## 本 Agent 专属约束
- **禁止来源**: 36kr、ResearchGate、新闻稿、媒体报道 — 不得作为团队背景的任何支撑
- 入职时间、实习经历等履历细节 → 除非作者主页明确写明，否则**不写**（不是标"待核实"，是不写）
- 每条断言标注具体来源页面路径（如 "xinjin.github.io > Publications"），不得笼统写"来源: 作者主页"
- **禁止定性评价**: "最具影响力""核心人物""标志性学者""领军团队"等 → 除非有官方奖项/头衔页面支撑，否则不写
- **宁缺毋滥**: 若无法从作者主页/Scholar/DBLP 验证，整段缩减为"作者所属机构 + 与本文直接相关的 1-2 篇前序工作"，禁止补充履历叙事。信息不足时本维度压缩为 2-4 句即可

你是学术情报分析专家。请研究论文《{paper_title}》作者团队的背景信息。

作者列表：{authors}
机构：{institution}
输出语言：{lang}

任务:
1. 使用 WebSearch 搜索以下关键词（至少搜 2 次不同关键词）:
   - "{第一作者} {institution} research site:scholar.google.com OR site:dblp.org OR site:semanticscholar.org"
   - "{通讯作者} homepage OR lab"
2. 使用 WebFetch 访问至少 1 个作者主页、Google Scholar 或 DBLP 页面（仅 L1/L2 来源）

整理以下信息:

   **核心作者介绍:**
   - 第一作者和通讯作者的背景（职位、研究方向、代表作）
   - 所属机构/实验室的研究特色
   - **仅列出可从官方主页或 Google Scholar 确认的荣誉/奖项**
   - 无法从 L1/L2 来源验证的信息 → **不写**（不是标"待核实"）

   **研究脉络:**
   - 团队在该领域的前序工作（列出具体论文名称和会议）
   - 与本论文的延续/突破关系

**特别注意**: 团队背景是最容易出现幻觉的维度。
- 所有奖项、项目归属、影响力定性（如"最具影响力"）必须有 L1/L2 来源支撑，否则不写
- **入职时间、实习经历、组织归属**等履历细节幻觉风险极高，除非在作者主页或论文首页明确写明，否则**不写**

{输出格式要求}

Agent 3: Technical Methods and Innovations

description: "Analyze the technical methods of {paper_title}"

prompt template:

{前置引用约束}

你是技术方法分析专家。请分析论文《{paper_title}》（{authors}, {year}）的核心技术方法和创新点。

论文摘要：{abstract}
论文原文（如有）：{paper_content}
论文定位映射：{paper_locator_map}
论文类型提示：{type_hint}
输出语言：{lang}

任务:
1. **优先从论文原文提取**：从 {paper_content} 中定位 Method / Approach / System Design 部分，提取：
   - 核心方法/算法/架构
   - 关键技术组件和工作流程
   - 与前人工作的区别（论文中的 comparison）
   - 标注具体 Section/Figure/Algorithm 编号
2. 使用 WebSearch 搜索补充解读（至少 1 次）:
   - "{paper_title} method explained" 或 "{paper_title} 方法解析"
   → 获取通俗解读作为补充（标注为 L3 外部解读）
3. 使用 WebFetch 访问至少 1 个技术解读页面

整理以下信息:

   **核心方法:**
   - 论文提出的核心方法/算法/架构（标注论文 Section/Figure）
   - 关键技术组件和工作流程
   - 数学/算法层面的核心思想（用直觉解释，避免纯公式）

   **创新点:**
   - 与前人工作的关键区别（列 2-3 点，标注论文中的论据）
   - 每个创新点为什么重要
   - 技术贡献的本质（新架构 / 新训练策略 / 新理论框架等）

{输出格式要求}

Agent 4: Experiments and Result Evaluation

description: "Analyze the experimental results of {paper_title}"

prompt template:

{前置引用约束}

你是实验评估专家。请分析论文《{paper_title}》（{authors}, {year}）的实验设计和结果。

论文摘要：{abstract}
论文原文（如有）：{paper_content}
论文定位映射：{paper_locator_map}
论文类型提示：{type_hint}
输出语言：{lang}

任务:
1. **必须从论文原文提取**：从 {paper_content} 中定位 Experiments / Evaluation / Results 部分，提取：
   - 实验设置（数据集、基线、指标）
   - 关键结果数字（标注具体 Table/Figure 编号）
   - 消融实验发现
2. 使用 WebSearch 搜索补充（至少 1 次）:
   - "{paper_title} results benchmark comparison"
   → 了解社区对实验结果的评价
3. 使用 WebFetch 访问至少 1 个 L1/L2 来源验证

整理以下信息:

   **实验设计:**
   - 使用的数据集/基准（标注论文 Section）
   - 基线方法（baseline）有哪些
   - 评估指标

   **关键结果:**
   - 主要实验结果（标注论文 Table/Figure 编号）
   - 与 baseline 的关键对比数字
   - 消融实验的主要发现（如有，标注 Table/Figure）
   - 结果是否充分验证了论文的核心 claim

**特别注意**: 实验数字是最需要精确的内容。所有数字必须能在论文原文的具体 Table/Figure 中找到，不得从博客或新闻稿中提取实验数据。

{输出格式要求}

Agent 5: Value Judgment and Boundary Analysis

description: "Analyze the value boundaries of {paper_title}"

prompt template:

{前置引用约束}

你是技术价值分析专家。请分析论文《{paper_title}》（{authors}, {year}）的实际价值边界——什么场景下有用，什么场景下没用，以及当前缺什么证据。

论文摘要：{abstract}
论文原文（如有）：{paper_content}
论文定位映射：{paper_locator_map}
论文类型提示：{type_hint}
输出语言：{lang}

任务:
1. **从论文原文提取**：定位 Discussion / Limitations / Conclusion / Applications 部分，提取：
   - 作者自述的局限性和适用边界（标注 Section）
   - 作者讨论的应用场景
2. 使用 WebSearch 搜索以下关键词（至少搜 2 次不同关键词）:
   - "{paper_title} limitations criticism use cases"
   - "{paper_title} review openreview" 或 "{paper_title} 不足 局限 应用"
3. 使用 WebFetch 访问至少 1 个包含论文评价的页面（优先 OpenReview 官方评审）

整理以下信息:

   **什么场景下有价值（列举 2-4 类具体场景）:**
   - 场景名 + 为什么这类方案在此场景下优于替代方案（标注论文中的支撑）
   - 越具体越好（如"跨文档多跳推理"而非"NLP 任务"）

   **什么场景下价值有限（列举 2-4 类具体场景）:**
   - 场景名 + 为什么现有替代方案（RAG、长上下文等）在此场景下更优
   - 标注作者自述（论文 Section）或外部评审指出的（来源层级）

   **当前最缺的证据:**
   - 列举 2-4 个使论文结论更可信所需但目前缺失的证据（如生产数据、独立复现、某类 benchmark）
   - 这些缺失如何影响对论文价值的判断

   **作者自述局限:**
   - 直接引用论文 Discussion/Limitations 段落（标注 Section）

{输出格式要求}

Agent 6: Community Feedback and Impact

description: "Research the community impact of {paper_title}"

prompt template:

{前置引用约束}

你是学术影响力分析专家。请研究论文《{paper_title}》（{authors}, {year}）的社区反馈和学术影响力。

论文摘要：{abstract}
输出语言：{lang}

任务:
1. 使用 WebSearch 搜索以下关键词（至少搜 2 次不同关键词）:
   - "{paper_title} site:scholar.google.com OR site:semanticscholar.org"
   - "{paper_title} github implementation" 或 "{paper_title} discussion reddit twitter"
2. 使用 WebFetch 访问：
   - **必须**: 至少 1 个 L2 来源（Google Scholar 或 Semantic Scholar）获取引用数据
   - **可选**: 社区讨论页面

整理以下信息:

   **学术影响力:**
   - 引用量（标注来源 + 抓取日期，如"Google Scholar, {date}, N citations"）
   - 被哪些重要后续工作引用（列出具体论文名称）
   - 是否获得最佳论文奖（仅从会议官方页面确认，无法确认则不写）

   **社区反馈:**
   - 业界/学界的主要评价（标注来源层级）
   - 开源实现情况（GitHub URL + star 数 + 抓取日期）
   - 工业界采用情况（仅写有 L1/L2 来源支撑的）
   - 争议或批评（标注来源）

**特别注意**:
- 本维度信息时效性最强，所有数字指标必须标注抓取日期和来源 URL
- 新论文（发布 < 6 个月）社区反馈可能不充分，如实标注"论文较新，数据有限"，不要编造
- L4 来源（新闻稿/聚合站）仅可作为"舆情信号"引用，不可作为事实依据，且必须标注"L4 低可信"
- 禁止从 L4 来源提取引用量、star 数等数字指标

{输出格式要求}

Agent 7: Glossary and Further Reading

description: "Compile the glossary and further reading for {paper_title}"

prompt template:

{前置引用约束}

你是学术知识整理专家。请整理论文《{paper_title}》（{authors}, {year}）的关键术语和推荐延伸阅读。

论文摘要：{abstract}
论文原文（如有）：{paper_content}
论文定位映射：{paper_locator_map}
论文类型提示：{type_hint}
输出语言：{lang}

任务:
1. **优先从论文原文提取**：从 {paper_content} 中提取：
   - 论文定义或频繁使用的关键术语
   - References 部分引用的核心论文
2. **历史谱系追踪**：从论文 Introduction / Related Work 中识别本文最直接的 2-3 篇前置工作（被引用最多、被明确对比的），梳理它们之间的演进关系（A 提出了 X → B 在此基础上改进了 Y → 本文进一步做了 Z），标注论文 Section 和 References 编号
3. 使用 WebSearch 搜索补充（至少 1 次）:
   - "{paper_title} related papers survey"
   → 了解该方向的前置和后续工作
4. 使用 WebFetch 访问至少 1 个 L1/L2 来源验证

整理以下信息:

   **术语索引表:**
   - 列出论文中 10-15 个关键术语/概念
   - 每个条目包含: 术语名称、简要定义（1-2 句）
   - 优先提取论文中有明确定义的术语（标注 Section）
   - 标注术语之间的关联关系

   **历史谱系（必须输出，供 Phase 2 合成技术地图）:**
   - 列出 2-3 篇最直接的前置工作，按时间排序
   - 每篇说明: 做了什么 → 本文相比它的关键差异是什么（标注论文 Section + References [N]）
   - 用箭头表示演进链: `[N1] 论文A (2022) → [N2] 论文B (2024) → 本文 (YYYY)`
   - 差异维度: 训练方式、检索粒度、扩展规模、工程架构等

   **延伸阅读:**
   - 前置阅读（理解本文需要先读的 2-3 篇论文，**必须从论文 References 中选取**，标注引用编号如 [12]）
   - 后续工作（基于本文的重要后续研究 3-4 篇，必须通过 WebSearch/WebFetch 确认真实存在）
   - 每篇包含: 标题、作者、年份、与本文的关联说明、**来源**（论文 References [N] / Google Scholar / Semantic Scholar）
   - **不得列出无法通过搜索确认的论文**——如果搜索不到，就不列

{输出格式要求}

Phase 2: Comprehensive Analysis (Main Thread)

After all 7 Agents have returned, complete the following analysis in the main thread.

Step 1: Information Aggregation and Fact Verification

Integrate the results of the 7 Agents and perform:

Source tier review: Check whether the assertions returned by each Agent are supported by qualified sources, and downgrade "facts" supported only by L4 sources to "to be verified"
Consistency signals: Multiple Agents mentioning the same conclusion from different sources = high confidence
Complementary relationships: Agent 1's motivation and Agent 3's method form a problem→solution correspondence
Contradictory signals: Inconsistent evaluations of a viewpoint from different sources = mark the controversy in the report
Organize allsource URLs, classified by tier

Step 1.5: Paper Location Mapping Table (Hard Gate)

Aggregate from the "assertion-location mapping tables" returned by all Agents, merge and deduplicate. Then classify and process according to the following rules:

Qualified entries(with Section/Table/Figure numbers): directly enter the report
Entries with only URLs and no numbers: Go back to the full text of {paper_content} to search and locate, and add the numbers
Entries that still cannot be located after searching: Mark as "location pending", and in the report change that number to a qualitative description (e.g., "significant improvement" instead of a specific multiple)
Entries from L4 sources: Remove from the facts paragraph; may only be cited in the community feedback paragraph

Summary format:

| 数字/断言 | 论文位置 | 来源 Agent | 状态 |
|-----------|---------|-----------|------|
| 1.87× 离线吞吐提升 | Table X, Section Y | Agent 4 | 合格 |
| 28.8× 算力增长 | Section 1, Figure Z | Agent 1 | 合格 |
| 某个数字 | 未定位 | Agent 3 | 降级为定性描述 |
| ... | ... | ... | ... |

Gating condition: When qualified entries < 10, it indicates the paper's original text was insufficiently extracted; you need to re-run WebFetch on the paper HTML and supplement the locations, rather than proceeding directly to Phase 3.

Step 2: Cross-correlation

Establish correlations among the 7 dimensions:

Motivation ↔ Method mapping: Map the background problem (Agent 1) one-to-one with the technical solution (Agent 3)
Method ↔ Experimental validation: Check whether the innovations (Agent 3) are fully validated in the experiments (Agent 4)
Value boundary ↔ Community feedback: Compare whether the value judgment (Agent 5) is consistent with the community evaluation (Agent 6)
Team background ↔ Research lineage: Annotate the relationship between the team's prior work (Agent 2) and this paper's continuation/breakthrough
Terminology ↔ Content anchoring: Verify the correspondence between the terminology (Agent 7) and the content of each dimension

Step 2.5: Technical map positioning

Based on the results of Agent 1 (background), Agent 3 (method), Agent 5 (value boundary), and Agent 7 (further reading), synthesize this paper's coordinates in the field:

Identify 3-5 competing routes: The main technical directions addressing the same core problem (e.g., window extension / external storage / compression / sparse retrieval / hybrid architecture); summarize each route's approach and cost in one sentence
Position this paper: Which route this paper is on, and what unique technical choices it made (extracted from Agent 3)
Historical lineage: Identify 2-3 of the most directly related prior works from Agent 7's prerequisite reading and the paper's References, and explain the key differences between this paper and them
Judge this paper's "recombination vs. originality" attribute: Does this paper's value come more from a new combination/recombination of existing ideas, or from a completely new principled breakthrough? This determines how to evaluate its contribution

Step 3: Distilling core conclusions

Based on the above analysis, write the core judgment (not a summary restatement, but a field-level positioning):

What this paper changed: What changed in our understanding of the field (not "proposed method X", but "moved problem Y from stage A to stage B")
Key technical choices: The 1-2 design decisions most worth noting and their trade-offs
Value boundary in one sentence: What scenarios it suits, and what scenarios it does not suit
Evidence maturity: What level of judgment the current evidence supports (strong signal / directional reference / to be verified)
What kind of reader it best suits

Phase 3: Output report

Output the final report in the following Markdown format:

# 《{paper_title}》论文深度分析

> **作者**: {authors} | **机构**: {institution} | **发表**: {venue} {year} | **分析日期**: {date}

## 核心判断

<!-- 硬格式：领域级定位，不是摘要复述 -->
- **这篇论文改变了什么**: {不是"提出了 X"，而是"让 Y 问题从 A 阶段进入 B 阶段"}（论文 Section X）
- **关键技术选择**: {1-2 个设计决策及其 trade-off}（论文 Section X, Table/Figure Y）
- **关键结果**: {一句话，含数字}（论文 Table X, Section Y）
- **价值边界**: {适合什么 / 不适合什么}
- **证据成熟度**: {强信号 / 方向参考 / 待验证} — {一句话理由}

---

## 一、背景与动机

### 研究领域现状

{该领域的发展脉络和现状，前人工作的贡献与局限}
{每个关键断言后标注来源，如（论文 Section 1, [2]）}

### 核心问题与动机

{论文要解决的核心问题、现有方法的不足、作者的切入角度}
{标注论文 Section}

### 为什么是现在

{使这项工作变得可能或必要的 2-3 条汇合趋势}
{每条趋势与本文的关系，标注来源}

---

## 二、团队与作者背景

### 核心作者

<!-- 仅写作者主页/Google Scholar/DBLP 可直接证实的内容，无法证实的不写 -->
{第一作者和通讯作者的背景、研究方向、代表作}
{所有荣誉/奖项标注验证来源（具体页面路径），无法从 L1/L2 验证的不写}

### 研究脉络

{团队前序工作与本文的关系、实验室研究特色}

---

## 三、技术方法与创新点

### 核心方法

<!-- 每个技术机制（如"使用 RDMA"、"采用 io_uring"）都必须标注描述该机制的论文 Section/Figure -->
{论文提出的核心方法/算法/架构（标注论文 Section/Figure）}
{关键技术组件和工作流程，每个组件标注论文 Section}

### 创新点

1. **{创新点一}**: {描述及为什么重要}（论文 Section X.Y）
2. **{创新点二}**: {描述及为什么重要}（论文 Section X.Y）
3. **{创新点三}**: {描述及为什么重要}（论文 Section X.Y）

### 技术地图定位

**该领域的主要路线:**

| 路线 | 核心思路 | 代价 | 代表工作 |
|------|---------|------|---------|
| {路线一} | {一句话} | {一句话} | {1-2 篇} |
| {路线二} | ... | ... | ... |
| **本文** | ... | ... | — |

**本文的历史谱系:**
- {最直接的前置工作 1} → 本文的差异: {一句话}
- {最直接的前置工作 2} → 本文的差异: {一句话}

---

## 四、实验与结果评估

### 实验设计

<!-- 论文位置列不可为空，无法定位则填"未定位" -->
| 项目 | 详情 | 论文位置 |
|------|------|---------|
| 数据集 | {使用的数据集/基准} | Section X（不可为空） |
| 基线方法 | {baseline 列表} | Section X（不可为空） |
| 评估指标 | {主要指标} | Section X（不可为空） |

### 关键结果

<!-- 硬格式：每条结果一行，数字紧跟引用，禁止段末堆链接 -->
- {指标名}: {数字}（论文 Table X, Section Y）
- {指标名}: {数字}（论文 Table X, Section Y）
- {指标名}: {数字}（论文 Figure X, Section Y）
- ...

### 消融实验

<!-- 硬格式：每条消融发现一行，同样逐个标注 -->
- {消融变量}: {去掉后的数字变化}（论文 Table X）
- {消融变量}: {去掉后的数字变化}（论文 Table X）
- ...

---

## 五、价值判断与边界

### 什么场景下有价值

<!-- 具体场景，不是泛泛的"NLP 任务" -->
1. **{场景一}**: {为什么这类方案在此场景下优于替代方案}（标注论文支撑）
2. **{场景二}**: ...
3. ...

### 什么场景下价值有限

1. **{场景一}**: {为什么替代方案更优，标注来源}
2. **{场景二}**: ...
3. ...

### 当前最缺的证据

- {证据缺口一}: {缺失如何影响判断}
- {证据缺口二}: ...

### 作者自述局限

{直接引用论文 Discussion/Limitations（标注论文 Section）}

---

## 六、社区反馈与影响力

### 学术影响力

{引用量（来源 + 抓取日期）、重要后续工作、获奖情况（仅官方确认的）}

### 社区反馈

**已验证事实:**
- {可独立验证的客观信息，如平台收录、讨论帖链接、开源仓库 URL + star + 抓取日期}

**外部讨论:**
- {来自社区的具体评论或观点}（来源层级, URL）

**分析判断:**
- 分析认为：{基于以上事实的推理和综合判断}

---

## 七、术语表与延伸阅读

### 术语索引

<!-- 论文位置列不可为空，无法定位则填"未定位" -->
| 术语 | 定义 | 论文位置 |
|------|------|---------|
| {术语1} | {简要定义} | Section X（不可为空） |
| {术语2} | {简要定义} | Section X（不可为空） |
| ... | ... | ... |

### 延伸阅读

**前置阅读（理解本文的基础）:**

<!-- 前置阅读必须来自论文 References，标注引用编号 -->
| 论文 | 作者 | 年份 | 关联说明 | 来源 |
|------|------|------|----------|------|
| {论文1} | {作者} | {年份} | {与本文的关联} | References [N] |

**后续工作（已发表的相关研究）:**

<!-- 后续工作必须通过搜索确认真实存在，不得编造 -->
| 论文 | 作者 | 年份 | 关联说明 | 来源 |
|------|------|------|----------|------|
| {论文1} | {作者} | {年份} | {与本文的关联} | Google Scholar / Semantic Scholar |

**潜在研究方向（未发表/作者设想）:**

- {方向1}：{简要描述}（来源：论文 Section X / 作者公开讨论）
- {方向2}：{简要描述}

---

## 判断

<!-- 不是摘要总结，是基于全文分析后的领域级判断 -->

{2-3 段，回答: 这篇论文真正提醒我们的是什么？它对该领域的问题分工/技术路线选择意味着什么？对于 builder/研究者，最值得持续追问的问题是什么？}

---

## 论文定位映射（paper_locator_map）

<!-- 必须原样输出 Phase 0 生成的 paper_locator_map，作为引用审计基准 -->
<!-- 正文中所有 `（论文 Section/Table/Figure X）` 引用的编号必须出现在此 map 中 -->

{paper_locator_map 原文}

---

## 断言与证据清单

<!-- 必须覆盖正文中出现的所有阿拉伯数字，无遗漏 -->
| # | 关键断言 | 论文位置（须匹配 paper_locator_map） | 外部来源 | 来源层级 | 置信度 | 性质 |
|---|---------|--------------------------------------|---------|--------|--------|------|
| 1 | {断言描述} | Section X, Table Y | — | L1 | High | 论文原文 |
| 2 | {断言描述} | — | {URL, 抓取日期} | L2 | Medium | 外部验证 |
| 3 | {断言描述} | — | — | — | Low | 分析推断 |
| ... | ... | ... | ... | ... | ... | ... |

> 置信度说明：**High** = L1 来源直接支撑；**Medium** = L2 来源支撑或多个 L3 来源交叉验证；**Low** = 仅有 L3/L4 来源或为分析推断
> **审计方法**: "论文位置"列中的每个 Section/Table/Figure 编号必须能在上方 paper_locator_map 中找到对应条目。找不到 → 该行标红待查。

---

## 来源登记表

<!-- 所有被引用的来源必须登记，便于审查"来源是否被滥用" -->
| # | 来源 | 层级 | 用于支撑 | 抓取日期 |
|---|------|------|---------|---------|
| 1 | 论文原文: [{paper_title}]({url}) | L1 | 方法、实验、局限性 | — |
| 2 | {作者主页 URL} | L1 | 团队背景 | {date} |
| 3 | {Google Scholar URL} | L2 | 引用量 | {date} |
| 4 | {技术博客 URL} | L3 | 外部解读 | {date} |
| ... | ... | ... | ... | ... |

> **L4 来源不得出现在此登记表中。** 如需引用 L4 来源，仅可在正文"社区反馈 > 外部讨论"中内联标注 `（L4 低可信）`，不登记。常见 L4 来源：`bitcoinethereumnews.com`、`aihaberleri.org`、`MEXC News`、`36kr`、`新浪科技`、`coindesk`（非技术原文时）。

---

*本笔记由 AI 自动生成。关键数字和事实已尽量标注论文原文位置，建议结合原文验证。"断言与证据清单"可快速定位需要核查的内容。*

Phase 3.5: Report validation and repair (must be executed before writing to file; cannot be skipped)

Step 1: Assertion citation scan

Scan the report line by line, forevery Arabic numeralandevery specific technical mechanism description:

Immediately followed by （论文 Section/Table/Figure X） and the number exists in paper_locator_map → pass
The paragraph ends with a link but the number/assertion itself has no annotation → fail(This is the "pile links at paragraph end" anti-pattern) → move the annotation after the number/claim
Has URL but no Section number → go back to the Step 1.5 mapping table to look it up, add the number or change it to a qualitative description
No annotation at all → add an annotation or delete the number/claim

Non-conforming patterns that must be caught(scan the following pattern character by character):

Table or Figure afterno number → add a number or change it to a Section reference
论文 Table, Section X — Table missing number (before the comma it should be Table N) → add the Table number or keep only the Section
（论文 Section X） X in not in paper_locator_map → recheck the paper, correct the number or mark "not located"
（来源: arxiv html） or （来源: 论文） — this is not a locator citation → replace with a specific Section/Table number
The referenced Table/Figure number does not exist in paper_locator_map → recheck the paper or delete the reference

Scanning scope with special attention to: core conclusions, key results, ablation experiments, and specific mechanism names in the technical method

Step 2: Source tier enforcement

Scan the entire report line by line (not just the "Source:" lines), searching for the following L4 domain keywords:

36kr / ResearchGate / MEXC / aihaberleri / bitcoinethereumnews / Eyerys / 新浪科技 / coindesk / techcrunch(when it is not a first-hand technical article)
The above domains appear inteam background, technical method, experimental evaluation, pros and consparagraph → remove the claims supported by that source, or move the claims into "Community Feedback > External Discussion" and mark （L4 低可信）
The above domains appear inreference sourcessection → remove; L4 does not enter the reference sources section

Step 3: Structural completeness

The experimental design table has a "Paper Location" column and it is non-empty
The terminology index table has a "Paper Location" column and it is non-empty
Value judgments are divided into: valuable scenarios / limited-value scenarios / evidence gaps / author-stated limitations
Community feedback is divided into: verified facts / external discussion / analytical judgment
The source registry exists, each row has a tier and a "used to support" column, and there are no L4 sources
The claims and evidence list exists and has ≥ 10 entries

Step 4: Team background review

No 36kr/ResearchGate/media article sources
Resume details that cannot be directly verified from the author's homepage/Scholar → delete (do not mark "pending verification")
"According to reports"/"According to ... reports"/"reported"/"according to media"/"it has been reported that" → if the immediately following content has no L1/L2 source parenthetical annotation → delete the entire sentence("According to reports" itself is an L4 signal phrase)
specific hardware procurement quantity(e.g., "procured N GPUs") → unless the source is an official company announcement or SEC filing → delete

Step 5: Other checks

Inferences are marked: comparative/evaluative expressions ("faster""significant""model example") have a source annotation, or are marked with "analysis suggests"
Further reading: prerequisite reading must come from the paper's References (marked [N]); follow-up work must have a source column (Google Scholar / Semantic Scholar); entries whose existence cannot be confirmed are deleted
Time-sensitive metrics: citation counts/stars/upvotes are annotated with the capture date and source URL
No placeholders/fabricated URLs: all source links are genuinely accessible

Step 6: Downgrade items that fail (if not fixed, downgrade; if not downgraded, delete)

For items that fail as identified in Steps 1-5, force downgrade according to the following rules:

Key numbers with no paper location → Delete the number, change to a qualitative description (e.g., "significantly improved")
Technical claims with no paper location → Delete the claim, or add an "analysis suggests" marker to downgrade it to an inference
Team background with only L3/L4 sources → Delete the claim, compress the entire paragraph to "insufficient information"
Team background uses qualitative evaluations("most influential", "key figure", etc.) and is not supported by official awards → Delete
Community opinions with only L4 sources → Rewrite as "media have reported on this work"; do not summarize industry evaluations
Further reading whose existence cannot be confirmed → Delete the item

Step 7: Evidence coverage check (final gate)

Calculate the following metrics; if they do not meet the standard, the final report must not be output:

Arabic numeral coverage: All Arabic numerals in the main text must be 100% included in the "Claims and Evidence List". If not met → complete them before outputting
L1 coverage of core conclusions: At least 3 of the 5 core conclusions must be supported by L1 sources. If not met → rewrite the core conclusions and delete items without L1 support
Experimental number localization rate: At least 80% of the numbers in the experimental section (key results + ablation) must have （论文 Table/Figure X）. If not met → add at the beginning of the report ⚠️ 实验结果定位不完整，建议对照原文验证
paper_locator_map reference consistency: All Section/Table/Figure numbers referenced in the "Paper Location" column of the Claims and Evidence List must exist in paper_locator_map. If there is a mismatch → correct the number or mark it "not located"
paper_locator_map existence: The final report must include a "Paper Location Mapping" section and it must be non-empty. If missing → add it from the Phase 0 output

Save the report as a file: ./{paper_name_en}-paper-analysis-{date}.md, saved in the current working directory.

File name rule: {paper_name_en} Use an abbreviated English form of the paper title (lowercase, hyphen-separated), for example:

Attention Is All You Need → attention-is-all-you-need
BERT: Pre-training of Deep Bidirectional Transformers → bert
GPT-4 Technical Report → gpt-4-technical-report

Error Handling

Agent timeout or failure: If an Agent does not return a result, mark that dimension in the report as "insufficient data; readers are advised to consult it themselves"; analyze the other dimensions as usual
Paper full text unavailable: Mark at the beginning of the report "⚠️ The full paper text was not obtained; the following analysis is based on the abstract and external sources and has low credibility", and set the confidence ceiling of all claims in the claims list to Medium
Paper cannot be identified: Prompt the user to confirm the paper title or provide a link; fuzzy matching suggestions may be given
Insufficient information for a new paper: If search results are very few (e.g., a just-released preprint), merge related dimensions and reduce repeated "insufficient information" prompts; mark the community feedback dimension as "the paper is relatively new, and community feedback is not yet sufficient"
Multiple versions of a paper: If the same paper has multiple versions (e.g., arXiv v1/v2), prioritize analyzing the latest version and note the version information

Notes

Each Agent's output is limited to 800 characters; when exceeded, prioritize retaining content supported by first-hand sources
The comprehensive analysis stage focuses on cross-correlation and fact-checking, rather than simply concatenating the results of 7 dimensions
The core conclusion should help readers judge "whether this paper is worth my time to read closely"
The output language follows {lang}, defaulting to Chinese
All source links must be genuinely accessible and must not be fabricated
The "assertions and evidence checklist" is the core value that distinguishes this skill from ordinary paper summaries—it lets readers know which content is trustworthy and which needs to be verified on their own

About this skill

When to Use

Source credibility rules (global, applicable to all Phases/Agents)

Primary source priority principle

Source layering

In-paragraph citations and assertion annotation

Phase 0: Paper positioning and input parsing

Input path

URL Input Processing

PDF Input Processing

Title/Keyword Input Processing

Obtain the original paper text (key step, cannot be skipped)

Generate the paper location map (required, cannot be skipped)

Determine key variables

Paper type adaptation tips

Phase 1: Parallel Data Collection (7 Subagents)

Agent 1: Background and Motivation

Agent 2: Team and Author Background

Agent 3: Technical Methods and Innovations

Agent 4: Experiments and Result Evaluation

Agent 5: Value Judgment and Boundary Analysis

Agent 6: Community Feedback and Impact

Agent 7: Glossary and Further Reading

Phase 2: Comprehensive Analysis (Main Thread)

Step 1: Information Aggregation and Fact Verification

Step 1.5: Paper Location Mapping Table (Hard Gate)

Step 2: Cross-correlation

Step 2.5: Technical map positioning

Step 3: Distilling core conclusions

Phase 3: Output report

Phase 3.5: Report validation and repair (must be executed before writing to file; cannot be skipped)

Step 1: Assertion citation scan

Step 2: Source tier enforcement

Step 3: Structural completeness

Step 4: Team background review

Step 5: Other checks

Step 6: Downgrade items that fail (if not fixed, downgrade; if not downgraded, delete)

Step 7: Evidence coverage check (final gate)

Error Handling

Notes