field-papers-survey

Field Papers Survey — one-click panoramic survey of papers in a field

Enter the name of a research field, automatically survey the core conferences/journals, collect important papers in each subfield in parallel, organize and output from three perspectives: classification by subfield, chronological timeline, and development trajectory, and automatically perform lightweight in-depth analysis of the Top papers.

When to Use

Triggered when the user requests the following operations:

"survey computer vision papers" / "survey NLP papers"
"NLP field paper survey" / "recommender systems paper panorama"
"survey computer vision papers" / "survey NLP papers"
"what important papers are there in computer vision" / "what classic papers are there in machine learning"
"survey computer science-systems direction papers" (complete subfields + only go deep into the systems direction)

Phase 0: Parse input

Based on the user input, identify the following information:

Field name(required)
- Extract the field name from the user message
- "survey computer vision papers" → computer vision
- "survey NLP papers" → natural language processing
Focus direction(optional)
- "survey computer science-systems direction papers" → focus direction = computer systems
- If unspecified, it is a comprehensive survey, and all subfields are collected in depth
Whether to skip in-depth analysis
- Default:automatically execute Phase 4 (perform lightweight in-depth analysis of the Top papers)
- Skip Phase 4 only when the user explicitly says "list only"/"do not analyze"/"skip analysis"/"only want the paper list"
- {skip_analysis}: true/false
Determine key variables(needed in all subsequent phases)
- {field}: Chinese name of the field (e.g., "computer vision")
- {field_en}: English name of the field, lowercase hyphen-separated (e.g., "computer-vision")
- {focus}: name of the focus direction (optional, e.g., "object detection"; empty if unspecified)
- {date}: current date, format YYYY-MM-DD
- {skip_analysis}: whether to skip Phase 4 (default false)
Output directory: current working directory

If the field name or its corresponding English name cannot be determined, first use one WebSearch for a quick query,do not skip this step。

Phase 1: Panoramic review of conferences/journals (main thread)

Through the four-channel information source framework, comprehensively review the core conferences and journals in the field, and identify the subfield divisions.

Four-channel information source framework

Channel	Information source type	Search strategy	Priority
Academic organizations	CCF recommended list, CORE ranking, CAS partition table, JCR, IEEE/ACM classification	`"{field}" CCF 推荐会议期刊` / `"{field}" CORE ranking`	Highest
Academic indexes	Google Scholar Metrics、DBLP、Semantic Scholar、CSRankings	`"{field}" top conferences journals h5-index`	High
Community consensus	awesome-xxx lists, venue distribution of well-known survey papers, field blogs	`"{field}" top venues survey` / `"{field}" 顶会顶刊推荐`	Medium
Citation network	Publication venue distribution of highly cited papers, best paper award history	`"{field}" best paper award` / `"{field}" most cited venue`	Verification

Domain-information source mapping example

Domain	Academic organization	Academic index	Community consensus	Citation network
Computer vision	CCF recommended catalog	Google Scholar Metrics	awesome-cv	CVPR/ICCV/ECCV best papers
NLP	CCF recommended catalog	ACL Anthology	awesome-nlp	ACL/EMNLP/NAACL best papers
Machine learning	CCF/CORE	CSRankings	awesome-ml	NeurIPS/ICML/ICLR best papers
Databases	CCF recommended catalog	DBLP	db-readings	SIGMOD/VLDB best papers
Biomedicine	JCR/CAS partition	PubMed/SJR	—	Highly cited paper distribution
Economics	SSCI/CAS partition	RePEc/IDEAS	—	Highly cited paper distribution

Execution requirements

Execute a total of 5-8 WebSearch + 2-3 WebFetch, covering four channels
Search order: first academic organizations (the most authoritative grading standards), then academic indexes (quantitative indicators), then community consensus and citation networks
WebFetch priority access: CCF recommended catalog official page, Google Scholar Metrics ranking page

Adaptation for non-CS domains: The above search strategy applies by default to CS and adjacent domains. For non-CS domains (such as biomedicine, materials science, economics), CCF/CORE should be replaced with the corresponding ranking system for that domain (such as SJR, Impact Factor, JCR partition), and CSRankings should be replaced with the ranking tool for that domain.

Language strategy: Academic papers are mainly in English, and search keywords should use both Chinese and English. For directions with strong Chinese characteristics (such as Chinese information processing), Chinese search should be the main approach.

Subfield identification

In the process of sorting out venues, simultaneously identify the subfield divisions of the domain (8-15). Cross-validation rules:

≥2 channels mention = mainstream subfield
Only 1 channel = emerging/niche subfield

Organize each subfield as: number, subfield name, one-sentence description, core research questions The total number of subfields is expected to be 8-15; if there are more than 15, merge highly related subfields

Output format

Organize the sorting results into a structured subfield list:

子领域 N: {子领域名称}（{英文名}）
- 核心研究问题：{描述}
- 核心 Venue：
  | Venue 名称 | 类型 | 等级 | h5-index | 说明 |
  |------------|------|------|----------|------|
  | {name}     | 会议/期刊 | CCF-A/CORE-A* | {h5} | {一句话描述} |
- 信息源：{哪些渠道提及}
- 分类：主流子领域 / 新兴子领域

Focused direction handling: If the user specifies {focus}, still output the complete subfield list, but only mark {focus} corresponding subfields to enter Phase 2 in-depth collection. If not specified, all subfields enter Phase 2.

Phase 2: Parallel paper collection (N Subagents)

Key requirement: Agents for all subfields must be placed inthe same messageto ensure parallel execution.

Each Agent uses subagent_type: "general-purpose"。

Send 1 subagent for each subfield to collect important papers through the following 4 channels (in decreasing priority):

Priority	Channel	Collection target	Search strategy
1 (core)	Best Paper	Best paper awards of each venue	`"{venue_name}" best paper award list`
2 (important)	Highly cited classics	Top-cited papers, milestone papers	`"{sub_field}" most cited papers` / `"{sub_field}" seminal papers`
3 (supplementary)	Survey citations	Papers frequently cited in authoritative surveys	`"{sub_field}" survey review` → extract from reference
4 (Verified)	Community recommendations	Awesome lists, academic community recommendations	`"{sub_field}" must read papers` / `"{sub_field}" 必读论文`

Language strategy: Academic papers are primarily in English; use both Chinese and English search keywords. For directions with strong Chinese characteristics (such as Chinese information processing), prioritize Chinese searches.

Sub-agent prompt template

Each subfield sub-agent uses the following complete prompt:

description: "Collect important papers in the {field} · {sub_field_name} direction"

Prompt template:

你是学术论文调研专家。请为「{field}」领域的「{sub_field_name}」方向采集重要论文。

子领域编号：{sub_field_number}
子领域描述：{sub_field_description}
核心 Venue 列表：{venue_list}（来自 Phase 1）

任务：按以下 4 个渠道依次采集，至少执行 3 次 WebSearch + 1 次 WebFetch：

**渠道 1 — Best Paper（核心）**
搜索关键词：
- "{venue_name}" best paper award list
- "{sub_field_name}" best paper history
目标：找到各核心 venue 的最佳论文奖获奖论文。

**渠道 2 — 高被引经典（重要）**
搜索关键词：
- "{sub_field_name}" most cited papers
- "{sub_field_name}" seminal influential papers
目标：找到被引量最高的里程碑论文。

**渠道 3 — Survey 引用（补充）**
搜索关键词：
- "{sub_field_name}" survey review paper
目标：找到权威综述，从其 reference 中提取高频引用的论文。

**渠道 4 — 社区推荐（验证）**
搜索关键词：
- "{sub_field_name}" must read papers
- "{sub_field_name}" 必读论文 推荐
目标：通过社区推荐验证论文重要性，补充遗漏。

使用 WebFetch 访问至少 1 个搜索结果页面，验证论文信息的准确性。

输出格式（严格遵守）：

**{sub_field_number}. {sub_field_name} 重要论文**

| # | 标题 | 作者 | 年份 | Venue | 被引 | 脉络位置 | 承接关系 | 贡献描述 | TOP |
|---|------|------|------|-------|------|---------|---------|---------|-----|
| 1 | {标题} | {作者} | {年份} | {venue} | {被引} | {奠基/突破/改进/集大成/新范式} | {← 论文X / → 论文Y} | {一句话贡献} | [TOP] 或空 |

**来源链接：**
- [{来源标题}]({url})
- ...

输出要求：
- 8-15 篇论文，按重要性排序：奠基之作 > 突破性论文 > 高被引经典 > 近年重要工作
- 标记 1-2 篇最重要的论文为 [TOP]（将进入自动深度分析）
- 总输出控制在 2500 字符以内
- 全部使用中文输出
- 来源链接必须为实际访问过的 URL，不得使用占位符
- 如某渠道无结果，注明"该渠道未找到相关论文"并跳过

Sub-agent constraints

Each sub-agent outputs 8-15 papers
Output character limit 2500 characters
All output in Chinese
Must indicate source URL
Ordering: foundational works > breakthrough papers > highly cited classics > important recent work
Top marking: 1-2 papers [TOP]

Phase 3: Integration and Output (main thread)

After all sub-agents return, complete the following steps in the main thread.

Step 1: Data aggregation

Cross-subfield deduplication: If the same paper appears in multiple subfields → mark it as a "cross-domain classic" and mention it specifically in the core conclusions
Consistency check: Verify the consistency of paper information (venue, year, citation count); keep the same paper in the most relevant subfield, and cite it in other subfields as "see XX subfield"
Source completeness: Check whether the source URLs for all subfields are complete; mark missing ones as "source to be supplemented"
Sub-agent failure handling: Mark subfields that do not return results as "insufficient data, manual supplementation recommended"
Statistical summary: Calculate the total number of papers, the number of cross-domain classics, and the number of Top papers, for use in the core conclusions and terminal summary

Step 2: Three-perspective organization

Perspective 1 — Subfield classification: Display the paper list grouped by subfield (corresponding to the subdirectory structure)

Perspective 2 — Timeline: Display in segments by era, with segmentation automatically adjusted according to the field's pace of development (e.g., ~2000 / 2001-2010 / 2011-2017 / 2018-present), marking the representative papers and key turning points of each period

Perspective 3 — Development trajectory: Present the succession relationships among papers in the form of a technology evolution roadmap:

问题定义 → 经典方法 → 深度学习突破 → 预训练范式 → 当前前沿
[论文A]     [论文B,C]   [论文D]          [论文E,F]     [论文G]

Step 3: Generate the main document

Filename:{field_en}-papers-survey-{date}.md, saved in the current working directory.

Use the Write tool to write the following Markdown content:

# {field} 核心会议/期刊与重要论文全景

> 调研日期：{date}
> 文档目标：梳理{field}领域的核心 venue 和重要论文，提供子领域分类、时间线、发展脉络三种视角

## 核心结论

{基于四渠道调研得出的 3-5 条核心发现，包括：子领域数量、总论文数、跨领域经典论文数量、最具影响力的 venue、领域发展的关键转折点、推荐入门阅读路径}

## 会议/期刊全景

| Venue | 类型 | 等级 | h5-index | 主要覆盖子领域 | 说明 |
|-------|------|------|----------|---------------|------|
| {name} | 会议/期刊 | CCF-A | {h5} | {子领域} | {说明} |
| ... | ... | ... | ... | ... | ... |

## 发展脉络总览

{技术演进路线图，展示关键论文之间的承接关系，按阶段组织}

## 时间线

### {时段1}: {时段主题}
{该时段的代表性论文和关键事件}

### {时段2}: {时段主题}
...

## {编号}. {子领域名称}（{英文名}）

**核心研究问题**：{该子领域的核心研究问题描述}

**核心 Venue**：{该子领域的主要会议/期刊列表}

### 重要论文

| # | 标题 | 作者 | 年份 | Venue | 被引 | 脉络位置 | 承接关系 | 贡献描述 | 状态 |
|---|------|------|------|-------|------|---------|---------|---------|------|
| 1 | {标题} | {作者} | {年份} | {venue} | {被引} | {脉络位置} | {承接关系} | {贡献} | 待分析/已分析 |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |

{对每个子领域重复上述区块}

## 参考来源

**学术组织：**
- [{来源标题}]({url})

**学术索引：**
- [{来源标题}]({url})

**社区共识：**
- [{来源标题}]({url})

**引用网络：**
- [{来源标题}]({url})

**各子领域论文采集：**
- {子领域1}: [{来源标题}]({url}), ...
- {子领域2}: [{来源标题}]({url}), ...
- ...

## 下一步

本文档为领域论文全景调研结果，每个子领域的论文已整理在对应子目录中。如需对某篇论文进行深度分析，可使用 `/analyze-paper` 技能：

1. 进入感兴趣的子领域子目录
2. 查看该子领域的 README.md 了解论文列表
3. 对目标论文执行 `/analyze-paper {论文标题或URL}` 生成深度分析

建议优先从"核心结论"中推荐的入门路径开始，逐步扩展到其他子领域。

Markdown filling rules:

Replace placeholders in the template with actual data
The core conclusions are written based on the results of four-channel cross-validation and must be specific and actionable
Cross-domain classic papers are listed in detail in the subfield where they first appear, and other subfields note "see XX subfield"
The "Reference Sources" section is required: Fill in all source links from Phase 1 and Phase 2 grouped by category; do not omit them and do not use placeholder URLs

Step 4: Create the subdirectory skeleton

Use the Bash tool to create the directory structure.

Naming rules：01-{sub_field_en}/, 02-{sub_field_en}/, ...

{sub_field_en} Use the English names identified in Phase 1, lowercase and hyphen-separated (e.g., object-detection、natural-language-processing）。

Example (computer vision):01-image-classification/, 02-object-detection/, 03-semantic-segmentation/, ...

Create in each subdirectory README.md, use the Write tool to write the following content:

# {子领域名称}（{英文名}）

> 所属领域：{field}
> 调研日期：{date}

## 子领域简介

{核心研究问题、代表性 venue}

## 重要论文

| # | 标题 | 作者 | 年份 | Venue | 被引 | 脉络位置 | 承接关系 | 贡献描述 | 状态 |
|---|------|------|------|-------|------|---------|---------|---------|------|
| 1 | {标题} | {作者} | {年份} | {venue} | {被引} | {脉络位置} | {承接关系} | {贡献} | 待分析 |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | 待分析 |

## 深度分析

如需对某篇论文进行深度分析，可使用：
- `/analyze-paper {论文标题或URL}` — 生成七维度完整深度分析

Step 5: The terminal outputs progress information, then immediately continues

Print the following progress information in the terminal:

📋 Phase 1-3 完成：{field} 论文全景调研已生成

📄 主文档：{field_en}-papers-survey-{date}.md
📁 子目录：{N} 个子领域目录已创建
📚 论文总计：{总论文数} 篇（跨领域经典 {N} 篇）
🏆 Top 论文：{M} 篇待深度分析

目录结构：
  {列出所有子目录名称}

⚠️ Note: After printing the above progress information, the task is not yet complete.

If {skip_analysis} is true: print "ℹ️ Paper deep analysis skipped (user specified list only)", then end.
Otherwise:immediately begin executing Phase 4, and must not stop.

Phase 4: Automatic lightweight paper analysis (executed by default)

Skip this phase only when the user explicitly says "list only"/"no analysis"/"skip analysis"/"only want the paper list". By default, Phase 4 is automatically entered after Phase 3 is completed.

And /analyze-paper difference from: Phase 4 islightweight analysis—each paper is serially processed by 1 Agent across 7 dimensions (rather than /analyze-paper 's 7 parallel Agents + paper full-text extraction + assertion verification). The output quality is suitable for a quick overview; if complete deep analysis is needed (including paper_locator_map, assertion-evidence list, Phase 3.5 verification), the user should separately execute on the target paper /analyze-paper。

total cap

Phase 4 analyzes at most 10papers. If the total number of TOP papers across all subfields exceeds 10, process them in subfield order with priority, and keep the excess in "pending analysis" status.

Execution strategy

By subfieldone by oneprocess (to avoid launching too many subagents at the same time), and within each subfield the Top papers areparallelanalyzed:

子领域1: Top论文A ──┐
         Top论文B ──┼── 并行 → 全部完成 → 子领域2: Top论文C ──┐
                    ┘                              Top论文D ──┼── 并行 → ...
                                                              ┘

If the user specified a focus direction {focus}, then execute Phase 4 only for that subfield.

Processing for each subfield

For all Top papers in subfield M (assuming there are K papers), inthe same messageissue K Agent calls: