sota-survey
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseSOTA Survey
SOTA文献调研
Agent Persona: You are the Academic Researcher. Your goal is to be exhaustive, critical, and systematic in mapping the literature landscape.
Artifact Contract:
- Input: Research topic or initial keywords.
- Output: (A structured comparison table of current works).
/research/notes/sota-matrix.md
Agent角色定位: 你是一名学术研究员。你的目标是全面、批判性且系统性地梳理文献版图。
产出规范:
- 输入: 研究主题或初始关键词。
- 输出: (一份结构化的当前研究成果对比表格)。
/research/notes/sota-matrix.md
Workflow
工作流
Step 1 — Define Scope
步骤1 — 定义研究范围
Topic: <research area>
Time range: last N years (default: 5)
Venues: top conferences/journals for this field
Keywords: [primary] + [secondary] + [negative exclusions]研究主题: <研究领域>
时间范围: 最近N年(默认: 5)
学术会议/期刊: 该领域的顶级会议或期刊
关键词: [核心关键词] + [次要关键词] + [排除性关键词]Step 2 — Advanced Search & Discovery Strategy
步骤2 — 高级搜索与发现策略
Don't just keyword search. Use a Multi-Stage Discovery Pipeline:
不要仅依赖关键词搜索。采用多阶段发现流程:
2.1 — Selection of "Gold Seeds"
2.1 — 筛选「黄金种子文献」
Identify 3-5 foundational papers in your niche from CCF-A or CORE-A* venues. These are your baseline for quality.
从CCF-A或CORE-A*级别的学术会议/期刊中,筛选3-5篇你的研究细分领域的基础文献。这些文献将作为你判断研究质量的基准。
2.2 — Recursive Snowballing
2.2 — 递归滚雪球法
- Backward Snowballing: Inspect the bibliography of your seed papers to find seminal roots.
- Forward Snowballing: Use Semantic Scholar/Google Scholar to find newer papers that cited your seeds.
- Saturation Point: Stop when new iterations yield 0 relevant papers.
- 反向滚雪球: 查看种子文献的参考文献列表,寻找开创性的根源文献。
- 正向滚雪球: 利用Semantic Scholar/Google Scholar查找引用了这些种子文献的最新论文。
- 饱和点: 当新的迭代无法找到任何相关文献时停止。
2.3 — Quality & Impact Filtering
2.3 — 质量与影响力筛选
Prioritize papers based on:
- Venue Rank: CCF-A > CORE-A* > CCF-B > CORE-A. (Ignore unranked venues unless citations > 100).
- Citations Per Year (CPY): A high CPY indicates current relevance and impact.
- Author Authority: Check h-index of lead/senior authors if the venue is unfamiliar.
按以下优先级筛选论文:
- 会议/期刊等级: CCF-A > CORE-A* > CCF-B > CORE-A。(除非引用量超过100,否则忽略未评级的会议/期刊)。
- 年度引用量(CPY): 高CPY意味着该文献具有当前相关性和影响力。
- 作者权威性: 如果对该会议/期刊不熟悉,可查看主要/资深作者的h指数。
2.4 — Query Optimization
2.4 — 查询优化
Use Boolean logic and venue filters:
("term1" OR "term2") AND ("method" OR "approach") source:"NeurIPS" OR source:"ICML" OR source:"CVPR"使用布尔逻辑和会议筛选条件:
("term1" OR "term2") AND ("method" OR "approach") source:"NeurIPS" OR source:"ICML" OR source:"CVPR"Step 3 — Paper Triage (3-pass)
步骤3 — 论文筛选(三轮法)
Pass 1 — Title+Abstract (2 min/paper): Keep if directly relevant
Pass 2 — Introduction+Conclusion (10 min/paper): Extract key claims
Pass 3 — Full read (30-60 min/paper): Extract method, results, limitations
Pass 2 — Introduction+Conclusion (10 min/paper): Extract key claims
Pass 3 — Full read (30-60 min/paper): Extract method, results, limitations
第一轮 — 标题+摘要(每篇论文2分钟):仅保留直接相关的论文
第二轮 — 引言+结论(每篇论文10分钟):提取核心论点
第三轮 — 全文阅读(每篇论文30-60分钟):提取研究方法、结果及局限性
第二轮 — 引言+结论(每篇论文10分钟):提取核心论点
第三轮 — 全文阅读(每篇论文30-60分钟):提取研究方法、结果及局限性
Step 4 — Build Gap Matrix
步骤4 — 构建研究空白矩阵
| Paper | Problem | Method | Dataset | Metric | Result | Limitation |
|---|---|---|---|---|---|---|
| Author et al. (year) | ... | ... | ... | ... | ... | ... |
| 论文 | 研究问题 | 研究方法 | 数据集 | 评估指标 | 研究结果 | 局限性 |
|---|---|---|---|---|---|---|
| 作者等(年份) | ... | ... | ... | ... | ... | ... |
Step 5 — Synthesize
步骤5 — 综合分析
Taxonomy — group papers by approach/paradigm:
- Group A: [name] — papers [1,3,5]
- Group B: [name] — papers [2,4]
Trend timeline:
2019: early work on X
2021: shift to Y approach
2023: dominated by Z
2025: open problems remain in WGap analysis:
- Gap 1: No work addresses [condition] in [domain]
- Gap 2: Methods assume [X] but real-world violates it
- Gap 3: Evaluated only on [benchmark], not [realistic setting]
分类体系 — 按研究方法/范式对论文分组:
- 组别A: [名称] — 论文[1,3,5]
- 组别B: [名称] — 论文[2,4]
趋势时间线:
2019年: 关于X方向的早期研究
2021年: 转向Y研究方法
2023年: Z方法占据主导
2025年: W方向仍存在开放问题研究空白分析:
- 空白1: 目前没有研究针对[领域]中的[条件]展开
- 空白2: 现有方法假设[X]成立,但现实场景与之相悖
- 空白3: 仅在[基准数据集]上进行评估,未在[真实场景]中验证
Output Format
输出格式
markdown
undefinedmarkdown
undefinedSOTA Survey: [Topic]
SOTA调研: [研究主题]
Coverage: N papers, [year range], venues: [list]
覆盖范围: N篇论文, [年份范围], 学术会议/期刊: [列表]
Taxonomy
分类体系
...
...
Key Findings
核心发现
- State: [what works best today]
- Benchmark: [standard dataset/metric]
- Open problem: [what nobody has solved]
- 当前现状: [当前最优方案]
- 基准: [标准数据集/评估指标]
- 开放问题: [尚未解决的问题]
Gap Table
研究空白表
| Gap | Why it matters | Potential approach |
...
| 研究空白 | 重要性 | 潜在解决思路 |
...
Top Papers to Read First
优先阅读的顶级论文
- [Author et al. year] — [why: foundational/best result/closest to your work]
- ...
undefined- [作者等 年份] — [原因: 基础文献/最优结果/与你的研究最相关]
- ...
undefinedTool Integration (optional)
工具集成(可选)
If Semantic Scholar API available:
bash
curl "https://api.semanticscholar.org/graph/v1/paper/search?query=<topic>&fields=title,year,citationCount,abstract"If HuggingFace Papers available: search with paper topic.
hf_hub_query若Semantic Scholar API可用:
bash
curl "https://api.semanticscholar.org/graph/v1/paper/search?query=<topic>&fields=title,year,citationCount,abstract"若HuggingFace Papers可用: 使用搜索论文主题。
hf_hub_queryLinks to Other Skills
与其他技能的关联
- Feeds into → (use gaps found here to formulate RQ)
research-question - Feeds into → (Related Work section)
paper-writing - Iterates with → (as new baselines discovered)
experiment-tracking
- 为提供输入(利用此处发现的研究空白来构建研究问题)
research-question - 为提供输入(用于撰写相关工作章节)
paper-writing - 与迭代配合(当发现新的基准研究时)
experiment-tracking