research-survey

Research Survey

Purpose

Turn a research direction into a rigorous survey organized around concrete problems, challenges, prior solutions, remaining gaps, and possible next research moves. Prioritize a field-level logic chain over a paper-by-paper summary.

Write the final survey in Chinese by default unless the user asks for another language.

Core Principle

Do not start from "what papers exist"; start from "what concrete problem must this specific direction solve." The direction must be narrow enough to make claims precise.

Good scope: physics-informed neural operators for long-horizon spatiotemporal forecasting with limited sensors Bad scope: deep learning, PDE, control

Effort Levels

Choose effort level from the user's wording or explicitly stated preference:

Level	Trigger	Scope	Paper Count	Depth
lite	"简单调研", "quick survey", "大概看看"	Single problem, 1-2 method families	5-10 papers	Challenge map + gap identification
standard	Default, "做个调研", "survey this"	Full problem chain, all major method families	15-30 papers	Full template with method comparison
deep	"深度调研", "exhaustive survey", "写综述"	Comprehensive coverage, historical evolution	30-60+ papers	Full template + timeline + open problems + detailed method cards

Phased Workflow

Phase 1: Scope Definition (必须先完成)

1. Restate the user-provided direction as a narrow research field
2. Specify: task, input/output, target scenario, assumptions, constraints, evaluation goal
3. If too broad, narrow it using research-context skill and state what scope you chose
4. Confirm scope with user before proceeding (unless user said to proceed autonomously)

Phase 2: Literature Discovery

Search Strategy (ordered by reliability):

1. Anchor papers: Find 2-3 seminal/SOTA papers via /arxiv or user input
2. Forward/backward expansion: From anchor papers, trace citations via /semantic-scholar
3. Keyword variants: Search multiple formulations via /openalex and /research-lit
4. Progressive reading: Use /deepxiv for layered access (search→abstract→section)
5. Benchmark/dataset pages: Find standard benchmarks and check leaderboards
6. Code repositories: Check papers with code (paperswithcode.com, GitHub)
7. Recent work: Use /arxiv explicitly for papers from the last 12 months

Literature Search Skills (from ARIS, use as needed):

Skill	When to use
/arxiv	Preprint search, PDF download, latest unrefereed work
/semantic-scholar	Published venue papers (IEEE, ACM, Springer), citation counts
/openalex	Broad open academic metadata, citation graphs
/deepxiv	Progressive section-level reading without loading full papers
/alphaxiv	Quick LLM-optimized summary of a single paper
/research-lit	Multi-source aggregated search (combines multiple sources)
/exa-search	AI web search beyond academic DBs (requires Exa key)

Search execution order for standard effort:

1. /arxiv "topic" → get recent preprints (top 10-15)
2. /semantic-scholar "topic" → get published venue papers (top 10)
3. Pick 2-3 anchors → /deepxiv for progressive reading
4. Check anchors' references → expand coverage
5. /openalex "topic" → fill gaps in coverage

For each paper, record:

• Title, authors, venue, year
• Method family classification
• Core innovation (1 sentence)
• Which challenge it addresses
• Datasets and metrics used
• Code availability
• Key limitation

Anti-Hallucination Rules:

• NEVER fabricate paper titles, authors, venues, or results
• If a paper cannot be verified via search skills, mark it as [未验证]
• Prefer citing papers you can retrieve metadata for (title, venue, year confirmed)
• When citing specific numbers or claims, note the source paper explicitly
• If you cannot find enough papers, say so rather than padding with uncertain references
• Use paper-digest skill only for papers you have actually accessed (PDF, abstract, or full text)

Phase 3: Problem Chain Construction

Before writing any output, build the logic chain internally:

Field → Concrete Problem → Why Hard (Challenges) → 
Prior Solutions (Method Families) → What Each Solves → 
What Remains (Gaps) → Possible Next Work

Verification checklist:

• Every challenge traces back to the concrete problem
• Every method family maps to at least one challenge
• Every gap is grounded in observed limitations or missing evaluations
• No orphan methods (methods that don't connect to any challenge)

Phase 4: Synthesis & Writing

Choose the survey lens:

• Liu Xuefeng method: proposal-oriented, ending in "what we should do"
• Wang teacher method: challenge-centered technical map
• Combined (default): problem → challenges → method-to-challenge mapping → gaps → our direction

Phase 5: Self-Verification

Before delivering the final survey:

• Field is narrow and concrete
• Core problem is specific enough to guide paper selection
• Every challenge corresponds to the core problem
• Every innovation maps to a challenge with mechanism explanation
• Limitations are field-level and evidence-backed
• No fabricated references (all papers verified or marked [未验证])
• Proposed direction directly targets remaining gaps
• Research-context skill consulted for transfer/inspiration sections

Method A: Liu Xuefeng Survey Method

Use when the survey should support research positioning, proposal writing, or "how should we do our work?"

1. 这个具体领域要解决什么问题

   • Define a specific field, not a broad umbrella
   • State the concrete problem: target task, current bottleneck, desired capability
   • Why this matters technically or practically
   • Prior paradigm or assumption being challenged

2. Related work 如何解决这些问题

   • Organize by solution strategy, not by paper
   • For each strategy: innovation, mechanism, what capability it adds
   • Representative papers with evidence
   • Method family comparison table

3. Related work 还有什么不足

   • Field-level limitations, not individual paper weaknesses
   • Be specific: which problem remains, under what setting
   • Separate evidence-proven limitations from inferred concerns
   • Identify contradictions: method A improves X but worsens Y

4. 我们要怎么做来解决这些不足

   • Direction targeting remaining gaps
   • Key idea + expected mechanism + why it could work
   • Minimum viable experiment
   • Risk: what may fail, what baseline would challenge the idea
   • Connect to user's research context

Method B: Wang Teacher Survey Method

Use when the survey should expose the technical skeleton and map innovations to challenges.

1. 这个具体领域要解决什么问题

   • Problem definition with task setting, I/O, evaluation objective
   • Why existing paradigms are insufficient
   • Split into subproblems if needed

2. 解决问题有什么挑战

   • Technical challenges decomposition
   • Why each challenge makes the problem hard
   • Why naive/existing methods struggle
   • Required capability for each challenge

3. 相关工作提出什么创新点解决什么挑战

   • Map each method family to challenges it addresses
   • Innovation as design decision + mechanism
   • Preserve technical details: architecture, loss, training, data construction
   • Evidence: what claim each experiment supports

4. 现在还有什么问题

   • Unresolved challenges after comparing all method families
   • Contradictions between methods
   • Missing evaluations and blind spots
   • Research opportunities from the challenge map

Tool & Skill Integration

• Use paper-digest for deep notes on key papers (seminal, SOTA, complex methods)
• Use research-context for user's research direction (for inspiration/transfer sections)
• Use web search for recent papers, benchmarks, leaderboards, code repos
• Use subagents for parallelizable subtasks when survey is large:
  • Agent 1: collect benchmarks/datasets/metrics landscape
  • Agent 2: read and summarize representative papers
  • Agent 3: search recent arXiv preprints (last 6 months)
• Cite sources when web research was used

Output Template

Use for standard and deep effort levels. Compress for lite level.

# [MMDD]·[Survey] 具体研究方向

> 核心摘要 (One-Liner):
该方向试图解决 [具体领域/任务] 中的 [具体问题]；现有工作主要通过 [主要方法族] 解决 [关键挑战]，但仍存在 [核心不足]。潜在突破口是 [方向]。

| 项目 | 内容 |
| --- | --- |
| **具体方向** | [狭义研究方向] |
| **核心问题** | [该方向最核心、最具体的问题] |
| **关键任务/场景** | [任务、数据、输入输出、约束] |
| **主要挑战** | [Challenge 1; Challenge 2; ...] |
| **代表论文/系统** | [代表作，含年份] |
| **可用代码/基准** | [代码、benchmark、dataset] |
| **调研深度** | lite / standard / deep |
| **论文覆盖** | [X 篇已验证, Y 篇未验证] |

---

## 1. 领域定义与核心问题

- **领域边界：** [为什么这个方向足够具体？]
- **核心问题：** [一句话定义]
- **问题的重要性：** [技术/工程/应用意义]
- **现有范式的不足：** [为什么传统方法不够？]

---

## 2. 关键挑战分解

| 挑战 | 为什么难 | 现有方法为何不足 | 需要的能力 |
| --- | --- | --- | --- |
| Challenge 1 | [原因] | [不足] | [能力] |
| Challenge 2 | [原因] | [不足] | [能力] |

---

## 3. Related Work 如何解决问题

### 3.1 方法族 A：[名称]

- **解决的挑战：** [Challenge X]
- **核心创新：** [设计决策]
- **技术机制：** [模型/算法/训练/数据/公式]
- **代表工作：** [论文 1 (venue'YY), 论文 2 (venue'YY)]
- **证据：** [实验支持了什么]
- **不足：** [该方法族仍未解决什么]

### 3.2 方法族 B：[名称]

[同上结构]

---

## 4. 挑战-创新-证据映射表

| 挑战 | 代表工作 | 创新点 | 解决机制 | 实验证据 | 剩余不足 |
| --- | --- | --- | --- | --- | --- |
| [Challenge] | [Paper] | [Innovation] | [Mechanism] | [Evidence] | [Gap] |

---

## 5. 方法对比分析 (standard/deep only)

| 维度 | 方法族 A | 方法族 B | 方法族 C |
| --- | --- | --- | --- |
| 核心思路 | | | |
| 解决的挑战 | | | |
| 计算成本 | | | |
| 数据需求 | | | |
| 泛化能力 | | | |
| 主要局限 | | | |

---

## 6. 现在还有什么问题

- **Field-level Gap 1：** [跨论文共同不足]
- **Field-level Gap 2：** [跨方法族的未解挑战]
- **Evaluation Gap：** [缺失数据/指标/场景]
- **Theory/Mechanism Gap：** [解释/保证/可控性不足]
- **矛盾与 Trade-off：** [方法 A 改善 X 但恶化 Y]

---

## 7. 我们要怎么做

- **研究切入点：** [从哪个未解问题切入]
- **核心想法：** [设计决策 + 为什么可能有效]
- **对应挑战：** [解决哪个 Challenge]
- **技术路线草案：** [输入→处理→输出→训练→评估]
- **最小可行实验：** [dataset/benchmark/baseline/metric]
- **预期贡献：** [问题/方法/实验/理论贡献]
- **风险与验证：** [可能失败处；什么结果才能证明有效]
- **与我的研究方向的关联：** [参考 research-context]

---

## 8. 论文清单与阅读优先级

| 优先级 | 论文 | Venue/Year | 为什么要读 | 用 paper-digest 深读 | 验证状态 |
| --- | --- | --- | --- | --- | --- |
| P0 | [核心/奠基/SOTA] | [venue'YY] | [原因] | 是/否 | ✓/[未验证] |
| P1 | [补充] | [venue'YY] | [原因] | 是/否 | ✓/[未验证] |

---

## 9. 总结

- **一句话判断：** [该方向是否值得做，为什么]
- **最值得突破的问题：** [具体问题]
- **下一步行动：** [搜索/精读/实验/复现/写proposal]

Quality Checklist

Before finalizing, verify:

• Field is narrow and concrete
• Core problem guides all paper selection
• Every challenge corresponds to the core problem
• Every innovation maps to a challenge with mechanism
• Limitations are field-level and evidence-backed
• No fabricated references
• Proposed direction targets remaining gaps
• Method comparison is fair (same evaluation dimensions)
• Paper list includes venue and year for all entries
• Transfer section references research-context skill