Overview
This skill provides a systematic research idea generation process to help researchers produce research ideas with novelty, feasibility, and impact from scratch. The process is divided into three stages:Divergence、Literature search verification、Convergence, ultimately producing a refined 1-page research proposal.
This skill is suitable for:
- Master's and doctoral students looking for thesis topics
- Researchers opening up new research directions
- Brainstorming ideas during interdisciplinary collaboration
- Research teams conducting regular brainstorming
Three-Stage Process
Overview
階段一：發散 階段二：文獻搜索驗證 階段三：收斂
───────────── ───────────────── ──────────
產生 10-20 個 搜索相關文獻 篩選至 1 個
候選 idea 確認新穎性 最佳 idea
建立文獻表
[10種發想框架] → [搜索策略 + 工具] → [評分矩陣 + 篩選標準]
Stage 1: Divergence — Generate 10-20 Candidate Ideas
Objective
Generate as many candidate ideas as possible without preset constraints. The key principle of this stage is **"quantity before quality"**—first seek breadth, then depth.
Preparation
- Define the scope of exploration: Determine your general research direction (e.g., natural language processing, computer vision, reinforcement learning)
- Collect materials: Review recently read papers, attended talks, and observations from experiments
- Set time limits: Spend 15-30 minutes on each framework to avoid getting stuck in details too early
10 Idea Generation Frameworks
The following provides ten idea generation frameworks from different angles. It is recommended to use at least 3-5 of them to generate candidate ideas. For detailed explanations of each framework, see references/brainstorming-frameworks.md。
Framework 1: Problem-First (Start from the Problem)
Start from real problems or pain points and look for research problems that have not yet been properly solved.
- Observe bottlenecks in existing systems
- Collect user complaints and needs
- Examine failure cases on benchmarks
Framework 2: Solution-First (Start from Technology and Find Applications)
Start from a new technology or method and explore scenarios where it can be applied.
- What other problems can a newly proposed algorithm solve?
- Can a breakthrough in one field be transferred to other fields?
Framework 3: Abstraction Ladder
Move between different levels of abstraction, rising from concrete problems to abstract principles and then descending to new concrete applications.
- Abstraction upward: What is the essence of this problem?
- Concretization downward: What can this principle be instantiated as?
Framework 4: Contradiction Hunting
Find contradictions, conflicting assumptions, or unquestioned presuppositions in existing methods.
- Which "common senses" actually lack evidence?
- Are the conclusions of two papers contradictory to each other?
Framework 5: Cross-Domain Transfer
Introduce successful methods or concepts from other fields into your own research field.
- Can mechanisms from biology inspire algorithm design?
- Can theories from the social sciences explain technological phenomena?
Framework 6: What Changed (Focus on Recent Technological Changes)
Focus on recent technological breakthroughs or environmental changes, and think about what new possibilities these changes have opened up.
- New hardware capabilities (larger models, faster inference)
- New data sources (multimodal data, synthetic data)
- New societal needs (privacy, fairness, explainability)
Framework 7: Failure Analysis
Look for opportunities in failed attempts, rejected papers, or underperforming methods.
- Why did this method fail? Can it be fixed?
- Does the reason for the failure point to a more fundamental problem?
Framework 8: Simplicity Test
Challenge complex existing methods with simpler methods, and test whether the complexity is truly necessary.
- Has the simple baseline been compared fairly?
- Which components of the complex method are truly necessary?
Framework 9: Stakeholder Rotation
Rethink the problem from the perspectives of different users and stakeholders.
- What do end users care about most?
- What is the biggest pain point for developers?
- What information do decision-makers need?
Framework 10: Combine / Decompose
Combine or break down existing methods or concepts to generate new research directions.
- Combine: What effect can Method A + Method B produce?
- Decompose: Which submodule of a complex system is worth studying independently?
Output of the Divergence Phase
After completing the divergence phase, you should have a list of 10-20 candidate ideas, each containing:
| Field | Description |
|---|---|
| Number | Sequential number |
| Title | One-sentence description |
| Framework used | Which ideation framework was used |
| Brief description | A preliminary concept in 2-3 sentences |
| Intuitive score | A preliminary intuitive score from 1-5 |
Phase Two: Literature Search Validation — Confirm Novelty
Objective
Through systematic literature search, validate the novelty of the candidate ideas and build a related literature table. The key to this phase isconfirming that your idea has not yet been done, while also finding related work that can be drawn upon.
Search Strategy
For the detailed search strategy, please refer to references/paper-search-strategy.md. The following is a brief process:
Step 1: Keyword Design
- Design 3-5 sets of search keywords for each candidate idea
- Use synonyms, hypernyms, and hyponyms to expand the search scope
- Combine English and Chinese keywords
Step 2: Multi-platform Search
- Semantic Scholar: Semantic search, suitable for exploratory search
- Google Scholar: Broad coverage, suitable for confirmatory search
- arXiv: Latest preprints, suitable for tracking the frontier
- DBLP: Complete index for the computer science field
- ACL Anthology: Dedicated to the NLP field
Step 3: Snowballing
- Forward tracking: after finding a key paper, see who cited it
- Backward tracking: see who the key paper cited
- Author tracking: view other work by key authors
Step 4: Novelty assessment
Perform a novelty assessment for each candidate idea:
| Assessment result | Description | Action |
|---|---|---|
| Brand new | No directly related work found | Keep, but confirm feasibility |
| Partially overlapping | There is related but not identical work | Identify the differences and refine the idea |
| Already done | Found nearly identical work | Eliminate or substantially revise |
Output of literature search verification
After completing this stage, you should have:
- Filtered candidate list: eliminate ideas that have already been done, usually leaving 5-10
- Literature table: each surviving idea corresponds to 5-10 related papers
- Novelty annotation: annotate each idea's novelty assessment result and differences
Stage 3: Convergence — Filter down to 1 best idea
Goal
From the surviving candidate ideas, select 1 best idea for in-depth development through systematic evaluation and filtering.
Convergence process
For the detailed convergence process, see references/idea-convergence.md. The brief process is as follows:
Step 1: Eliminate
Quickly eliminate clearly infeasible ideas:
- Technically impossible to complete within a reasonable time
- Lack of necessary data or computational resources
- Insufficient novelty
Step 2: Merge
Check whether there are candidate ideas that can be merged:
- Are two ideas addressing different aspects of the same problem?
- Can merging them produce a stronger research contribution?
Step 3: Refine
Refine the remaining candidate ideas:
- Narrow the research scope to make the problem more specific
- Clarify the research contribution (Technical Contribution)
- Conceive a preliminary experimental design
Step 4: Scoring matrix
Use a scoring matrix for quantitative evaluation. For the detailed scoring matrix template, see references/evaluation-matrix.md。
Scoring dimensions:
| Dimension | Weight | Description |
|---|---|---|
| Novelty | 25% | Degree of difference from existing work |
| Feasibility | 25% | Likelihood of completion with existing resources |
| Impact | 20% | Potential contribution to the field |
| Timeliness | 15% | Whether it aligns with current research trends |
| Personal interest | 15% | Personal passion and motivation |
Step 5: Final decision
- Choose the idea with the highest weighted total score
- But also consider: does your intuition agree? Do you have enough passion to carry it out?
- If reason and intuition conflict, take time to clarify why
Common pitfalls
During the convergence stage, beware of the following cognitive biases:
- Confirmation bias: searching only for evidence that supports your preference
- Sunk cost: Unwilling to abandon a poor idea because time has already been invested
- Novelty bias: Over-pursuing novelty while ignoring feasibility
- Bandwagon effect: Following a trend just because a direction is popular
- Perfectionism: Waiting for the “perfect” idea and never starting
Final output
After completing the three stages, you should produce the following four documents:
1. Candidate list
A complete list of all candidate ideas, including status markers (alive/eliminated/merged).
2. Literature table
A list of relevant literature for each surviving idea, format:
## Idea: [標題]
1. [作者] (年份). [論文標題]. [會議/期刊].
- 相關性：[高/中/低]
- 與本 idea 的關係：[簡述]
2. ...
3. Scoring matrix
Quantitative scores for all surviving ideas, including scores for each dimension and the weighted total score.
4. 1-Page proposal
A one-page proposal for the final selected idea, including the following sections:
# [研究標題]
## 問題陳述（Problem Statement）
- 要解決什麼問題？為什麼重要？
## 現有方法與不足（Related Work & Gap）
- 現有方法怎麼做？有什麼不足？
## 提議方法（Proposed Approach）
- 你打算怎麼做？核心想法是什麼？
## 預期貢獻（Expected Contributions）
- 預期的研究貢獻是什麼？
## 初步實驗計畫（Preliminary Experiment Plan）
- 打算用什麼資料集？什麼指標？什麼 baseline？
## 風險與備案（Risks & Mitigation）
- 主要風險是什麼？備案是什麼？
## 時間表（Timeline）
- 預計的里程碑與時間安排
Process checklist
Divergence stage
- Set the exploration scope and time limit
- Use at least 3 ideation frameworks
- Generate at least 10 candidate ideas
- Each idea has a title, brief description, and intuitive rating
- No premature filtering at this stage
Literature search validation stage
- Design search keywords for each idea
- Use at least 2 search platforms
- Use snowballing to track key literature
- Mark the novelty assessment for each idea
- Build a literature table
Convergence stage
- Eliminate clearly infeasible ideas
- Check for ideas that can be merged
- Refine the remaining candidates
- Complete the scoring matrix
- Select the final idea
- Write the 1-page proposal
- Check for cognitive biases
Time recommendations
| Stage | Recommended time | Description |
|---|---|---|
| Divergence | 2-4 hours | Can be spread over 2-3 days |
| Literature search validation | 4-8 hours | About 30-60 minutes per idea |
| Convergence | 2-3 hours | Recommended to complete in one sitting |
| Write proposal | 1-2 hours | Do it immediately after convergence |
| Total | 9-17 hours | About 1-2 weeks |
Reference resources
- Detailed explanation of ideation frameworks
- Literature search strategies
- Convergence process and screening criteria
- Scoring matrix template
Advanced recommendations
Build an Idea Bank
- When reading papers normally, jot down the ideas they trigger
- Regularly review the idea bank and combine old ideas with new knowledge
- Use a tagging system to categorize (by field, method, problem type)
Interact with others
- Explain your idea to people in different fields and collect feedback
- Join reading groups or seminars and spark new ideas in discussion
- Find an “idea buddy” to exchange ideas with regularly
Iterative ideation
- Do not expect to find the perfect idea in one go
- Allow yourself to return to the divergent stage and re-brainstorm with new information
- Good ideas are often the result of multiple iterations
Recording and reflection
- Record each brainstorming process, including ideas that were ultimately not selected
- Reflect on which frameworks work best for you
- Build a personal brainstorming process and preference profile