Back to Details

brainstorming-research-ideas

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Research Idea Brainstorming

研究构思头脑风暴

Structured frameworks for discovering the next research idea. This skill provides ten complementary ideation lenses that help researchers move from vague curiosity to concrete, defensible research proposals. Each framework targets a different cognitive mode—use them individually or combine them for comprehensive exploration.
用于发掘下一个研究创意的结构化框架。本Skill提供十个互补的构思视角，帮助研究人员从模糊的好奇心转向具体、站得住脚的研究提案。每个框架针对不同的认知模式——可单独使用，也可组合使用以进行全面探索。

When to Use This Skill

何时使用本Skill

  • Starting a new research direction and need structured exploration
  • Feeling stuck on a current project and want fresh angles
  • Evaluating whether a half-formed idea has real potential
  • Preparing for a brainstorming session with collaborators
  • Transitioning between research areas and seeking high-leverage entry points
  • Reviewing a field and looking for underexplored gaps
Do NOT use this skill when:
  • You already have a well-defined research question and need execution guidance
  • You need help with experimental design or methodology (use domain-specific skills)
  • You want a literature review (use 
    scientific-skills:literature-review
    )
  • 开启新的研究方向，需要结构化探索时
  • 当前项目陷入瓶颈，想要寻找新视角时
  • 评估一个尚未成型的创意是否真正具备潜力时
  • 准备与合作者开展头脑风暴会议时
  • 在不同研究领域间过渡，寻找高影响力切入点时
  • 梳理某一领域，寻找未被充分探索的空白时
请勿在以下场景使用本Skill：
  • 你已经有明确的研究问题，需要执行指导时
  • 需要实验设计或方法论帮助时（使用特定领域的Skill）
  • 需要文献综述时（使用
    scientific-skills:literature-review
    ）

Core Ideation Frameworks

核心构思框架

1. Problem-First vs. Solution-First Thinking

1. 问题优先 vs 解决方案优先思维

Research ideas originate from two distinct modes. Knowing which mode you are in prevents a common failure: building solutions that lack real problems, or chasing problems without feasible approaches.
Problem-First (pain point → method):
  • Start with a concrete failure, bottleneck, or unmet need
  • Naturally yields impactful work because the motivation is intrinsic
  • Risk: may converge on incremental fixes rather than paradigm shifts
Solution-First (new capability → application):
  • Start with a new tool, insight, or technique seeking application
  • Often drives breakthroughs by unlocking previously impossible approaches
  • Risk: "hammer looking for a nail"—solution may lack genuine demand
Workflow:
  1. Write down your idea in one sentence
  2. Classify it: Is this problem-first or solution-first?
  3. If problem-first → verify the problem matters (who suffers? how much?)
  4. If solution-first → identify at least two genuine problems it addresses
  5. For either mode, articulate the gap: what cannot be done today that this enables?
Self-Check:
  • Can I name a specific person or community who needs this?
  • Is the problem I am solving actually unsolved (not just under-marketed)?
  • If solution-first, does the solution create new capability or just replicate existing ones?
研究创意源自两种截然不同的模式。明确自己所处的模式可以避免一个常见误区：构建缺乏真实问题的解决方案，或是追逐没有可行方法的问题。
问题优先（痛点→方法）：
  • 从具体的失败、瓶颈或未被满足的需求入手
  • 自然能产出有影响力的研究，因为动机是内在的
  • 风险：可能局限于增量改进，而非范式转变
解决方案优先（新能力→应用）：
  • 从寻求应用场景的新工具、洞见或技术入手
  • 通常通过解锁此前不可能的方法推动突破
  • 风险：“为锤子找钉子”——解决方案可能缺乏真实需求
工作流程：
  1. 用一句话写下你的创意
  2. 分类：这是问题优先还是解决方案优先？
  3. 如果是问题优先→验证该问题是否重要（谁会受影响？影响程度如何？）
  4. 如果是解决方案优先→确定它能解决的至少两个真实问题
  5. 无论哪种模式，都要明确空白：当前无法实现、而该创意能实现的是什么？
自我检查：
  • 我能说出需要这个创意的具体人群或社区吗？
  • 我解决的问题真的是未被解决的（而非只是宣传不足）吗？
  • 如果是解决方案优先，该方案是创造了新能力还是只是复制现有能力？

2. The Abstraction Ladder

2. 抽象阶梯

Every research problem sits at a particular level of abstraction. Deliberately moving up or down the ladder reveals ideas invisible at your current level.
DirectionActionOutcome
Move Up (generalize)Turn a specific result into a broader principleFramework papers, theoretical contributions
Move Down (instantiate)Test a general paradigm under concrete constraintsEmpirical papers, surprising failure analyses
Move Sideways (analogize)Apply same abstraction level to adjacent domainCross-pollination, transfer papers
Workflow:
  1. State your current research focus in one sentence
  2. Move UP: What is the general principle behind this? What class of problems does this belong to?
  3. Move DOWN: What is the most specific, constrained instance of this? What happens at the extreme?
  4. Move SIDEWAYS: Where else does this pattern appear in a different field?
  5. For each new level, ask: Is this a publishable contribution on its own?
Example:
  • Current: "Improving retrieval accuracy for RAG systems"
  • Up: "What makes context selection effective for any augmented generation system?"
  • Down: "How does retrieval accuracy degrade when documents are adversarially perturbed?"
  • Sideways: "Database query optimization uses similar relevance ranking—what can we borrow?"
每个研究问题都处于特定的抽象层级。刻意上下移动阶梯，会发现当前层级看不到的创意。
方向行动成果
向上移动（概括）将特定结果转化为更广泛的原则框架类论文、理论贡献
向下移动（实例化）在具体约束下测试通用范式实证类论文、意外失败分析
横向移动（类比）将同一抽象层级的模式应用到相邻领域跨领域融合、迁移类论文
工作流程：
  1. 用一句话说明你当前的研究重点
  2. 向上移动：这背后的通用原则是什么？它属于哪类问题？
  3. 向下移动：它最具体、受约束的实例是什么？极端情况下会发生什么？
  4. 横向移动：这种模式还出现在哪些其他领域？
  5. 针对每个新层级，问：这本身是否可作为可发表的研究贡献？
示例：
  • 当前："提升RAG系统的检索准确率"
  • 向上："是什么让上下文选择对任何增强生成系统都有效？"
  • 向下："当文档受到对抗性干扰时，检索准确率会如何下降？"
  • 横向："数据库查询优化使用类似的相关性排序——我们可以借鉴什么？"

3. Tension and Contradiction Hunting

3. 张力与矛盾挖掘

Breakthroughs often come from resolving tensions between widely accepted but seemingly conflicting goals. These contradictions are not bugs—they are the research opportunity.
Common Research Tensions:
Tension PairResearch Opportunity
Performance ↔ EfficiencyCan we match SOTA with 10x less compute?
Privacy ↔ UtilityCan federated/encrypted methods close the accuracy gap?
Generality ↔ SpecializationWhen does fine-tuning beat prompting, and why?
Safety ↔ CapabilityCan alignment improve rather than tax capability?
Interpretability ↔ PerformanceDo mechanistic insights enable better architectures?
Scale ↔ AccessibilityCan small models replicate emergent behaviors?
Workflow:
  1. Pick your research area
  2. List the top 3-5 desiderata (things everyone wants)
  3. Identify pairs that are commonly treated as trade-offs
  4. For each pair, ask: Is this trade-off fundamental or an artifact of current methods?
  5. If artifact → the reconciliation IS your research contribution
  6. If fundamental → characterizing the Pareto frontier is itself valuable
Self-Check:
  • Have I confirmed this tension is real (not just assumed)?
  • Can I point to papers that optimize for each side independently?
  • Is my proposed reconciliation technically plausible, not just aspirational?
突破往往来自解决被广泛接受但看似冲突的目标之间的张力。这些矛盾不是缺陷——它们就是研究机会。
常见研究张力：
张力对研究机会
性能 ↔ 效率我们能否用10倍更少的算力达到SOTA水平？
隐私 ↔ 实用性联邦/加密方法能否缩小准确率差距？
通用性 ↔ 专业化微调何时优于提示，原因是什么？
安全性 ↔ 能力对齐能否提升而非削弱能力？
可解释性 ↔ 性能机制洞见能否催生更好的架构？
规模 ↔ 可及性小型模型能否复制涌现行为？
工作流程：
  1. 选择你的研究领域
  2. 列出3-5个最核心的需求（所有人都想要的东西）
  3. 找出通常被视为权衡的需求对
  4. 针对每一对，问：这种权衡是根本性的，还是当前方法的产物？
  5. 如果是当前方法的产物→调和这种矛盾就是你的研究贡献
  6. 如果是根本性的→刻画帕累托边界本身就有价值
自我检查：
  • 我是否确认这种张力是真实存在的（而非只是假设）？
  • 我能否指出分别针对每一侧优化的论文？
  • 我提出的调和方案在技术上是否可行，而非只是空想？

4. Cross-Pollination (Analogy Transfer)

4. 跨领域融合（类比迁移）

Borrowing structural ideas from other disciplines is one of the most generative research heuristics. Many foundational techniques emerged this way—attention mechanisms draw from cognitive science, genetic algorithms from biology, adversarial training from game theory.
Requirements for a Valid Analogy:
  • Structural fidelity: The mapping must hold at the level of underlying mechanisms, not just surface similarity
  • Non-obvious connection: If the link is well-known, the novelty is gone
  • Testable predictions: The analogy should generate concrete hypotheses
High-Yield Source Fields for ML Research:
Source FieldTransferable Concepts
NeuroscienceAttention, memory consolidation, hierarchical processing
PhysicsEnergy-based models, phase transitions, renormalization
EconomicsMechanism design, auction theory, incentive alignment
EcologyPopulation dynamics, niche competition, co-evolution
LinguisticsCompositionality, pragmatics, grammatical induction
Control TheoryFeedback loops, stability, adaptive regulation
Workflow:
  1. Describe your problem in domain-agnostic language (strip the jargon)
  2. Ask: What other field solves a structurally similar problem?
  3. Study that field's solution at the mechanism level
  4. Map the solution back to your domain, preserving structural relationships
  5. Generate testable predictions from the analogy
  6. Validate: Does the borrowed idea actually improve outcomes?
从其他学科借鉴结构性创意是最具生成性的研究启发法之一。许多基础技术都是这样诞生的——注意力机制源自认知科学，遗传算法源自生物学，对抗训练源自博弈论。
有效类比的要求：
  • 结构保真：映射必须在底层机制层面成立，而非只是表面相似
  • 非显而易见的关联：如果关联已广为人知，就没有新颖性了
  • 可测试的预测：类比应能生成具体的假设
机器学习研究的高价值源领域：
源领域可迁移概念
神经科学注意力、记忆巩固、分层处理
物理学基于能量的模型、相变、重正化
经济学机制设计、拍卖理论、激励对齐
生态学种群动态、生态位竞争、共同进化
语言学组合性、语用学、语法归纳
控制理论反馈回路、稳定性、自适应调节
工作流程：
  1. 用领域无关的语言描述你的问题（去掉行话）
  2. 问：还有哪个领域解决过结构相似的问题？
  3. 从机制层面研究该领域的解决方案
  4. 将解决方案映射回你的领域，保留结构关系
  5. 从类比中生成可测试的预测
  6. 验证：借鉴的创意是否真的能改善结果？

5. The "What Changed?" Principle

5. “什么发生了变化？”原则

Strong ideas often come from revisiting old problems under new conditions. Advances in hardware, scale, data availability, or regulations can invalidate prior assumptions and make previously impractical approaches viable.
Categories of Change to Monitor:
Change TypeExampleResearch Implication
ComputeGPUs 10x fasterMethods dismissed as too expensive become feasible
ScaleTrillion-token datasetsStatistical arguments that failed at small scale may now hold
RegulationEU AI Act, GDPRCreates demand for compliant alternatives
ToolingNew frameworks, APIsReduces implementation barrier for complex methods
FailureHigh-profile system failuresExposes gaps in existing approaches
CulturalNew user behaviorsShifts what problems matter most
Workflow:
  1. Pick a well-known negative result or abandoned approach (3-10 years old)
  2. List the assumptions that led to its rejection
  3. For each assumption, ask: Is this still true today?
  4. If any assumption has been invalidated → re-run the idea under new conditions
  5. Frame the contribution: "X was previously impractical because Y, but Z has changed"
优秀的创意往往来自在新条件下重新审视旧问题。硬件、规模、数据可用性或法规的进步可能会使先前的假设失效，让以前不切实际的方法变得可行。
需要关注的变化类别：
变化类型示例研究启示
算力GPUs提速10倍曾因成本过高被摒弃的方法现在变得可行
规模万亿级token数据集在小规模下不成立的统计论证现在可能成立
法规EU AI Act、GDPR催生对合规替代方案的需求
工具新框架、API降低复杂方法的实现门槛
失败案例高知名度系统故障暴露现有方法的空白
文化用户新行为改变最受关注的问题类型
工作流程：
  1. 选择一个知名的负面结果或被放弃的方法（3-10年前的）
  2. 列出导致其被否决的假设
  3. 针对每个假设，问：这在今天仍然成立吗？
  4. 如果任何假设已失效→在新条件下重新验证该创意
  5. 阐述贡献：“X此前因Y不切实际，但Z已经发生了变化”

6. Failure Analysis and Boundary Probing

6. 失败分析与边界探测

Understanding where a method breaks is often as valuable as showing where it works. Boundary probing systematically exposes the conditions under which accepted techniques fail.
Types of Boundaries to Probe:
  • Distributional: What happens with out-of-distribution inputs?
  • Scale: Does the method degrade at 10x or 0.1x the typical scale?
  • Adversarial: Can the method be deliberately broken?
  • Compositional: Does performance hold when combining multiple capabilities?
  • Temporal: Does the method degrade over time (concept drift)?
Workflow:
  1. Select a widely-used method with strong reported results
  2. Identify the implicit assumptions in its evaluation (dataset, scale, domain)
  3. Systematically violate each assumption
  4. Document where and how the method breaks
  5. Diagnose the root cause of each failure
  6. Propose a fix or explain why the failure is fundamental
Self-Check:
  • Am I probing genuine boundaries, not just confirming known limitations?
  • Can I explain WHY the method fails, not just THAT it fails?
  • Does my analysis suggest a constructive path forward?
了解方法失效的场景往往和展示其有效的场景同样有价值。边界探测系统地揭示了公认技术失效的条件。
需要探测的边界类型：
  • 分布偏移：输入为分布外数据时会发生什么？
  • 规模变化：方法在典型规模的10倍或0.1倍时性能会下降吗？
  • 对抗性干扰：能否刻意破坏该方法？
  • 组合性：组合多种能力时性能是否保持稳定？
  • 时间漂移：方法性能会随时间下降吗（概念漂移）？
工作流程：
  1. 选择一个被广泛使用且报告结果优异的方法
  2. 明确其评估中的隐含假设（数据集、规模、领域）
  3. 系统性地违反每个假设
  4. 记录方法失效的场景和方式
  5. 诊断每次失效的根本原因
  6. 提出修复方案或解释为何失效是根本性的
自我检查：
  • 我是否在探测真正的边界，而非只是确认已知的局限性？
  • 我能否解释方法失效的原因，而非只是指出它失效了？
  • 我的分析是否提出了建设性的改进方向？

7. The Simplicity Test

7. 简洁性测试

Before accepting complexity, ask whether a simpler approach suffices. Fields sometimes over-index on elaborate solutions when a streamlined baseline performs competitively.
Warning Signs of Unnecessary Complexity:
  • The method has many hyperparameters with narrow optimal ranges
  • Ablations show most components contribute marginally
  • A simple baseline was never properly tuned or evaluated
  • The improvement over baselines is within noise on most benchmarks
Workflow:
  1. Identify the current SOTA method for your problem
  2. Strip it to its simplest possible core (what is the one key idea?)
  3. Build that minimal version with careful engineering
  4. Compare fairly: same compute budget, same tuning effort
  5. If the gap is small → the contribution is the simplicity itself
  6. If the gap is large → you now understand what the complexity buys
Contribution Framing:
  • "We show that [simple method] with [one modification] matches [complex SOTA]"
  • "We identify [specific component] as the critical driver, not [other components]"
在接受复杂性之前，先问问更简单的方法是否足够。有时领域会过度关注复杂解决方案，而精简的基线方法就能达到竞争力。
不必要复杂性的警示信号：
  • 方法有许多超参数，且最优范围狭窄
  • 消融实验显示大多数组件的贡献微乎其微
  • 简单基线从未被正确调优或评估
  • 在大多数基准测试中，相对于基线的提升处于误差范围内
工作流程：
  1. 确定你的问题当前的SOTA方法
  2. 将其精简到最简单的核心（核心创意是什么？）
  3. 通过严谨的工程实现这个最小版本
  4. 公平比较：相同算力预算、相同调优投入
  5. 如果差距很小→贡献本身就是简洁性
  6. 如果差距很大→你现在理解了复杂性的价值
贡献阐述：
  • “我们展示了[简单方法]通过[一项修改]就能匹配[复杂SOTA]”
  • “我们发现[特定组件]是关键驱动因素，而非[其他组件]”

8. Stakeholder Rotation

8. 利益相关者视角切换

Viewing a system from multiple perspectives reveals distinct classes of research questions. Each stakeholder sees different friction, risk, and opportunity.
Stakeholder Perspectives:
StakeholderKey Questions
End UserIs this usable? What errors are unacceptable? What is the latency tolerance?
DeveloperIs this debuggable? What is the maintenance burden? How does it compose?
TheoristWhy does this work? What are the formal guarantees? Where are the gaps?
AdversaryHow can this be exploited? What are the attack surfaces?
EthicistWho is harmed? What biases are embedded? Who is excluded?
RegulatorIs this auditable? Can decisions be explained? Is there accountability?
OperatorWhat is the cost? How does it scale? What is the failure mode?
Workflow:
  1. Describe your system or method in one paragraph
  2. Assume each stakeholder perspective in turn (spend 5 minutes per role)
  3. For each perspective, list the top 3 concerns or questions
  4. Identify which concerns are unaddressed by existing work
  5. The unaddressed concern with the broadest impact is your research question
从多个视角审视系统会发现不同类别的研究问题。每个利益相关者看到的摩擦、风险和机会都不同。
利益相关者视角：
利益相关者核心问题
终端用户它好用吗？哪些错误是不可接受的？延迟容忍度是多少？
开发者它可调试吗？维护负担有多大？它的组合性如何？
理论家它为什么有效？有哪些形式化的保障？空白在哪里？
攻击者如何利用它？攻击面有哪些？
伦理学家谁会受到伤害？嵌入了哪些偏见？谁被排除在外？
监管者它可审计吗？决策可解释吗？是否有问责机制？
运营者成本是多少？可扩展性如何？失效模式是什么？
工作流程：
  1. 用一段话描述你的系统或方法
  2. 依次代入每个利益相关者的视角（每个角色花5分钟）
  3. 针对每个视角，列出最核心的3个关注点或问题
  4. 找出现有研究未解决的关注点
  5. 影响最广泛的未解决关注点就是你的研究问题

9. Composition and Decomposition

9. 组合与分解

Novelty often emerges from recombination or modularization. Innovation frequently lies not in new primitives, but in how components are arranged or separated.
Composition (combining existing techniques):
  • Identify two methods that solve complementary subproblems
  • Ask: What emergent capability arises from combining them?
  • Example: RAG + Chain-of-Thought → retrieval-augmented reasoning
Decomposition (breaking apart monolithic systems):
  • Identify a complex system with entangled components
  • Ask: Which component is the actual bottleneck?
  • Example: Decomposing "fine-tuning" into data selection, optimization, and regularization reveals that data selection often matters most
Workflow:
  1. List the 5-10 key components or techniques in your area
  2. Compose: Pick pairs and ask what happens when you combine them
  3. Decompose: Pick a complex method and isolate each component's contribution
  4. For compositions: Does the combination create emergent capabilities?
  5. For decompositions: Does isolation reveal a dominant or redundant component?
新颖性往往来自重组或模块化。创新通常不在于新的原语，而在于组件的排列或拆分方式。
组合（结合现有技术）：
  • 找出解决互补子问题的两种方法
  • 问：将它们结合会产生什么涌现能力？
  • 示例：RAG + 思维链 → 检索增强推理
分解（拆分单体系统）：
  • 找出组件相互纠缠的复杂系统
  • 问：哪个组件是真正的瓶颈？
  • 示例：将“微调”分解为数据选择、优化和正则化，发现数据选择往往是最重要的
工作流程：
  1. 列出你所在领域的5-10个核心组件或技术
  2. 组合：挑选成对技术，问结合后会发生什么
  3. 分解：挑选一个复杂方法，分离每个组件的贡献
  4. 对于组合：结合是否产生了涌现能力？
  5. 对于分解：分离是否揭示了主导或冗余组件？

10. The "Explain It to Someone" Test

10. “向他人解释”测试

A strong research idea should be defensible in two sentences to a smart non-expert. This test enforces clarity of purpose and sharpens the value proposition.
The Two-Sentence Template:
Sentence 1 (Problem): "[Domain] currently struggles with [specific problem], which matters because [concrete consequence]." Sentence 2 (Insight): "We [approach] by [key mechanism], which works because [reason]."
If You Cannot Fill This Template:
  • The problem may not be well-defined yet → return to Framework 1
  • The insight may not be clear yet → return to Framework 7 (simplify)
  • The significance may not be established → return to Framework 3 (find the tension)
Calibration Questions:
  • Would a smart colleague outside your subfield understand why this matters?
  • Does the explanation stand without jargon?
  • Can you predict what a skeptic's first objection would be?
优秀的研究创意应该能用两句话向聪明的非专业人士解释清楚。这个测试能确保目标清晰，强化价值主张。
两句话模板：
第一句（问题）：“[领域]当前面临[具体问题]，这很重要，因为[具体后果]。” 第二句（洞见）：“我们通过[关键机制]采用[方法]，之所以有效是因为[原因]。”
如果无法填写该模板：
  • 问题可能尚未明确定义→回到框架1
  • 洞见可能不清晰→回到框架7（简洁性测试）
  • 重要性可能未确立→回到框架3（寻找张力）
校准问题：
  • 你所在子领域之外的聪明同事能理解它的重要性吗？
  • 不用行话也能解释清楚吗？
  • 你能预测怀疑者的第一个反对意见是什么吗？

Integrated Brainstorming Workflow

整合式头脑风暴工作流程

Use this end-to-end workflow to go from blank page to ranked research ideas.
使用这个端到端的工作流程，从空白页面到筛选出优质研究创意。

Phase 1: Diverge (Generate Candidates)

阶段1：发散（生成候选创意）

Goal: Produce 10-20 candidate ideas without filtering.
  1. Scan for tensions (Framework 3): List 5 trade-offs in your field
  2. Check what changed (Framework 5): List 3 recent shifts (compute, data, regulation)
  3. Probe boundaries (Framework 6): Pick 2 popular methods and find where they break
  4. Cross-pollinate (Framework 4): Pick 1 idea from an adjacent field
  5. Compose/decompose (Framework 9): Combine 2 existing techniques or split 1 apart
  6. Climb the abstraction ladder (Framework 2): For each candidate, generate up/down/sideways variants
目标：生成10-20个候选创意，不做筛选。
  1. 扫描张力（框架3）：列出你所在领域的5个权衡
  2. 检查变化（框架5）：列出3个近期变化（算力、数据、法规）
  3. 探测边界（框架6）：挑选2个流行方法，找出它们失效的场景
  4. 跨领域融合（框架4）：从相邻领域挑选1个创意
  5. 组合/分解（框架9）：结合2种现有技术或拆分1种技术
  6. 攀登抽象阶梯（框架2）：针对每个候选创意，生成向上/向下/横向变体

Phase 2: Converge (Filter and Rank)

阶段2：收敛（筛选与排序）

Goal: Narrow to 3-5 strongest ideas.
Apply these filters to each candidate:
FilterQuestionKill Criterion
Explain-It Test (F10)Can I state this in two sentences?If no → idea is not yet clear
Problem-First Check (F1)Is the problem genuine and important?If no one suffers from this → drop it
Simplicity Test (F7)Is the complexity justified?If a simpler approach works → simplify or drop
Stakeholder Check (F8)Who benefits? Who might object?If no clear beneficiary → drop it
FeasibilityCan I execute this with available resources?If clearly infeasible → park it for later
目标：缩小到3-5个最优质的创意。
对每个候选创意应用以下筛选条件：
筛选条件问题淘汰标准
解释测试（框架10）我能用两句话说明它吗？不能→创意尚未清晰
问题优先检查（框架1）问题是否真实且重要？没有人受此影响→淘汰
简洁性测试（框架7）复杂性是否合理？更简单的方法有效→简化或淘汰
利益相关者检查（框架8）谁受益？谁可能反对？没有明确受益者→淘汰
可行性我能用现有资源执行吗？明显不可行→暂时搁置

Phase 3: Refine (Sharpen the Winner)

阶段3：优化（打磨最优创意）

Goal: Turn the top idea into a concrete research plan.
  1. Write the two-sentence pitch (Framework 10)
  2. Identify the core tension being resolved (Framework 3)
  3. Specify the abstraction level (Framework 2)
  4. List 3 concrete experiments that would validate the idea
  5. Anticipate the strongest objection and prepare a response
  6. Define a 2-week pilot that would provide signal on feasibility
Completion Checklist:
  • Two-sentence pitch is clear and compelling
  • Problem is genuine (problem-first check passed)
  • Approach is justified (simplicity test passed)
  • At least one stakeholder clearly benefits
  • Core experiments are specified
  • Feasibility pilot is defined
  • Strongest objection has a response
目标：将最优创意转化为具体的研究计划。
  1. 撰写两句话的提案（框架10）
  2. 确定要解决的核心张力（框架3）
  3. 明确抽象层级（框架2）
  4. 列出3个能验证创意的具体实验
  5. 预判最强烈的反对意见并准备回应
  6. 定义一个2周的试点项目，验证可行性
完成 checklist：
  • 两句话提案清晰且有吸引力
  • 问题真实（通过问题优先检查）
  • 方法合理（通过简洁性测试）
  • 至少有一个利益相关者明确受益
  • 核心实验已明确
  • 可行性试点已定义
  • 已准备好回应最强烈的反对意见

Framework Selection Guide

框架选择指南

Not sure which framework to start with? Use this decision guide:
Your SituationStart With
"I don't know what area to work in"Tension Hunting (F3) → What Changed (F5)
"I have a vague area but no specific idea"Abstraction Ladder (F2) → Failure Analysis (F6)
"I have an idea but I'm not sure it's good"Explain-It Test (F10) → Simplicity Test (F7)
"I have a good idea but need a fresh angle"Cross-Pollination (F4) → Stakeholder Rotation (F8)
"I want to combine existing work into something new"Composition/Decomposition (F9)
"I found a cool technique and want to apply it"Problem-First Check (F1) → Stakeholder Rotation (F8)
"I want to challenge conventional wisdom"Failure Analysis (F6) → Simplicity Test (F7)
不确定从哪个框架开始？使用这个决策指南：
你的情况从这里开始
“我不知道该研究哪个领域”张力挖掘（框架3）→ 什么发生了变化（框架5）
“我有模糊的方向但没有具体创意”抽象阶梯（框架2）→ 失败分析（框架6）
“我有一个创意但不确定它好不好”解释测试（框架10）→ 简洁性测试（框架7）
“我有一个好创意但需要新视角”跨领域融合（框架4）→ 利益相关者视角切换（框架8）
“我想结合现有工作创造新东西”组合/分解（框架9）
“我发现了一个很酷的技术想应用它”问题优先检查（框架1）→ 利益相关者视角切换（框架8）
“我想挑战传统认知”失败分析（框架6）→ 简洁性测试（框架7）

Common Pitfalls in Research Ideation

研究构思的常见陷阱

PitfallSymptomFix
Novelty without impact"No one has done X" but no one needs XApply Problem-First Check (F1)
Incremental by defaultIdea is +2% on a benchmarkClimb the Abstraction Ladder (F2)
Complexity worshipMethod has 8 components, each helping marginallyApply Simplicity Test (F7)
Echo chamberAll ideas come from reading the same 10 papersUse Cross-Pollination (F4)
Stale assumptions"This was tried and didn't work" (5 years ago)Apply What Changed (F5)
Single-perspective biasOnly considering the ML engineer's viewUse Stakeholder Rotation (F8)
Premature convergenceCommitted to first idea without exploring alternativesRun full Diverge phase
陷阱症状解决方法
有新颖性但无影响力“没人做过X”但也没人需要X应用问题优先检查（框架1）
默认增量改进创意在基准测试上提升2%攀登抽象阶梯（框架2）
崇拜复杂性方法有8个组件，每个贡献都很小应用简洁性测试（框架7）
回音室效应所有创意都来自阅读相同的10篇论文使用跨领域融合（框架4）
陈旧假设“这个试过了，没用”（5年前）应用“什么发生了变化”（框架5）
单一视角偏见只考虑ML工程师的视角使用利益相关者视角切换（框架8）
过早收敛执着于第一个创意，未探索其他选项完成完整的发散阶段

Usage Instructions for Agents

Agent使用说明

When a researcher asks for help brainstorming research ideas:
  1. Identify their starting point: Are they exploring a new area, stuck on a current project, or evaluating an existing idea?
  2. Select appropriate frameworks: Use the Framework Selection Guide to pick 2-3 relevant lenses
  3. Walk through frameworks interactively: Apply each framework step-by-step, asking the researcher for domain-specific inputs
  4. Generate candidates: Aim for 10-20 raw ideas across frameworks
  5. Filter and rank: Apply the Converge phase filters to narrow to top 3-5
  6. Refine the winner: Help articulate the two-sentence pitch and define concrete next steps
Key Principles:
  • Push for specificity—vague ideas ("improve efficiency") are not actionable
  • Challenge assumptions—ask "why?" at least three times
  • Maintain a written list of all candidates, even rejected ones (they may recombine later)
  • The researcher makes the final call on which ideas to pursue; the agent facilitates structured thinking
当研究人员请求帮助构思研究创意时：
  1. 确定他们的起点：他们是在探索新领域、当前项目陷入瓶颈，还是在评估现有创意？
  2. 选择合适的框架：使用框架选择指南挑选2-3个相关视角
  3. 交互式引导框架应用：逐步应用每个框架，向研究人员询问领域特定的输入
  4. 生成候选创意：目标是通过多个框架生成10-20个原始创意
  5. 筛选与排序：应用收敛阶段的筛选条件，缩小到前3-5个创意
  6. 优化最优创意：帮助撰写两句话的提案并定义具体的下一步
核心原则：
  • 推动具体化——模糊的创意（如“提升效率”）无法落地
  • 挑战假设——至少问三次“为什么？”
  • 保留所有候选创意的书面记录，即使是被淘汰的（它们以后可能会被重组）
  • 研究人员最终决定追求哪些创意；Agent仅负责引导结构化思考