Research Ideation

Research Ideation

From research goal to ranked ideas and a detailed proposal.

Step 0: Load evo-memory (M_I)
    ↓
Step 1: Define Goal
    ↓
Step 2: Literature Grounding (MUST use paper-navigator scripts)
    ↓
Step 3: Generate Ideas (3 directions × 3 personas)
    ↓
Step 4: Refine Ideas (3 tracks × N iterations)
    ↓
Step 5: ELO Tournament → Present Top-3 to User
    ↓
Step 6: Update evo-memory (IDE)
    ↓
User Selects
    ↓
Step 7: Expand into Proposal
    ↓
Step 8: Validate and Iterate

When to Use

• User wants to find a research direction or brainstorm research ideas
• User wants to evaluate whether an idea is novel or worth pursuing
• User wants to rank or compare multiple research ideas
• User wants to generate a research proposal from an idea

When NOT to Use

• Finding/reading papers → use paper-navigator
• Literature survey report → use research-survey
• Planning a paper (story design, experiment plan) → use paper-planning

Step 0: Load Prior Knowledge from evo-memory

Before any ideation begins, load Ideation Memory (M_I) from prior research cycles:

1. Read M_I at /memory/ideation-memory.md (refer to evo-memory skill)
2. Select the top-2 entries (k_I=2) most relevant to the user's current goal by comparing each entry's Summary and Retrieval Tags against the goal
3. Feasible directions from prior cycles → use as seeds in Step 3 (incorporate as candidate research directions alongside new ones; Step 3 still selects exactly 3 directions)
4. Unsuccessful directions marked as fundamental failures → use during idea pruning in Step 4 (prune any idea that matches a fundamental failure pattern)
5. If M_I doesn't exist yet (first cycle), skip this step

This step prevents repeating known dead ends and builds on prior successes across research cycles.

Step 1: Define a Long-Term Research Goal

Start with a goal that has both scientific and practical value. Ambitious enough for multiple papers, concrete enough to guide daily decisions.

Ask: "What is the ultimate form of this research direction? What would the world look like if this problem were fully solved?"

Step 2: Literature Grounding (via paper-navigator)

Invoke paper-navigator (Workflow 9: Ideation Support) to collect 30-50 relevant papers. Do NOT skip this step or substitute with general knowledge — ideas must be grounded in real papers.

CRITICAL: All paper discovery in this step MUST use the paper-navigator skill and its scripts (scholar_search, citation_traverse, arxiv_monitor, recommend, etc.). Using WebSearch, WebFetch, or any generic web search tool for finding papers is PROHIBITED. Generic web search returns blog posts, news articles, and low-quality results — only paper-navigator provides access to Semantic Scholar, arXiv, citation graph traversal, and academic recommendations needed for rigorous literature grounding.

Build Challenge-Insight Tree

From the collected papers, construct a challenge-insight tree — a many-to-many mapping between technical challenges and the insights/techniques that address them:

• Extract challenges: From each paper, what technical problem does it solve?
• Extract insights: What technique or key idea does it use?
• Map connections: Which insights address which challenges?

How this drives ideation:

• Challenges with few insights → unsolved problem (candidate for Step 3)
• Insights not yet applied to a challenge → cross-domain transfer opportunity (candidate for Step 4)
• Challenges with many insights → well-studied, avoid unless you have a fundamentally new angle

Also generate a condensed literature review synthesis as context for idea generation (for full surveys use research-survey).

See references/literature-tree.md for construction methodology.

Execution rule: Do NOT generate ideas without real paper grounding. The tree must reference actual papers with titles, authors, and findings. Paper search MUST go through paper-navigator — never use WebSearch/WebFetch as a shortcut.

Step 3: Generate Ideas

Generate 3 initial research ideas from 3 distinct research directions, grounded in the literature.

Three Personas

Process

1. Analyze literature + challenge-insight tree → identify 3 fundamentally different research directions
2. Generate one idea per direction using Innovator persona
3. Each idea must follow one of two methodological paths:
   • Path 1 (Focused Contribution): Single new component; clean hypothesis
   • Path 2 (System Contribution): Tight causal interaction between components; emergent capability

Idea Format

# Research Idea: [Concise Title]

## Core Idea
[One paragraph: the proposal + which research direction it addresses]

## Validation Plan
[Concrete experiment outline: datasets, baselines, metrics]

Step 4: Refine Ideas

Run 3 parallel refinement tracks — one per initial idea. Each track uses all 3 personas.

For each track:
  For N=3 iterations:
    1. Evaluate current best idea (novelty, feasibility, impact, alignment)
    2. All 3 personas generate refined versions based on evaluation
    3. Pick the best refinement as seed for next iteration
  Track champion = the refinement picked in the final iteration

5 Evolution Strategies

1. Enhancement through Grounding: Strengthen with literature citations
2. Improving Coherence: Fix logical flaws in the mechanism
3. Inspiration and Combination: Combine with a different concept from literature
4. Simplification: Strip down to a clean, testable hypothesis
5. Literature-Driven Pivot: Abandon the mechanism; propose a new approach from literature

Critical rule: If evaluation says the approach is a dead-end, the persona MUST pivot — refinement is not restricted to patching.

Logical Cohesion Principles

• Too many variables → Focus via Subtraction: isolate the most promising variable
• Disconnected components → Justify via Strong Correlation: build explicit causal links

Step 5: ELO Tournament → Present Top-3

Rank the 3 track champions through pairwise comparison, then present all three, ranked, to the user for selection. The tournament orders the champions and records why; it does not eliminate any — each champion is a different research direction, and the user chooses between directions.

Four Dimensions

Tournament

• Starting Elo: 1500 | K-factor: 32
• Round-robin: every champion meets the other two once (3 matches)
• Per match: score both ideas on the four dimensions → higher composite wins → update Elo
• Sort by final Elo. If ratings are equal, or each champion won exactly one match, order by mean composite score instead, then by track order
• See references/elo-ranking-guide.md for rubric and formula

Present Top-3 to User

After the tournament, present the top-3 ideas with both a comparison table and the full refined idea for each. This ensures the user sees the concrete, actionable version of each idea — not just a summary.

Part 1: Comparison Table

## Top-3 Research Ideas (ranked by ELO)

| Rank | Title | Core Mechanism | Novelty | Feasibility | Relevance | Clarity | ELO |
|------|-------|---------------|---------|-------------|-----------|---------|-----|
| 1 | ... | ... | 9 | 7 | 8 | 8 | 1531 |
| 2 | ... | ... | 7 | 9 | 8 | 7 | 1500 |
| 3 | ... | ... | 8 | 6 | 9 | 7 | 1469 |

Each dimension column is the idea's mean score across its two matches.

Part 2: Full Refined Ideas

For each of the top-3, present the refined idea using the same structured format as Step 3, plus a refinement summary:

# Refined Idea [Rank]: [Concise Title]

## Core Idea
[One paragraph: the refined proposal — this should reflect ALL changes from Step 4 refinement,
not the original Step 3 version]

## Validation Plan
[Concrete experiment outline updated with refinement insights: datasets, baselines, metrics,
key ablations identified during refinement]

## Refinement Summary
[Brief paragraph summarizing what changed from the initial idea and why:
- What was simplified or removed (and why)
- What was added or concretized (and why)
- Which persona drove the most impactful change
- Key risk mitigations added during refinement]

This section is mandatory — do NOT skip the full refined ideas or collapse them into the comparison table. The user needs to see the complete, refined version to make an informed selection.