research-innovation-explorer

Research Innovation Explorer

This search-first workflow turns a structured paper pool and an A+B matrix into an evidence-grounded landscape of research questions, uncertainties, and next checks for researcher review.

中文文档

Why This Exists

Most research-idea workflows fail in one of three ways:

• they rely on vague intuition instead of systematic search
• they generate combinations but cannot explain why the combination matters
• they lose the evidence and uncertainty that should guide the next research decision

research-innovation-explorer is built to close those gaps with one coherent workflow:

1. Search broadly and repeatedly.
2. Decompose papers into reusable capabilities.
3. Generate and rank candidate combinations for review.
4. Check both combination directions against source evidence.
5. Return a provisional candidate landscape with supporting evidence, uncertainty, and next checks.

Theory framing, experiment planning, and publication-oriented reporting remain available as explicit follow-up layers for a researcher-selected candidate.

Core Methodology

This skill is built around one explicit research-production loop:

1. Collect roughly 40 relevant, high-quality papers with enough detail to support comparison.
2. Build a pairwise combination matrix over those papers.
3. Keep one row per unique paper pair; 40 papers produce 40 x 39 / 2 = 780 rows with the current generator.
4. Use matrix scores to build a review queue, then examine both A -> B and B -> A through focused source, prior-art, code, or benchmark checks.
5. Keep a provisional set of questions with evidence, contrary evidence, uncertainty, and recommended next actions.

This is the operational core of the workflow, not a side note. The point is not to wait for a single flash of inspiration. The point is to search comprehensively, force structured combination, validate aggressively, and only then keep the few ideas that survive contact with evidence.

What You Get

Workflow

flowchart LR
    A[Search Pass] --> B[Paper Pool]
    B --> C[Capability Decomposition]
    C --> D[Idea Matrix]
    D --> E[Review Queue]
    E --> F[Evidence Review]
    F --> G[Candidate Landscape]
    G -. on request .-> H[Theory Framing]
    G -. on request .-> I[Experiment Plan]
    G -. on request .-> J[Extended Report]

Design Principles

1. Search First

The skill assumes that current literature claims should not come from memory alone when search is available.

2. Dynamic Review

Each review round targets the uncertainty most likely to change the recommendation. Unknown, incomplete, and conflicting evidence remain visible in the output.

3. Evidence-Carrying Reports

The default candidate landscape includes:

• citations
• observed facts and agent inferences kept separate
• candidate comparison
• uncertainty and next checks

Matrix scores are triage signals. They do not establish novelty, feasibility, publishability, or expected research success.

4. Host Neutrality

The workflow is portable across different agent hosts and even manual use. The repo does not depend on one specific runtime.

Quick Start

1. Prepare the search pack

python scripts/build_search_queries.py \
  --topic "long-context reasoning" \
  --keywords "memory routing, verifier head, benchmark"

2. Build the paper pool

Start from:

• assets/templates/search-log.csv
• assets/templates/paper-pool.csv

3. Generate the idea matrix

python scripts/build_idea_matrix.py \
  assets/templates/paper-pool.csv \
  --output work/idea-matrix.csv

4. Review candidates and optionally generate a report

After generating the matrix, copy assets/templates/candidate-review.yaml for candidates in the review queue. Read references/post-matrix-review.md and record source-linked facts, inferences, status, confidence, and next checks.

The report script remains available as a matrix-overview scaffold. Generate static figures and a Markdown overview only when they help the current review:

Generate static figures first when the final research output should include academic paper-style data visuals:

python scripts/build_research_figures.py \
  --paper-pool assets/templates/paper-pool.csv \
  --idea-matrix work/idea-matrix.csv \
  --output-dir work/figures \
  --topic "Long-Context Reasoning" \
  --prefix long_context

python scripts/build_markdown_report.py \
  --topic "Long-Context Reasoning" \
  --paper-pool assets/templates/paper-pool.csv \
  --idea-matrix work/idea-matrix.csv \
  --search-log assets/templates/search-log.csv \
  --figure-dir work/figures \
  --figure-prefix long_context \
  --output work/report.md

Optional Report Style

The reporting layer is intentionally designed for GitHub-native reading:

• Mermaid flowcharts for process explanation
• static PNG heatmaps for matrix snapshots and worked examples
• Mermaid pie charts for quick distribution views
• Markdown evidence tables for claim tracing
• compact narrative sections for executive summary and detailed analysis

This makes an optional overview readable as a working note and a shareable artifact. The default deliverable remains the provisional candidate landscape.

Example Output

Exploring LLM Training Directions

This worked example uses frontier large language model training r