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.

Stage	What to do	What comes out
Paper pool	Gather around 40 relevant papers with reproducible detail	a reusable capability inventory
Combination pass	Enumerate every unique paper pair	780 pair rows for a 40-paper pool
Post-matrix review	Check both directions, inspect evidence, and identify the highest-value unresolved question	candidate review records
Provisional landscape	Group promising, unresolved, parked, weak, and excluded candidates with reasons	a researcher-ready shortlist

What You Get

Layer	What it does
SKILL.md	Defines the default exploration workflow, evidence rules, and optional expansion paths
scripts/build_search_queries.py	Generates structured query packs for topic scan, novelty checks, and failure analysis
scripts/build_idea_matrix.py	Builds a scored pairwise candidate matrix from the paper pool
scripts/build_research_figures.py	Generates publication-style literature heatmaps, scoring heatmaps, and analysis panels from the research artifacts
scripts/build_markdown_report.py	Scaffolds a Markdown matrix overview; reviewed evidence is added afterward
references/	Contains the search playbook, theory framing rules, reporting rules, and ethics boundaries
assets/templates/	Provides CSV, candidate-review, idea-brief, experiment-plan, and report templates

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 research as the target domain. It starts from a search-backed pool of roughly 40 recent papers, builds the combination matrix, and then reviews selected candidates to produce a provisional landscape with evidence and open questions.

At the survey level, the workflow turns the literature into a readable interaction matrix instead of a prose dump:

At the decision level, the workflow uses the matrix as a screening view and records source-based questions, concerns, and next checks separately:

What this example demonstrates:

• search is used during collection and during analysis, not only at the beginning
• the matrix-to-review-queue process is explicit and inspectable
• GitHub README pages and Markdown reports can show the screening logic with direct images, without depending on host-side math rendering

The bundled example images live in assets/examples/llm-training/ and can be regenerated with scripts/build_llm_training_example_figures.py.

Repository Layout

.
├── SKILL.md
├── README.md
├── README.zh-CN.md
├── agents/
│   └── openai.yaml
├── assets/
│   ├── examples/
│   │   └── llm-training/
│   └── templates/
├── references/
└── scripts/
    ├── build_idea_matrix.py
    ├── build_llm_training_example_figures.py
    ├── build_markdown_report.py
    ├── build_research_figures.py
    └── build_search_queries.py

Recommended Use Cases

• discovering literature-grounded research questions worth further review
• mapping literature around a topic before starting implementation
• checking whether an A+B combination already exists in prior work or code
• producing a provisional candidate landscape with citations and visual summaries
• training literature review, abstraction, evaluation design, and research writing habits

Documentation

• Core workflow: SKILL.md
• Search protocol: references/search-playbook.md
• Post-matrix review: references/post-matrix-review.md
• Candidate template: assets/templates/candidate-review.yaml
• Theory framing: references/framing-and-theory.md
• Reporting rules: references/reporting-and-visualization.md
• Report template: assets/templates/analysis-report-template.md

Notes

• If your host cannot render Mermaid, keep the Markdown tables and replace Mermaid blocks with static images or plain-text summaries.
• If your host has no search capability, use the workflow manually and explicitly downgrade confidence in current-literature claims.

Community

For broader discussion around tools, workflows, and AI-native building, visit linux.do.

License

This repository is released under the MIT License.