The world's first literature-driven AI research assistant. Methods come from real papers, not hardcoded templates.
- Why ResearchX?
- Architecture Overview
- Core Innovation: Literature-Driven Methods
- Modules Deep Dive
- Workflow Examples
- Quick Start
- Installation
- Output Artifacts
- FAQ
- Contributing
| Tool | Limitation |
|---|---|
| ChatGPT / GPT-4 | Uses pre-training cutoff knowledge, can't find latest methods |
| Paper chatbots | Only summarize, don't help you design new research |
| Research Copilots | Hardcode methods ("use Transformer"), not grounded in current literature |
| Literature tools | Find papers but can't synthesize into actionable research plans |
Other tools: "For image segmentation, use U-Net or Transformer."
↑ Pre-defined, may be outdated
ResearchX: "For crop mapping from satellite imagery (2025),
the SOTA methods found via literature search are:
- SwinUNet (94.2% OA, 2024)
- SSTFormer (95.1% F1, 2025)
- GeoFM (emerging foundation model, 2025)
The gap is: none of these handle cloud-heavy regions well.
Innovation opportunity: [cross-domain idea]"
↑ Grounded in real 2023-2026 literature
- PhD students & researchers — From literature review to paper submission
- Masters students — Thesis topic discovery and experimental design
- Professors & PIs — Grant proposals and research planning
- Industry R&D — Method selection and innovation discovery
graph TB
subgraph "User Input Layer"
U[User Request]
PDF[PDF Upload]
DOI[DOI / Paper Link]
end
subgraph "ResearchX Core"
R[Task Router]
R --> P[Paper Analysis]
R --> G[Gap Mining]
R --> M[Method Mining]
R --> E[Experiment Design]
R --> W[Manuscript Writer]
R --> S[SCI Upgrade]
R --> RV[Peer Review]
R --> V[Visual Generation]
R --> GR[Grant Proposal]
R --> LR[Lit Review Builder]
end
subgraph "Knowledge Sources"
WS[web_search ~ 2023-2026]
LS[Local Scripts]
end
U --> R
PDF --> P
DOI --> P
P --> WS; G --> WS; M --> WS; W --> WS; GR --> WS; LR --> WS
M --> LS; V --> LS
subgraph "Output"
O1[analysis_*.md] --> R
O2[topics_*.md] --> R
O3[methods_*.md] --> R
O4[manuscript_*.md] --> R
O5[review_report.md] --> R
O6[grant_proposal.md] --> R
end
style WS fill:#e1f5fe,stroke:#0288d1
style R fill:#f3e5f5,stroke:#7b1fa2
flowchart LR
subgraph "End-to-End: Vague Idea → Published Paper"
A1[Clarify Topic] --> A2[Gap Mining]
A2 --> A3[Method Mining]
A3 --> A4[Experiment Design]
A4 --> A5[Manuscript Writing]
A5 --> A6[SCI Upgrade]
A6 --> A7[Review Simulation]
A7 --> A8[Visual Generation]
end
subgraph "Paper Analysis → Improvement"
B1[Paper Analysis] --> B2[Method Comparison]
B2 --> B3[Missing Experiments]
B3 --> B4[Upgrade Plan]
end
A1 -.->|or| B1
sequenceDiagram
participant U as User
participant RX as ResearchX
participant WS as Web Search
participant LS as Local Scripts
U->>RX: "I want to study [topic]"
RX->>WS: Search reviews + surveys
RX->>WS: Search challenges + gaps
RX->>WS: Search SOTA methods
WS-->>RX: Paper data
RX->>LS: analyze_methods.py
LS-->>RX: Method taxonomy + trends
RX->>RX: Identify research gaps
RX->>U: 5 topics with literature-backed methods
U->>RX: "I like topic 3, write it up"
RX->>WS: Targeted method search
RX->>LS: Generate method landscape
RX->>U: Manuscript + citation list
RX->>U: "Shall I continue to SCI upgrade?"
U->>RX: "Yes"
RX->>U: Upgrade plan + peer review
This is the single most important design decision in ResearchX.
User says: "I want to study crop mapping using deep learning"
STEP 1 — web_search: "crop mapping deep learning review 2024 2025"
→ Finds: 2024 survey in RSE, 2025 benchmark paper
STEP 2 — web_search: "crop mapping challenge limitation"
→ Finds: Cloud cover problem, limited generalization, label scarcity
STEP 3 — web_search: "crop mapping state-of-the-art 2025"
→ Finds: SSTFormer, CropFormer, GeoFM foundation model
STEP 4 — Method landscape built:
| Method Family | Best Performer | Year | Limitation |
|-------------------|----------------|------|----------------------|
| CNN-based | DeepCropNet | 2023 | Limited context |
| Transformer-based | SSTFormer | 2025 | Heavy computation |
| Foundation Models | GeoFM | 2025 | Data-hungry |
STEP 5 — Gap + Innovation identified:
"None of these methods handle heavy cloud cover.
Innovation: Combine SSTFormer's accuracy with a cloud-imputation module from meteorology literature."
↑ Cross-domain transfer from meteorology
- Always current — Not limited to pre-training data
- Always grounded — Every claim cites a real paper
- Cross-domain — Can borrow methods from any adjacent field
- Falsifiable — User can check the cited papers
Input: PDF / DOI / Paper title
Output: analysis_[Paper]_YYYYMMDD.md
- One-sentence summary
- Research logic chain (RQ → Data → Method → Results)
- Academic thought analysis (Why done? Why this approach?)
- Strengths & Weaknesses table
- Missing experiments detected
- Literature context (latest SOTA comparison)
- Improvement suggestionsInput: Research domain or topic
Output: topics_[Topic]_YYYYMMDD.md
Produces a gap taxonomy (Solved / Unsolved / Controversial / Unexplored) and generates 5 concrete SCI topic proposals with:
- Specific research question
- Innovation statement (citing concrete papers)
- Data requirements
- Literature-backed method recommendation
- Target journal & feasibility score
Input: Research topic
Output: methods_[Topic]_YYYYMMDD.md
| Component | Description |
|---|---|
| Method Families | Extracted from literature, not assumed |
| Evolution Timeline | 2022 → 2023 → 2024 → 2025-2026 |
| Method Landscape Table | Performance, pros, cons (from papers) |
| Innovation Opportunity | What's missing? What can be transferred? |
Input: Proposed method
Output: Full experimental protocol
Designs: Baseline comparison (3-5 SOTA), Ablation study, Parameter sensitivity, Generalization tests (×3), Robustness tests, Statistical significance.
Then detects: "What experiments are missing that reviewers would ask for?"
Input: Research plan
Output: manuscript_[Topic]_YYYYMMDD.md
Generates any section on demand with journal-adapted style:
| Section | Template |
|---|---|
| Abstract | Background(1s) → Problem(1s) → Method(2-3s) → Results(2s) → Implication(1s) |
| Introduction | Broad → Specific → Gap → Our work → Contributions |
| Methods | Data → Model → Implementation → Evaluation |
| Results | Overall → Ablation → Parameter → Generalization |
| Discussion | Findings → Comparison → Limitations → Future |
Input: Draft paper
Output: Upgrade action plan
Scores 4 dimensions (Novelty, Experiments, Writing, Related Work) and produces a ranked priority action list with specific literature citations. Recommends target journals with predicted acceptance rates.
Input: Paper draft
Output: review_report.md + response letter
Generates 3-5 major comments + 2-3 minor comments from a selected reviewer persona. Includes a revision strategy and a point-by-point response letter.
Input: Research topic
Output: Prompts for GPT Image / Midjourney / Flux
Generates:
- Graphical abstract prompts (3 platforms, 3 journal styles)
- Research poster layouts (A0, A1, 9:16, conference)
- Mermaid workflow diagram code
Input: Research topic
Output: literature_matrix_[Topic]_YYYYMMDD.md
Builds a structured literature matrix table → generates a coherent "Related Work" section with thematic grouping and critical comparison.
Input: Research topic
Output: grant_proposal_[Topic]_YYYYMMDD.md
Full proposal structure (NSFC-compatible and international) with literature-backed claims, scientific questions, technical route, innovation points, and expected outcomes.
User: "I'm a PhD student in remote sensing. I want to study crop type mapping
using deep learning, but I need a specific topic and method."
ResearchX auto-chains: Gap Mining → Method Mining → Experiment Design
1. Clarify: "What region? What crops? What data do you have access to?"
2. Search: "crop type mapping deep learning review 2024 2025"
"crop mapping challenge limitation"
"crop mapping state-of-the-art 2025"
3. Gap analysis: identifies "few-shot learning for rare crops" as an open problem
4. Propose 5 topics with literature-backed methods
5. User picks one: "Few-shot crop mapping using foundation models"
6. Search methods specifically for this topic
7. Design full experimental protocol
8. Save: topics_crop_mapping_20250610.md, methods_few_shot_crop_20250610.md
User: "I have a draft paper on urban land use classification using ResNet.
I want to publish in RSE (Q1). Help me upgrade it."
ResearchX auto-chains: Paper Analysis → Method Mining → SCI Upgrade
1. Analyze the paper, extract its logic chain
2. Search: "urban land use classification 2024 2025"
3. Compare: The paper uses ResNet-50, while SOTA now uses Vision Transformers
4. Identify: Missing experiments (cross-city generalization, ablation)
5. Generate upgrade plan:
Priority 1: Replace backbone with Swin Transformer (cite 3 papers)
Priority 2: Add cross-city generalization test
Priority 3: Add ablation study for each module
Priority 4: Update related work with 8 missing 2024 papers
6. Simulate a review from RSE reviewer perspective
7. Save: upgrade_plan_urban_land_use_20250610.md, review_report.md
User: "I need to write an NSFC grant on ecological monitoring with AI."
ResearchX auto-chains: Gap Mining → Method Mining → Grant Proposal
1. Search: "ecological monitoring deep learning 2024 2025" (English)
"生态监测 深度学习 2024 2025" (Chinese)
2. Find current gaps: lack of multi-sensor fusion methods
3. Propose research focus
4. Search methods for the chosen focus
5. Generate full grant proposal with:
- 立项依据 (Research Rationale) — literature cited
- 科学问题 (Scientific Questions)
- 技术路线 (Technical Route) — Mermaid diagram
- 创新点 (Innovation Points) — compared to existing work
- 预期成果 (Expected Outcomes)
6. Save: grant_ecological_monitoring_20250610.md
Just mention any research-related need:
# Analyze a paper
"Please analyze this paper for me: [paste DOI or upload PDF]"
# Find a research topic
"I need a research topic in [your domain]. Find me gaps and propose 5 topics."
# Find methods
"What methods should I use for [your problem]?"
# Write a paper
"Write the introduction and methods sections for a paper on [topic]."
# Upgrade to Q1
"I have a paper draft. Help me assess its quality and plan Q1 upgrades."
# Generate visuals
"Create a graphical abstract prompt for my paper on [topic]."
Every module produces a structured markdown file. Example: topics_crop_mapping_20250610.md
## Topic 3: Few-Shot Crop Mapping via Foundation Model Fine-Tuning
**Research Question**: Can a pre-trained foundation model (e.g., GeoFM) achieve
>90% crop classification accuracy with fewer than 50 labeled samples per class?
**Innovation**: First systematic evaluation of foundation model fine-tuning
strategies for crop mapping in data-scarce regions (cf. [Liu, 2024] which
requires 500+ samples).
**Data Needed**: Sentinel-2 imagery + Cropland Data Layer (CDL) labels
**Proposed Methods**:
- Baseline: SSTFormer (SOTA at 95.1% with full data) [Chen, 2025]
- Proposed: GeoFM + LoRA fine-tuning [adapting from NLP, 2024]
- Also evaluate: Prompt-based learning [new, not yet applied to crops]
**Target Journal**: Remote Sensing of Environment, IF 13.5, Q1
**Feasibility**: ★★★★☆ResearchX is available on 10+ platforms and registries:
| Platform | How to Find | Install Command |
|---|---|---|
| GitHub | Search "xingguangYan/ResearchX" | git clone https://github.com/xingguangYan/ResearchX.git |
| Codex (OpenAI) | codex plugin add researchx@personal |
Auto-discovers via triggers |
| Claude Desktop | MCP: researchx-mcp | npx researchx-mcp |
| Claude Code | Copy CLAUDE.md to project root | Auto-reads CLAUDE.md |
| npm Registry | npm search researchx |
npm install researchx-mcp |
| PyPI | pip search researchx-mcp |
pip install researchx-mcp |
| GPT Store | Search "ResearchX" | GPT config in mcp-server/ |
| Cline / Roo Code | MCP: researchx-mcp | Copy .clinerules |
| Continue.dev | MCP: researchx-mcp | Copy .continuerules |
| Cursor | Copy .cursorrules | Auto-reads |
| Windsurf | Copy .windsurfrules | Auto-reads |
Search GitHub for: ResearchX or ai research operating system
Any MCP client can use ResearchX:
{
"mcpServers": {
"researchx": {
"command": "npx",
"args": ["researchx-mcp"]
}
}
}For full instructions: ResearchX/DISCOVERY.md and ResearchX/platforms/PLATFORMS.md.
# From this repo
Copy-Item -Recurse "ResearchX" "$env:USERPROFILE\.codex\skills\ResearchX"git clone https://github.com/xingguangYan/ResearchX.git
Copy-Item -Recurse "ResearchX/ResearchX" "$env:USERPROFILE\.codex\skills\ResearchX"# Windows (Admin PowerShell)
New-Item -ItemType Junction -Path "$env:USERPROFILE\.codex\skills\ResearchX" `
-Target "C:\path\to\ResearchX\ResearchX"# Check the skill is recognized
Get-ChildItem "$env:USERPROFILE\.codex\skills\ResearchX"
# Should show: SKILL.md, agents/, scripts/, references/, assets/ResearchX works across all major AI agent platforms:
| Platform | Config File | Installation | Auto-Discovery |
|---|---|---|---|
| Codex (OpenAI) | ResearchX/SKILL.md | ~/.codex/skills/ResearchX |
Trigger terms |
| Claude Code | platforms/CLAUDE.md | Copy to project root | Auto-reads CLAUDE.md |
| GitHub Copilot | AGENTS.md | Repo root | Auto-reads AGENTS.md |
| Cursor | platforms/.cursorrules | Copy to project root | Auto-reads .cursorrules |
| Cline / Roo Code | platforms/.clinerules | Copy to project root | Auto-reads .clinerules |
| Continue.dev | platforms/.continuerules | Copy to project root | Auto-reads .continuerules |
| Windsurf | platforms/.windsurfrules | Copy to project root | Auto-reads .windsurfrules |
| OpenAI GPTs | platforms/mcp.json | GPT Builder config | Manual |
| MCP Clients | platforms/mcp.json | MCP server config | Manual |
git clone https://github.com/xingguangYan/ResearchX.git
cd ResearchX
# Codex (PowerShell):
Copy-Item -Recurse "ResearchX" "$env:USERPROFILE\.codex\skills\ResearchX"
# Claude/Cursor/Cline/Continue/Windsurf (Bash):
cp platforms/CLAUDE.md ./CLAUDE.md
cp platforms/.cursorrules ./.cursorrules
cp platforms/.clinerules ./.clinerules
cp platforms/.continuerules ./.continuerules
cp platforms/.windsurfrules ./.windsurfrulesFor full instructions, see ResearchX/platforms/PLATFORMS.md.
Each module saves structured files to the working directory:
| File Pattern | Module | Contains |
|---|---|---|
analysis_[Paper].md |
§1 Paper Analysis | Full paper breakdown |
topics_[Topic]_*.md |
§2 Gap Mining | 5 topic proposals |
methods_[Topic]_*.md |
§3 Method Mining | Method landscape |
manuscript_[Topic]_*.md |
§5 Manuscript Writing | Full paper draft |
upgrade_plan_[Paper]_*.md |
§6 SCI Upgrade | Action plan |
review_report.md |
§7 Peer Review | Review + response |
grant_proposal_[Topic]_*.md |
§10 Grant Proposal | Full proposal |
literature_matrix_[Topic]_*.md |
§9 Lit Review | Matrix + related work |
A: Yes. ResearchX is domain-agnostic. It searches literature for ANY scientific topic. The method mining protocol is the same whether you study remote sensing, medicine, materials science, or ecology.
A: ResearchX detects this and falls back to first-principles reasoning, clearly stating the limitation. It will ask you to verify assumptions.
A: Yes. The Python scripts are helper tools for analysis. The core SKILL.md works with web_search alone. The scripts provide deeper analysis (method taxonomy, trend scoring).
A: ChatGPT relies on its pre-training knowledge cutoff. ResearchX actively searches for the LATEST papers (2023-2026) and builds method landscapes from real literature. It also chains multiple research workflows together automatically.
A: No. While the writing templates target SCI journals, the gap mining, method discovery, and experiment design modules work for ANY research output (conference papers, theses, technical reports).
A: If you upload a PDF in Codex, ResearchX will analyze it using the Paper Analysis Engine. It extracts metadata, logic chain, and compares with latest literature.
A: Pull the latest from this repo and copy the folder again:
git pull
Copy-Item -Recurse "ResearchX" "$env:USERPROFILE\.codex\skills\ResearchX"ResearchX aims to be the world's best AI research tool. Contributions are welcome!
- New module: Data visualization, statistical analysis, code generation
- Writing templates: More journal styles (ACM, IEEE, Nature family)
- Domain-specific protocols: Medicine (clinical trial design), Chemistry (lab protocols)
- Scripts: Citation analysis, network graphs, topic modeling
- Translations: Multi-language support for non-English researchers
- Fork the repository
- Create a feature branch
- Make your changes
- Submit a pull request
MIT License — free to use, modify, and distribute.
Built for the global research community.
ResearchX — Moving research assistance from "what I know" to "what the field knows."