"Automation for whom? Researchers operate under different resource configurations, hold different methodological preferences, and target different output formats. A system that produces uniform outputs regardless of these differences will systematically under-serve every individual user."

— NanoResearch (Xu et al., 2026)

NanoResearch takes a one-line research idea — in any scholarly field — and rides it through ideation, planning, experimentation, analysis, writing, and review to produce a downloadable LaTeX paper, while learning your preferences so the next run feels more like you.

✨ Highlights

• 💬 Single chat surface. No buttons. Tell it your field and a topic; it narrates the pipeline back to you and pauses for feedback at every stage.
• 🌍 Field-agnostic. Biology, social sciences, engineering, mathematics, computer science — prompts adapt to the field's conventions (regressions vs. proofs vs. case studies vs. ablations).
• 🧠 Tri-level co-evolution. Per-user Skill Bank (procedural rules), Memory Module (project-specific facts), and a planner adapter trained via SDPO from your free-form feedback.
• 🔬 Real artefacts. Generates and runs a Python project for empirical fields, parses results, then assembles a section-by-section LaTeX paper that compiles to PDF (when pdflatex is installed).
• 🔐 Azure AD auth. Talks to your private GPT-5.1 deployment via DefaultAzureCredential — no API keys in .env.

flowchart LR
  T[Topic] --> O((Orchestrator))
  O -->|retrieve| SB[(Skill Bank 𝒮)]
  O -->|retrieve| MM[(Memory Module ℳ)]
  O --> I[Ideation]:::s1
  I --> P[Planning]:::s1
  P --> C[Coding +<br/>Execution]:::s2
  C --> A[Analysis]:::s2
  A --> W[Writing]:::s3
  W --> R[Review]:::s3
  R --> Paper[paper.pdf]
  W -.->|distil| SB
  W -.->|distil| MM
  I -.->|narrations| U[Chat]
  P -.-> U
  C -.-> U
  A -.-> U
  W -.-> U
  U -.->|feedback ℱ| O
  classDef s1 fill:#1e3a8a,stroke:#3b82f6,color:#fff
  classDef s2 fill:#92400e,stroke:#f59e0b,color:#fff
  classDef s3 fill:#065f46,stroke:#10b981,color:#fff

🗂 What's in the box

🚀 Quickstart

Prerequisites

• Python 3.11+ (3.12 tested)
• Node 18+ (Vite + assistant-ui)
• An Azure OpenAI / Foundry deployment of GPT-5.1 (or a compatible reasoning model)
• az login performed locally; your account needs the Cognitive Services OpenAI User role
• (optional) pdflatex or tectonic for PDF compilation — otherwise the paper ships as .tex
• (optional) Apple-Silicon Mac with 32 GB+ unified RAM for the local Qwen planner (SDPO)

Setup

# 1. Clone + venv
git clone https://github.com/saadmsft/nanoresearch.git
cd nanoresearch
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# 2. Configure Azure (AAD auth — no API keys)
cp .env.example .env
# edit AZURE_OPENAI_ENDPOINT + AZURE_OPENAI_DEPLOYMENT
az login

# 3. Backend
nanoresearch serve            # http://127.0.0.1:8000

# 4. Frontend (separate terminal)
cd ui && npm install && npm run dev   # http://localhost:5173

Open http://localhost:5173 and say hi.

Optional: local SDPO planner

pip install -e ".[local]"     # torch, transformers, peft, accelerate
huggingface-cli download Qwen/Qwen2.5-7B-Instruct \
  --local-dir data/models/Qwen2.5-7B-Instruct

💬 What it looks like

You ▸ I'm Mia, an ecologist. I prefer field studies, 6-month timeline.
      Start a run on canopy cover and breeding-bird richness in city parks.

NanoResearch ▸ Nice to meet you, Mia. Starting on canopy cover + bird richness.
NanoResearch ▸ 🔎 Searching scholarly databases…
NanoResearch ▸ 📚 Done. 12 papers.
NanoResearch ▸ 💡 Drafted hypotheses (n=6). Checking novelty next.
NanoResearch ▸ 🎯 Going with: Canopy × heterogeneity interaction predicts richness.
NanoResearch ▸ ⏸ Paused at ideation — what should I emphasise or change?

You ▸ Keep the design simple and proceed.

NanoResearch ▸ 📐 Drafting an experiment blueprint…
NanoResearch ▸ 👀 Running an internal peer review of the blueprint…
NanoResearch ▸ 🧪 Writing a small experiment project to test the plan…
NanoResearch ▸ ▶️ Running the experiment…
NanoResearch ▸ 📈 Run finished (ok=True exit=0 dur=3.2s).
NanoResearch ▸ 📊 Analysing results…
NanoResearch ▸ ✍️ Drafting the introduction / method / experiments / … sections.
NanoResearch ▸ 👓 Reviewing the paper draft.
NanoResearch ▸ 📄 Paper compiled. [Download PDF] — or the [LaTeX source].

🏛 Architecture

NanoResearch is a stage pipeline orchestrated around two persistent stores and one trainable planner.

flowchart TB
  subgraph User["👤 User"]
    Chat["💬 Chat UI"]
  end

  subgraph API["⚡ FastAPI"]
    Intent["/api/intent<br/>NL → action"]
    RunMgr["RunManager<br/>background thread"]
    Narr["Narrator<br/>event → English"]
    SSE["SSE /stream<br/>+ narration"]
    Files["paper.pdf<br/>paper.tex"]
  end

  subgraph Pipe["🔬 Pipeline (Orchestrator)"]
    direction LR
    I[Ideation] --> P[Planning]
    P --> C[Coding]
    C --> An[Analysis]
    An --> W[Writing]
  end

  subgraph Stores["💾 Per-User Stores"]
    Profile[(Profile)]
    Skill[(Skill Bank 𝒮)]
    Mem[(Memory ℳ)]
    LoRA[(LoRA adapter)]
  end

  subgraph LLMs["🤖 Models"]
    Azure[Azure OpenAI<br/>GPT-5.1]
    Qwen[Qwen2.5-7B<br/>local · planner only]
  end

  Chat <-->|HTTP| Intent
  Chat <-->|EventSource| SSE
  Chat -->|download| Files
  Intent --> RunMgr
  RunMgr --> Pipe
  Pipe --> Narr --> SSE
  Pipe <-->|retrieve / distil| Stores
  Pipe -->|complete| Azure
  Pipe -->|plan| Qwen
  Qwen <-->|SDPO LoRA| LoRA

📖 Full architecture deep-dive: docs/architecture.md · paper §3 mapping · SDPO math

🧪 Testing

pytest -m "not azure and not local_model"   # 61 offline tests
pytest -m azure                              # AAD smoke
pytest -m local_model                         # Qwen MPS smoke

📂 Repository layout

nanoresearch/
├── src/nanoresearch/
│   ├── agents/           # Stage I-III stage controllers + prompts + artefacts
│   ├── api/              # FastAPI app, RunManager, intent classifier, narrator
│   ├── cli/              # `nanoresearch serve`, `health`, `settings`
│   ├── config/           # pydantic-settings
│   ├── literature/       # OpenAlex client + evidence extraction
│   ├── llm/              # Azure (AAD) + local Qwen backends, agent-role router
│   ├── logging/          # structlog + per-run JSONL manifest
│   ├── orchestrator/     # Retrieve → Plan → Dispatch → Reflect → Update
│   ├── planner/          # Qwen wrapper + LoRA + SDPO trainer (Eq. 14-15)
│   ├── schemas/          # Profile / Skill / Memory pydantic models
│   └── stores/           # SkillBank + MemoryStore + Profile (JSON-backed)
├── ui/                   # React + TypeScript + Tailwind chat
├── docs/                 # Jekyll site (GitHub Pages)
├── tests/                # Pytest suite
├── runs/                 # ← created at runtime (event logs + papers/<run>/paper.tex)
└── data/users/<id>/      # ← created at runtime (profile, skills, memories, lora)

🗺 Roadmap

• Phase 0–4 — bootstrap, stores, planner+SDPO, orchestrator, Stage I (Ideation + Planning)
• Phase 5 — Stage II (Coding + sandboxed exec + debug loop) + Analysis
• Phase 6 — Stage III (Writing + Reviewer + LaTeX/PDF)
• FastAPI + React/Vite UI with live SSE narrations
• Phase 7 — Compliance/Novelty/Writing judges (paper §8–10) + 20-topic benchmark harness
• Phase 8 — CLI ergonomics (nanoresearch run, nanoresearch eval)
• Docker sandbox upgrade for Stage II
• Per-section figure generation + bibliography auto-fill

📜 Citation

If this implementation is useful in your research, please cite the original paper:

@misc{xu2026nanoresearch,
  title  = {NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automation},
  author = {Xu, Jinhang and Zhu, Qiyuan and Wu, Yujun and Wang, Zirui and Zhang, Dongxu and others},
  year   = {2026},
  eprint = {2605.10813},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url    = {https://arxiv.org/abs/2605.10813}
}

📄 License

Apache 2.0 — see LICENSE.

Original NanoResearch paper © Xu et al., 2026.
This implementation is independent and not affiliated with the original authors.